Energy Systems and Control

[CE295]

Book of Projects 2018

Instructor: Professor Scott Moura
Grad. Stud. Instructor: Bertrand Travacca
 Preface

We are extremely proud to present the first book of projects made for the
CE295 class (Energy Systems & Control). CE95 is a class given every year in the
spring at UC Berkeley in the department of Civil & Environmental Engineering.
This class gives an introduction to control system tools for students interested in
energy system applications. Applications of interest include batteries, electric
vehicles, renewable energy, power systems, smart buildings/homes and more!
Technical tools include system modeling, state-space representations, stability,
parameter identification, state observers, feedback control, and optimization.
Each year students engage in a semester-long course projects. For many
years we have noticed the quality of the projects. The class of 2018 displayed
such excellence that we decided to collect all the projects here. This book is
aimed at future students, who will find essential material to get inspiration and
build upon previous work made in the class. Note that this book could also
be of interest to people from industry or academia as truly innovative ideas are
presented.
We organized the book in five different thematic parts: Microgrids (part I),
Electric Vehicles and Transportation (part II), Building Energy Management
(part III), Batteries and Energy Storage (part IV), Health and Environment
(part V).
We wish you a pleasant reading,

Professor S.Moura,
B.Travacca

UC Berkeley, CE295
 Table of Contents

Part I: Microgrids p.1-60

• Microgrid Control Platform p.2-22
• Grid Optimization and Resiliency Study
for the Puerto Rican Electricity Grid p.24-39
• Robust Optimal Sizing and Operation of a Microgrid with Electric Ve-
hicle Charging and Renewable Energy Generation Grid Optimization
and Resiliency Study
for the Puerto Rican Electricity Grid p.41-59

Part II: Electric Vehicles and Transportation p.61-135

• The Application of Estimation Techniques to Autonomous Vehicles Re-
lated Object Tracking Problems p.62-81
• Nanjing Metro traveler entrance modeling and minimization of system-
wide wait time by optimizing train distribution p.83-99
• Minimizing Ridesplitting Vehicles Miles Traveled p.101-115
• Electric Vehicle Charging Controller p.117-134

Part III: Building Energy Management p.136-170

• Predicting Building Electric Loads under Climate Change p.137-153
• Air Conditioning Control System with Prediction of Occupant Flow
p.155-169

Part IV: Batteries and Energy Storage p.171-200

• Minimization of cumulative aging in batteries: a grid-based approach
p.172-183
 • States Estimation of Li-ion Battery p.185-199

Part V: Health and Environment p.201-268

• Mathematical Modeling of an Ecosystem Network p.202-2017
• Blood Glucose Prediction p.219-233
• Optimization of Rainwater Harvesting Systems for Single-Family Resi-
dential Buildings in California p.235-250
• Wildfire Modeling: A Case Study on the 2013 Rim Fire p.252-267
1

Part I

Microgrids

UC Berkeley, CE295
2

Microgrid Control Platform

Ramon Crespo, Ioanna Kavvada, Guo Jun Li and Dieter Smiley

Abstract
This study designed, implemented and analyzed an optimization program as a Microgrid
Control Platform (MCP). The MCP takes into account the state of each of the Microgrid’s
assets, Photovoltaic (PV) generation, electric vehicle (EV) demand and building load de-
mand, as well as future costs to electricity. The objective of the optimization program
is to minimize operational costs. The forecast states of PV generation, EV demand and
building load demand are based on forecasts made from individual data sets, using Markov
chains, ARX machine learning models and random forest regression in each of the respec-
tive modeling processes. Ultimately, results show that utilizing these forecasting methods in
a moving-time-horizon linear optimization program may yield additional revenue beyond a
simplified rules-based control scheme. The results also indicate that the optimal time hori-
zon for standard operations is 12 hours. Future research should focus on shortcomings of the
MCP designed, which includes accounting for seasonal variation and improving robustness
to forecast errors.

Introduction
Motivation and Background
The world energy demand is projected to rise through 2050 with an average annual
growth about 0.9 percent, as reported in [1], and is expected to increase energy-related
carbon dioxide emissions. Energy generating units based on renewable energy sources are
major components of the strategy to reduce harmful emissions and deal with depleting energy
resources. However, they are less reliable as compared to the conventional fossil fuel-based
power generating systems due to their intermittent nature, as stated in [2]. Integration of
different renewable energy sources coupled with energy storage system can add reliability in
the power systems.
In light of this trend, a significant shift in the electricity structure of communities is
needed. These communities will require grid resiliency and reliability, consideration of
carbon-free energy resources while also taking into account the economics and profitabil-
ity of buying and selling electricity in a shared grid. One such vision is a decentralized
electricity structure where communities are semi-self sufficient through distributed energy
resources locally while still possessing a connection to the grid, also known as smart grids.

UC Berkeley, CE295
3

Smart grids consist of numerous components and controls that are still being developed
and tested. According to [3], the full implementation of the smart grid will happen in the
next decade, when technology is mature enough to be implemented in a commercial scale.
As the technology and efficiency of the components of the smart grid improves, the need for
the right control tools increases. The MPC seeks to fill this gap. The platform will calculate
the optimal use of resources to maximize energy efficiency and minimize CO2e emissions
by taking into account forecasts of pricing, building load demand, other load demands of a
community as well as variability in renewable generation.

Focus of this Study

The focus of the study is the design and implementation of a MCP that takes into account
PV generation forecasts, building energy demand and EV energy demand to optimize the use
of electricity, from and economic and environmental point of view. In particular, emphasis
is placed on cost performance, the time horizon required for adequate performance and the
MCP’s robustness to error.

Literature review
This project will use the design of the EcoBlock project in Oakland, California to test the
MPC. EcoBlock is an existing project that seeks to design, build and test a solar-powered
urban system that uses PV energy generation and flywheel storage to meet the electric
demand of the buildings and EV fleet. Though the MCP will be designed using an existing
project, it will be easily scalable for implementation on other systems.
The definition of a Microgrid may vary dependent upon the article, but generally tends
to contain three important characteristics: a group of interconnected load and distributed
energy resources (DERs); a clear point of common coupling with the larger electric grid;
and an ability to connect and island itself from the grid. More generally, [4] describes that
Microgrids contain sources of generation, load, and components capable of energy storage
that are connected to the macro grid at a single point. Energy storage is an important
consideration for the optimal performance of a Microgrid, as it enables an energy system to
balance supply and demand, and enables the shifting of dispatchable energy in time.
In addition to the necessary physical equipment for generation, storage and dispatching
electricity, other components are required to design an optimal Microgrid. An essential
ingredient of Microgrids and in optimal energy management is the utilization of energy
controllers, or sensors capable of providing accurate information in real time and making
decisions. According to [5], the capability of electricity networks to “facilitate the fluent
interaction of all users connected to it” enables bi-directional power flow and is a vague
definition of the smart grid concept. Controller and sensors allow for consumption-driven
adjustment by monitoring energy use performance in real time or in discrete time intervals.
This data is also used to generate predictions for future energy consumption and generation.
Many methods of forecasting electricity demand and generation exist and the optimal
method may vary dependent upon application and prediction horizon. As in [6], some
common forecasting methods reviewed include time series, regression, and autoregressive
integrated moving average. Beyond attaining data and controls, to provide optimal energy

UC Berkeley, CE295
4

management there exists a need for real-time computational intelligence methods such as an
optimization program, as in [7]. The inclusion of correct data for the programs application
are critical to the utility of the program. Appropriate information may include historic and
forecasted weather data, expected load profiles for the microgrid, and knowledge or predic-
tions of the pricing schedule for retail electricity rates. The objective of the optimization
algorithm influences what may be considered relevant data. For example, other considera-
tions and goals of a Microgrid which are not in the scope of this report may include peak
load shaving or voltage regulation, as described in [8]. Different objective functions will,
necessarily, use different optimization algorithms.
Cost reduction and DER power generation maximization are two main objectives when
planning for the deployment of renewable energy sources. Different optimization techniques
have been proposed to achieve these two objectives. [9] describes various methods of op-
timization techniques for various renewable energy sources. Other papers, as in [2], show
various optimization models, such as renewable energy models, emission reduction models,
energy planning models, energy supply and demand models, forecasting models and control
models for efficient utilization of the renewable energy sources. Ultimately, the real time
demand response model over a twenty-four hour planning horizon found in [5] served as a
framework for our team’s energy management controls.
Renewable energy sources can be used in grid connected as well as in stand-alone modes.
Grid connected hybrid renewable energy sources are promising areas of research as these are
expected to provide the same level of power supply reliability that can be achieved from the
conventional fossil fuel based energy systems.

Key contributions
Our research has contributed to the understanding of time horizon selection when at-
tempting to minimize computational power, as well as a better understanding on how to
forecast Microgrid components and apply them to a simple MCP program.

1 Technical Description of MCP

Here we will outline in detail (i) the technical components of the Microgrid studied;
(ii)how our MCP optimizes for cost-minimizing Microgrid operations; and, (iii) how the
forecast of each Microgrid component of interest is generated.

1.1 Descpription of Microgrid

The MicroGrid (Figure 1) consists of 63 residential buildings, 24 electric vehicle supply
equipment (EVSE) each serving one electric vehicle, PV area of 27,750 f t2 and 480 kWh of
flywheel storage. Solar power generated by the solar panels is collected at a single point and
can be either stored in the flywheel, sold to the grid or used to satisfy EV and house load
demand. The MicroGrid acts as a unit with a single connection to the grid.

UC Berkeley, CE295
5

Figure 1: Schematic description of the microgrid

1.2 Optimization Program Formulation

The feasibility and environmental-friendliness of a MicroGrid is ultimately based upon
the cost-effectiveness of the Microgrid, or how much energy needs to be bought or sold
from the electrical grid. Hence, the MCP designed and implemented in this study focuses
on minimizing the overall operations cost over a set time horizon. This is done through a
moving-time-horizon linear optimization program. The program setup is listed below.

Objective function:
k+L
X k+L
X
min cI (i)GI (i)∆t − rE (i)GE (i)∆t (1)
i=k i=k

Subject to: f or i = k, k + 1, ..., k + L

S(i) + Bd (i) + GI (i) = L(i) + Bc (i) + GE (i) + C2,1 (i) (2)

0 ≤ GI (i) ≤ GI,max , 0 ≤ GE (i) ≤ GE,max (3)

1
E(i + 1) = E(i) + [ηc Bc (i) − ( )Bd (i)]∆t (4)
ηd
E(k) = Emeas (5)
E(k + L) = Ef inal (6)
0 ≤ E(i) ≤ Emax (7)
0 ≤ Bc (i) ≤ Bmax , 0 ≤ Bd (i) ≤ Bmax (8)
Z2,1 (i + 1) = Z2,1 (i) + η2 C2,1 (i)∆t (9)
Z2,1, (k) = Z2,1,k,meas (10)
Z2,1,min (i) ≤ Z2,1 (i) (11)
0 ≤ C2,1 (i) ≤ C2,1,max (12)
A list of variables and their definitions are provided in the Appendix.

UC Berkeley, CE295
6

In essence, this optimization program minimizes the total cost (cost minus revenue) of
importing or exporting electricity over the next L time periods, as shown in (1). In (1),
we see that cost or revenue at each time step is determined by term G(i), which, in (2),
we see is simply the difference between supply of energy and consumption of energy within
the Microgrid. In order to control the outflow and inflow of energy based on forecast price,
generation and consumption, the flywheel storage, or Bc and Bd terms in (2), are used as
the control for the system.
The remainder of the optimization program is limited by power and energy constraints,
as well as the dynamics of each component. The power limits of the grid are included in
the optimization through (3). The flywheel storage dynamics are shown through (4), where
Bc and Bd are the charging and discharging mechanisms of the storage system. The initial
and final conditions for the flywheels are shown through (5) and (6), respectively, and the
limits of the energy state of the flywheel storage is shown through (7). Charging constraints,
or the maximum charging can occur at each time step, is represented through (8). Finally,
EV demand is characterized by a cumulative EV energy demand, represented by Z2,1 , and
a charging power, represented by C2,1 . From the constraints, (9) represents the dynamics of
EV charging, while (10) shows the initial condition of EV charging. The energy and power
limits constrain the optimization problem via the inequality constraints (11) and (12).
Note that there is no need to model the conditions of the flywheel in our study, as flywheel
losses are assumed to be very low compared to the energy being transferred on an hourly
basis. The flywheel storage is simply computed iteratively at each time step given the value
at previous time step and charging or discharging commands. The charging and discharging
efficiencies are 0.95 and 0.95, respectively, and are found in [10].

1.3 Microgrid Component Forecasting

In order for the supply and consumption of energy by the Microgrid to be balanced as in
(2), each component must be modeled separately. Here, we show how we used prior data to
model each of the Microgrid’s assets.

1.3.1 Electric Vehicle Demand

Electric vehicle demand is taken from [11]. The year-long data acquired was used to
create a model that predicts the future EV demand based on previous data. Ultimately, two
models were used and compared to see what differences could be observed through using a
variation of an Average Model and a Machine-Learning model.
For the average model, the average EV demand for the month in question (May) was
used. Then, a normally-distributed error is injected into each hour. This is shown in the
formulation below.
!
Eavg (i) − E(i) |Eavg (i) − E(i)|
E(i + 1) = Eavg (i) + N , (13)
2 8

Essentially, as shown in (13), the model attempts to correct for deviations from the average
by compensating with a normally distributed error that is centered closer to the average. The
standard deviation of the normally-distributed component in (13) was chosen arbitrarily.

UC Berkeley, CE295
7

The resultant graph of predictions for a 24-hour time-horizon is shown in Figure 2.

Figure 2: Average Model forecasting of EV Demand (all colored lines) as compared to actual
EV Demand (solid black line)

The second model used was a 3-hour previous data input, Autoregressive with Exogenous
Inputs (ARX) Model. Here, year-long data from [11] was used as training data to fit variables
to a linear function that forecasts the next time step by linearly combining the monthly
average with the three previous data points. This is formulated below.

E(i) = Eavg (i) + α1 E(i − 1) + α2 E(i − 2) + α3 E(i − 3) (14)

From (14), we can then use ordinary least squares regression to find the α coefficients in
front of each previous term injected into the model. The resultant coefficients generated
from the regression is shown in (15).
" α # " 0.3386 #
1
α = α2 = -0.1353 (15)
α3 -0.1014

The resulting 24-hour forecast from this machine learning model is shown in Figure 3.

Figure 3: ARX Model forecasting of EV Demand (all colored lines) as compared to actual
EV Demand (solid black line)

UC Berkeley, CE295
8

From Figures 2 and 3, we observe that the Average Model creates predictions that have
high random noise and does not reflect the true nature of EV demand as often as the machine
learning model does. This is reflected by the RMSE for each hour predicted by each model.
The Average Model’s RMSE values hovers around 0.35 kWh constantly, while the RMSE
for the machine learning model begins at 0.34 and extends higher as time progresses. The
slightly lower RMSE values for the ARX model, combined with the overall better shape of the
predictions as shown in Figure 3 shows that the ARX model is a better forecast mechanism
for the data at hand.
Ultimately, because of these reasons, the ARX model was used as the input data for the
MCP in this study.

1.3.2 Photovoltaic Generation

Solar energy, which comes from the sun in the form of solar irradiance, can be directly
converted to electricity by using photovoltaic (PV) technology. PV technology uses solar cells
made of semiconductors to absorb the irradiance from the sun and convert it to electrical
energy. The typical solar PV cell has an efficiency of about 15–20% [12].
The characteristics of a PV module can be demonstrated by power–voltage or cur-
rent–voltage curves. Figure 4 shows the power–voltage curve of a PV module for different
conditions of solar irradiance and cell temperature. The PV output power is dependent on
solar irradiance and cell temperature. Low irradiance leads to low power, and high temper-
ature causes a reduction in output power. The point on the curve at which the PV module
delivers maximum power to the load is known as maximum power point (MPP). In the PV
system, we assume that a maximum power point tracker will be used.

Figure 4: Power–voltage curve of a PV module for different conditions of solar irradiance

and cell temperature [12].

The maximum power output is presented by [13]:

P = η × S × I × (1 − βref × (T − Tref )) (16)

Where η is the conversion efficiency of the solar cell array (%) at the reference temperature
Tref at solar radiation flux of 1000 W/m2 as seen in [14], S is the array area (m2 ), βref is
the temperature coefficient at reference conditions, I is the solar radiation (kW/m2 ) and T
is the outside air temperature (◦ C).

UC Berkeley, CE295
9

The average value of βref reported in [14] is 0.0045 (◦ C)−1 at Tref = 25◦ C. In this
analysis, the conversion efficiency of the solar cell array was considered to be 15%.
Hourly data regarding the Global Horizontal Irradiance (GHI), the total amount of short-
wave radiation received from above by a surface horizontal to the ground, and the outside
temperature for a typical meteorological year was retrieved from the National Renewable
Energy Laboratory (Sacramento Municipal Utility District station) for the years 1998-2016.
Using equation 16, historical solar power generation data was generated.
Each year was split into 4 seasons that exhibited similar solar power generation profiles
(Season 1: May, June and July, Season 2: November, December, January and February,
Season 3: October and March, Season 4: April, August and September). For each season, the
hourly solar power generation values are not normally distributed about the hourly averages
(Figure 5) and the hourly PV generation values are correlated to power generation of the
previous time steps. Therefore, to take into account the correlation with previous time steps,
a discrete-time Markov Chain model was selected as the most suitable forecasting method
for PV power generation.

Figure 5: Non-normal solar power generation distribution (summer-time)

The transition probabilities for each time k were defined as follows:

pijk = P r[Sk+1 = S j |Sk = S i , k] ∀S i , S j ∈ S, k = 0, ..., 23 (17)

Variables S i , S j are different levels of solar power generation within a set S. Set S was created
by discretizing the solar power generation values into 14 evenly spaced levels as follows:

S = {Smin + 0 × (Smax − Smin )/12,

Smin + 1 × (Smax − Smin )/12,
..., (18)
Smin + 11 × (Smax − Smin )/12,
Smin + 12 × (Smax − Smin )/12}

The transition matrices were then used to create the cumulative transition probability
matrices for each time of day k. An example of the PDF and CDF for a specific initial
solar power generation level and time of day is presented in Figure 6. Using the CDFs
and a uniform random number generator in the range 0 to 1, given the current solar power
generation level and the current time step k, the solar power generation level of the next
time step k+1 was predicted. Each time step has an initial solar power generation level and

UC Berkeley, CE295
10

a uniformly distributed pseudorandom number that has been generated in the range from
0 to 1. For the initial power generation lever, the pseudo random number is mapped onto
the CDF to find the corresponding power generation level of the next time step by using
linear interpolation. The predicted solar power generation level then became the current
one for the next iteration. By performing N iterations, solar power generation levels can be
predicted for a N-hour time horizon. Figure 7 presents the forecasted solar power generation
values using a 6h-time-window for a 4 day period.

Figure 6: Transition probabilities for a specific solar power generation level and specific time
of day

Figure 7: Forecasted solar power generation values using a 6h-time-window (red) and actual
solar power generation values (black)

1.3.3 Building Load Demand

The data compiled to create prediction models and forecasts for building energy con-
sumption was obtained by the EcoBlock team. This data was gathered from historic electric
consumption recorded by smart meters and accessed through [15]. The historic electric de-
mand readings varied among residents, depending on when their service with PG&E started
and when their smart meter was installed, with some data sets dating back to the first
quarter of 2016.
From [15], energy consumption of 23 individual housing units were obtained and filtered
to ensure feasibility. In our study, we assume that at no time can a building realistically have

UC Berkeley, CE295
11

zero demand, so any hours containing zero energy consumption were removed. Initially, for
every hour, an arithmetic average and median were determined to provide an initial estimate
load profile, per unit, for any given day. The average and median of the raw data is shown
in 8, and the accompanying mean and standard deviation are shown in the corresponding
table in the Appendix. The data indicates that these residential building experienced two
peak loads in a given day, with the maximum of the two occurring in the evening hours.

Figure 8: Average and Median Hourly Energy Demand for EcoBlock

Random forest regression was used in creating a prediction model for estimating building
energy demand. According to [16], Random Forest Regression (RFR) is a tree-based method
of supervised machine learning that fits prediction models to training data by using decision
trees, random sampling, and a randomly selected subset of available parameters from which
to sample. In building the decision tree, RFR stratifies the predictor space, the set of all
possible values for parameters X1 , X2 ,..., Xp into J smaller, distinct subsets, R1 , R2 ,...,RJ .
For each subset, RFR evaluates the mean response variable yR , which in the case of building
energy demand forecasting is kilowatts consumed in an hour.The RFR methodology iterates
through the available parameters and regions at each decision node and determines the
parameter, X, and splitting point, s, within the parameter, that provides the lowest total
error in prediction, in the form of residual sum squared (RSS).
This is neatly described by [16] in (19) and (20).

R1 (j, s) = {X|Xj < s}, R2 (j, s) = {X|Xj > s}, ∀j, s (19)
X X
min (yi − yR1 )2 + (yi − yR2 )2 (20)
i:xi ∈R1 (j,s) i:xi ∈R2 (j,s)

Where yR1 is the mean response of training observations in R1 , yR2 is the mean response of
training observations in R2 and yi is the actual observed response in the data.
The process of creating predictive models using RFR followed a systematic process: as-
semble variables of interest from available data that were expected to have strong explanatory
power of energy consumption, use the training data to fit a random forest regression model in
Python, and use that model to predict energy consumption for a test data set of two weeks.
Examples of parameters expected to be powerful predictors and used within the regression
include several hours of previous energy consumption of a building, building index within
the data set, date and the time of the data point. If temperature data was available, it

UC Berkeley, CE295
12

would have been included, as [17] has shown that temperature has high predictive power for
building energy consumption.
After generating predictions for various models, the mean absolute error (MAE) was
evaluated for each model relative to the actual energy consumed from the test data. The
best model, Model 36, was selected as that model, with it having the lowest MAE, 0.16 kWh
(see orange line in 9. 9 compares Model 36 to actual data and average energy consumption
for hour and day. The best performing model included building index number, hour of the
day, day of the week, month of the year, and importantly, the previous two hours of actual
energy consumption.

Figure 9: Forecast model output for building load demand compared to actual test data and
average from all historic data.

1.3.4 Component Forecast Error

An important step in determining prediction models is establishing criteria for model
evaluation. For each forecast method described, error was necessarily considered with respect
to the test data set, i.e. the true values which the team is attempting to accurately predict.
Typical error metrics considered are the mean absolute error, mean square error, and root
mean square error.
Each component forecast method was designed with the ability to make predictions for
every hour over a subsequent 24 hour time horizon. We decided to evaluate the mean absolute
error of each model for every hour in the moving prediction horizon over the two week period
of the test data. Figures 10, 11, and 12 describe the errors of each model. To enable easier
comparison, the average mean absolute error of each predictive model is described as a
percentage of its respective maximum hourly energy consumption or generation throughout
the test period.
It appears that, for all models, accuracy of the predictions tend to worsen for estimates
nearer the end of the prediction horizon. This indicates it may be more useful to run
the optimization program more frequently than once every 24 hours. By evaluating the
optimization program every hour, we expect to maintain a lower error between the decision
variable implemented due to the forecast and the true energy need due to the actual energy
balance from all microgrid components.

UC Berkeley, CE295
13

Figure 10: Forecast Error for Each Hour in the Prediction Horizon. MAE is 11.7% of Test
Maximum Hourly Energy Demand

Figure 11: Forecast Error for Each Hour in the Prediction Horizon. MAE is 12.3% of Test
Maximum Hourly Energy Generation

Figure 12: Forecast Error for Each Hour in the Prediction Horizon. MAE is 24.6% of Test
Maximum Hourly Energy Demand

2 MCP Performance and Discussion

Here we outline the three major results of this study: (i) the performance of the MCP
in general conditions, (ii) the minimum time horizon needed to reach standard operating
conditions and (iii) the robustness of the MCP to error.

UC Berkeley, CE295
14

2.1 Performance of MCP

The performance of the MCP was compared between two main modes of operation: (i)
under the assumption of complete clairvoyance and (ii) under the forecast data that was
provided using the tools described in sections 1.3.1 through 1.3.3. The results are shown in
Figures 13 and 14.

Figure 13: MCP Strategy implemented using a 12-hour time horizon and forecast tools.

Figure 14: MCP Strategy implemented using 12-hour time horizon assuming full clairvoyance

It is worth mentioning that the optimization program never corrects for errors in the
forecast, meaning that the actual values of consumption and generation are not known and
the simulation is completely guided by the forecast values. Nonetheless there is an adjustment
in the forecast values to the true values of consumption and generation that occurs in the
forecast algorithms where the forecast are adjusted each hour to correct for actual conditions.
Figure 13 contains valuable insight into the operations and strategy choice of the MPC.
First and foremost, the shape resembles that of the case of perfect clairvoyance in Figure 14.
It is apparent that the strategy of the forecast system is similar to that of full clairvoyance - to
charge the flywheels during periods of high solar generation while discharging them during
periods of low generation and high energy cost, as illustrated by the changes in flywheel
state-of-charge in the figure. This concurs with intuition as it is more economically feasible

UC Berkeley, CE295
15

to perform arbitrage when knowledge of the future is possible. This occurs over both days
in Figure 13.
By applying this to a full 14-day period, we were able to obtain 41% of the revenues
possible from full clairvoyance. A breakdown of this can be found in the Appendix.

2.2 Selection of Time Horizon

Computing power is not always a luxury, and so one performance metric assessed was the
minimum forecast time horizon needed to achieve benchmark performance (at standard con-
ditions). In our study, we varied the time horizon from 2 to 24 hours to see how performance
changes. The results were normalized and produced in figure 15.

Figure 15: Normalized performance of MCP when varying time horizons used.

Here, normalized performance is computed by comparing earnings via the time horizon
as a percentage of maximum theoretical earnings, assuming full clairvoyance. As figure 15
shows, performance decreases as time horizon shortens, but reaches a maximum threshold at
around a 12-hour time horizon. Here, the reasoning behind this performance change matches
with intuition - as time horizon decreases, the MCP is unable to fully utilize its storage to
account for future changes in conditions of its assets. As time horizon increases, the MCP
is able to better predict the future (it will know more) and be able to account for changes
better. Further assessment of these trends can be found in the Appendix.

2.3 Error Sensitivity Analysis

Finally, error sensitivity is a large part of this MCP as the optimization program used
does not account for actual values. As formulated, the program sees no value in keeping
charge in the flywheels in case of large errors in the forecast. This is apparent in figure 13,
during the second period at time 10am when there is a higher than usual energy demand
with a low energy generation. At this point it was necessary to import energy from the grid
at a point when energy cost was high. To address the limitation, a penalty term would need
to be included in the formulation of the optimization equation, where the term would force
the flywheels to possess some charge at every point of the day that would only be used as

UC Berkeley, CE295
16

backup for unforseen moments of low generation with high energy demand. This could be
valuable during the more variable seasons of the year.

3 Conclusions and Findings

Through this study, we have demonstrated that the design and implementation of a
simple moving-time-horizon linear optimization program as an MCP can produce revenues
for Microgrids with variable generation, energy storage and EV demand. Given data of
PV generation, EV demand and building load demand, we have used Markov chains, ARX
models and Random Forest Regression to forecast future data points. This has produced
results that have reached 41% of maximum theoretical revenues.
The shortcomings of our MCP are aligned with the future work that can be done. The
MCP used in our study lacks the ability to correct for high variability in actual energy
consumption or production, and also does not account for variability in season. Further
research can be done to account for seasonal changes in energy consumption, as well as
improve on forecasting algorithms to ensure that actual data is taken into account during
the forecasting process.

Executive Summary
Energy consumption is predicted to rise greatly over the next century, and the integration
of renewables into the current electricity mix is needed to lower energy-related carbon emis-
sions. Much of this can be done through Microgrids, though a cost-minimizing Microgrid
Controls Platform (MCP) is needed. Here, we have designed an MCP that uses a moving
time horizon linear optimization program to make decisions on whether a Microgrid should
store, sell or buy energy from the grid based on generation and consumption forecasts of
the Microgrid’s assets. These forecasts are generated using different methods for each asset:
Random forest regression for building load forecasting, ARX and machine learning models for
EV demand forecasting and Markov chains for PV generation forecasting. Each method was
chosen after experimentation with different models and picking the model with the smallest
error and deviation from actual data.
When operating with only forecast data, the MCP produces decisions that closely align
with the case of full clairvoyance. The strategy chosen matches with the optimization objec-
tive chosen - the program typically decides to perform arbitrage in order to gain economic
benefit. One major shortcoming is that the program fails to address large deviations from
predictions, as there is no current means of re-injecting actual data back into the model.
The optimal time horizon that would balance computational power and operational stability
was found to be 12 hours, and was used as the best-case operations scenario in our study.
Future work should focus on accounting for seasonal variance in energy consumption as well
as robustness against errors in forecasting.

UC Berkeley, CE295
17

References
[1] “Annual energy outlook 2018,” U.S. Environmental Information Agency, 2018. [Online].
Available: https://www.eia.gov/outlooks/aeo/

[2] A. H. Fathima and K. Palanisamy, “Optimization in microgrids with hybrid energy

systems - a review,” Renewable and Sustainable Energy Reviews, vol. 45, pp. 431–446,
2015.

[3] “What is the smart grid,” U.S. Department of Energy, 2018. [Online]. Available:
www.smartgrid.gov/the smart grid/smart grid.html.

[4] C. Sun, F. Sun, and S. Moura, “Data enabled prediction energy management of a pv-
battery smart home nanogrid,” American Control Conference, pp. 1023–1028, 2016.

[5] A. Conejo and L. Morales, J.and Baringo, “Real-time demand response model,” IEEE
Transactions on Smart Grid, vol. 1, pp. 236–242, 2011.

[6] L. Suganthi and A. Samuel, “Energy models for demand forecasting - a review,” Re-
newable and Sustainable Energy Review, vol. 16, pp. 1223–1240, 2012.

[7] C. Colson, M. Nehrir, and S. Pourmousavi, “Towards real-time microgrid power man-
agement using computational intelligence methods,” 2010.

[8] “IEEE 1547 standard for interconnecting distributed resources with electric
power systems,” IEEE Standards Association, 2018. [Online]. Available: http:
//grouper.ieee.org/groups/scc21/1547/1547 index.html

[9] R. Banos, F. Manzano-Agugliaro, F. Montoya, A. Alcayde, and J. Gomez, “Optimiza-

tion methods applied to renewable and sustainable energy: A review,” Renewable and
Sustainable Energy Reviews, vol. 15, pp. 1753–1766, 2011.

[10] S. Schoenung, “Energy storage systems cost update,” SANDIA Report, 2011.

[11] “Dataport API data set,” Pecan Street Inc., 2018. [Online]. Available: http:
//www.pecanstreet.org/

[12] R. Maghami, H. Hizam, C. Gomes, A. Radzi, I. Rezadad, and S. Hajighorbani, “Power

loss due to soiling on solar panel: A review, renewable and sustainable energy reviews,”
Renewable and Sustainable Energy Reviews, vol. 59, pp. 1307–1316, 2016.

[13] S. X. Chen, H. B. Gooi, and M. Q. Wang, “Sizing of energy storage for microgrids,”
IEEE Transactions on Smart Grid, vol. 3, no. 1, 2012.

[14] P. J. Skoplaki, E., “On the temperature dependence of photovoltaic module electrical
performance: A review of efficiency/power correlations,” Solar Energy, vol. 83, pp.
614–624, 2009.

[15] “Utility API data,” UtilityAPI, 2018. [Online]. Available: https://utilityapi.com/

UC Berkeley, CE295
18

[16] G. James, D. Witten, T. Hastie, and R. Tibshirani, Tree-Based Methods, G. Casella,

F. S., and I. Olkin, Eds. New York, NY: Springer Science and Business Media, 2013.

[17] H. Zhao and F. Magoules, “A review on the prediction of building energy consumption,”
Renewable and Sustainable Energy Reviews, vol. 16, pp. 3588–3592, 2012.

Appendix
Variable Definitions in Optimization Program
The variables and their corresponding definitions are enumerated in the table below.

Variable Units Definition

GI (i), GE (i) [kW] Power imported or exported to/from the Grid
GI,max , GE,max [kW] Maximum grid import and export power
S(i) [kW] Power generated from solar PV
L(i) [kW] Load demand of the facility
Bd (i), Bc (i) [kW] Power discharged from/charged to flywheel storage
E(i) [kWh] Energy level of flywheel storage
C2,1 (i) [kW] Charging power of Level-2 EV charging station
Z2,1 (i) [kWh] Cumulative energy delivered to Level-2 EV charg-
ing station
Z2,1,min (i) [kWh] Minimum energy required for EVs at timestep i
η2 [-] Charging efficiency for Level-2 EV charging sta-
tions
C2,1,max [kW] Maximum charging capacity of Level-2 EV charg-
ing stations
cI (i) [$USD/kW] Time-of-use price of grid-imported power
rE (i) [$USD/kW] Time-of-sale revenue of exported power to grid
∆t [hr] Time step
ηc , ηd [-] Charging and discharging efficiencies for flywheel
storage
Emeas (i) [kWh] Measured initial flywheel storage energy level
Emax [kWh] Maximum energy stored in flywheel storage
Bmax [kW] Maximum charging capacity of flywheel storage
system

Building Load Demand Data Table

The following data table corresponds with 9 in section 1.3.3.

UC Berkeley, CE295
19

Figure 16: Table of Mean and Sample Standard Deviation of hourly energy demand.

Pacific Gas and Electric Residential Electricity Rates

Figure 17: Data for Summer Time-of-Use Residential Electricity Rates from PG&E

Figure 18: Data for Winter Time-of-Use Residential Electricity Rates from PG&E

Summary of the results obtained from the MCP

The results of the MCP are summarized as follows:

UC Berkeley, CE295
20

*Represents the cost of energy the microgrid would pay under the existing conditions. That
is no generation or storage.
** Represents the cost of energy if the microgrid had PV generation but no energy storage
device.
*** Represents the earnings possible with energy storage, PV generation and clairvoyance
of all demand.

Figure 19: Summary of the results obtained from the MCP

Further Analysis of Time Horizon Selection Implications

Since shorter time horizons mean that the program is unable to fully predict the future,
the MCP will choose to discharge all of the flywheel’s charge as soon as possible to make
profit. This is seen in Figure 20 as compared to the flywheel state-of-charge response in
Figure 21.

UC Berkeley, CE295
21

Figure 20: MCP operations simulation with only 3-hour time horizon

Figure 21: MCP operations simulation with full 12-hour time horizon

Acknowledgments
The team wishes to thank Professor Scott Moura for his guidance and insight throughout
the course of this semester and project. The group also wishes to thank Bertrand Travacca
for his ideas and discussions regarding alternative approaches for predictive modeling control.
Additionally, the team acknowledges the entire CE 295 class for adding value to the overall
curriculum of CE 295 through thoughtful and inquisitive questions and comments.

UC Berkeley, CE295
22

About the authors

Ramon Crespo M.S. Civil and Environmental Engineering: Energy, Civil Infrastructure
and Climate.

Ioanna Kavvada M.S. Civil and Environmental Engineering: Energy, Civil Infrastructure
and Climate.

Guo Jun Li M.S. Civil and Environmental Engineering: Energy, Civil Infrastructure and
Climate.

Dieter Smiley M.S. Civil and Environmental Engineering: Energy, Civil Infrastructure
and Climate.

UC Berkeley, CE295
23

UC Berkeley, CE295
24

Grid Optimization and Resiliency Study for the Puerto

Rican Electrical Grid
Dylan Kato, Anoush Razavian, Shane Gallagher, Dane de Wet

Abstract
In order to meet carbon emission targets in the fight to stop climate change, many
energy grids around the world are switching to clean energy sources. These sources help
make modern society more sustainable, but come with their own challenges. Wind and solar
power suffer from variability. Hydroelectric generators are geographically restricted, and
nuclear power plants are inflexible to rapid load changes. This study uses Puerto Rico as a
case study for designing a clean energy grid that balances these challenges to meet demand,
without compromising on energy prices, reliable power generation, or grid resiliency. First,
the minimal cost generating mix was found, while meeting the required physical constraints
for a variety of wind energy grid penetrations. This was done by formulating a linear program
to minimize total cost, subject to various demand and ramp rate constraints. From this, a
generating mix was selected to conduct a grid resiliency study of a basic grid network.

Introduction
Motivation and Background
Climate change has become one of the greatest challenges facing the world today, resulting
in an increased frequency of natural disasters and extreme weather events. These natural
events are particularly devastating in areas with vulnerable infrastructure, like Puerto Rico.
Recently, Hurricane Maria left much of Puerto Rico without power for over four months,
crippling daily life and risking the health of millions of people. In order to prevent disasters
like this in the future, it is important that power grids be designed for resilience to worsening
natural disasters while eliminating the emissions that are causing climate change in the first
place.
A large body of scientific work has helped show that greenhouse gas emissions are leading
to cascading negative effects that will greatly affect the Earth and the future of people around
the globe. Power generation is one of the primary causes of these emissions. Many grids
around the world would like to switch to cleaner energy sources, but need to address the
challenges associated with them. This project aims to help address that issue by designing a
modernized grid for Puerto Rico that will greatly reduce the environmental impact of power
generation while making it resilient to natural disasters.

UC Berkeley, CE295
25

Such resilience comes at a cost, since energy grid management can become cumbersome
when trying to balance low cost power production with resiliency during extreme weather
conditions. Control schemes need to be carefully thought out so that the grid can bal-
ance natural disaster resilience with needless inefficiencies created through over-fortification.
Additional challenges arise when designing an energy grid that has a high penetration of
renewable energy.
It is vital that we face the challenge of designing clean energy grids while maintain-
ing resilient and economic energy production. This project aims to address the challenges
associated with creating a grid that is resilient, economic, and clean.

Focus of this Study

For this study, we found the optimal power generation mix for a redesigned Puerto Rican
power grid to reliably meet demand using clean generating sources while minimizing cost.
The optimal generating mix was then used for a grid resiliency study to better understand
how to better design a modernized grid to withstand natural disasters like Hurricane Maria.

Literature review
Ferris of E&E news reported that the aftermath of Hurricane Maria in a months-long
blackout, despite the fact that the damage potential of the storm was similar to Hurricane
Sandy, which in comparison, left residents without power for at most several weeks. These
devastating impact on Puerto Rico caught the attention of officials, motivating them to find
solutions superior to the traditional grid network infrastructure. Most are in agreement
that microgrids can provide a reasonably compelling solution. Grid redesign does face some
bureaucratic obstacles—even though many companies have lent helping hands to implement
renewables, many have concerns about creating a microgrid system in Puerto Rico because
of the potential it has to undermine the monopoly the Puerto Rico Electric Power Authority
(PREPA) has over utilities. Robust and resilient redesign may require the aid of external
factors to become a reality.[1]
De Jonghe et al. in their 2011 paper on grid optimization with high renewable penetration
use many techniques that we wish to build upon for determining an optimal and resilient grid
design. De Jonghe incorporates renewables as a “negative demand”, shifting the variability
of renewables from the supply side to the demand side, resulting in a “net demand” to be
met by controllable energy sources. Ramp rates from traditional energy sources were then
binned into Base, Mid, Peak, and High Peak ramp rates. A minimum amount of flexibility in
ramp rate is necessary at any given time due to the variability of high renewable penetration.
This constraint can be softened by renewable curtailment, but is done at an opportunity cost
of not using renewable power. Additionally, energy storage was also considered for softening
this constraint. [2]
Egbue’s paper details how achieving resilience can be done through increasing system
flexibility and robustness, namely through introducing microgrids that combat the fragility
of the interconnected system. Installing microgrids is in essence the same as introducing
defensive islanding. Microgrids allow for distributed generation, which allows for power
to still be delivered via renewables on the microgrid in the case that power-demanding

UC Berkeley, CE295
26

infrastructure is cut off from the main source. The paper goes into discussion about the
challenges in microgrid operation, control, cybersecurity, and other barriers. It also discusses
their utility of microgrids in light of natural disasters, relevant to both the motivation of our
project and solution we plan to attempt.[3]
Panteli’s paper shows the need for resilient grid infrastructure to combat the damage
caused by extreme weather events, namely the aftermath of Hurricane Sandy, motivated this
paper. This is highly relevant to the state of Puerto Rico’s grid redesign, as the destruction
of Hurricane Maria was similar to, though greater in magitude than that of Hurricane Sandy
The paper proposes a grid resiliency approach known as defensive islanding, where the
objective is to maintain stability of resulting subsystems to reduce total losses across the
board. The general method the paper suggests is to begin by modelling the impact of
the weather event using a fragility curve, conducting risk assessment as suggested by the
procedure outlined in the paper that can be applied to any specific weather event, use the
defensive islanding algorithm which uses power flow data to split the system into islands
and isolate the vulnerable components and apply the appropriate islanding solution, and
finally, use the weather dependent failure probabilities obtained from fragility curves and
a uniformly distributed random number to determine which components will trip due to
the weather event. The paper applies this method by doing a case study for a simplified
transmission network in Great Britain. They use this analysis to conclude that as the
“Severity Risk Index” (SRI) they define increases, the more beneficial the application of the
defensive islanding becomes.[4]

Key contributions
One aspect all these references had in common was the suggestion of distributed gen-
eration, which can be achieved through microgrid operation and control. In our study of
resilience, we have thus far conducted testing to see how our simplistic macrogrid will func-
tion when branches are no longer functional. Some of these branch breaks effectively causing
the system to behave like a microgrid. This is also similar to the defensive islanding strat-
egy suggested by several of the references, as the isolated island operation will be key to
our findings. Some of these studies in the literature were conducted prior to the extreme
weather events in Puerto Rico, as microgrid operation seems to have been seen as a valuable
solution to improve resilience for a while, but in the face of disaster, is finally being taken
more seriously than ever before.
The De Jonghe paper contributed largely to the formulation of the optimization problem
used to determine the optimum energy mix. We built upon this paper by adjusting it to
nuclear, natural gas, and hydro power. It was adjusted to incorporate maximum hydroelec-
tric power, and physical constraints related to the ramp rate of each specific technology.
We ran the optimization for a spectrum of different wind penetrations, and calculated the
total emissions for the time period analyzed for each. These were then plotted to gain an
understanding of the influence wind has on optimal controllable energy mix on the grid.

UC Berkeley, CE295
27

Technical Description
Generation Mix Optimization
Part I: Problem Formulation

In order to find the optimal mix of generating technologies for the new Puerto Rican
grid, we formulated an optimization problem based on the work by De Jonghe [2]. This
formulation treats the variable generation by wind as a negative demand to create a net
demand (1). The capacity of the controllable generators (natural gas, nuclear, and hydro)
are then optimized to minimize the cost of the grid based on the fixed and variable costs
of each generation technology (2), (3), and (4). The constraints included a requirement for
generation to meet demand (5), as well as physical constraints related to the ramp rates (9
through 14) , must run capacity (7), and periodic maintenance factors (6). The capacity
of hydro power is also constrained due to limited availability of hydro power expansion (8).
The time based constraints were set for a large dataset of different demand and wind power
generation data. In this optimization formulation, wind power was not curtailed, as the data
likely already included curtailment by the Finish grid operator.

∀j ∈ J : N ET DEM AN Dj = DEM AN Dj − (W Pj ∗ W CAP ) (1)

∀i ∈ I : Fi = IN Vi + F OMi (2)

∀i ∈ I : Vi = F Ui + V OMi (3)
X X
minimize : Fi ∗ capi + Vi ∗ gij (4)
i ij
X
subject to : ∀j ∈ J : gij = N ET DEM AN Dj (5)
i

∀i ∈ I, j ∈ J : gij ≤ capi ∗ P Mi (6)

∀i ∈ I, j ∈ J : gij ≥ M Ri ∗ capi (7)

capHY DRO ≤ M AXHY DRO (8)

∀i ∈ I, j ∈ J : f lex upij ≤ RAM P Ci ∗ gij−1 + RAM P N Ci ∗ (capi − gij ) (9)

∀i ∈ I, j ∈ J : f lex upij ≤ capi − gij−1 (10)

∀i ∈ I, j ∈ J : f lex downij ≤ RAM P Ci ∗ gij−1 + RAM P N Ci ∗ (capi − gij−1 ) (11)

UC Berkeley, CE295
28

∀i ∈ I, j ∈ J : f lex downij ≤ gij−1 (12)

∀i ∈ I, j ∈ J : gij ≤ gij−1 + f lex upij (13)

∀i ∈ I, j ∈ J : gij ≥ gij−1 − f lex downij (14)

Part II: Calculating Parameters

Data for power generation in Puerto Rico was unavailable for this study, so we used data
which came from Finland in the early part of 2017 [5]. We normalized the demand data to
the maximum value for the whole dataset, and scaled it to the size of the Puerto Rican power
grid. The wind power data was also normalized to its maximum value, and then scaled for
the desired grid penetration. In order to get accurate numbers, data was obtained from a
large number of sources [6].
Table 1 outlines the data found to determine parameter values used in the grid opti-
mization. There are several things to note about the parameters used. For the proposed
grid, it is desired to use advanced nuclear reactor designs as opposed to conventional reactor
designs. The costs for these are not readily available, so several adjustments had to be made.
The investment cost for nuclear power was set to the twice the value of the claimed price
by General Electric, giving room for first-of-a-kind construction problems [7]. The ramp
rates came from a integrated resource plan for CAISO, using the values they use for long
term grid planning, as it was desired to minimize emissions while still meeting peak demand.
Base load natural gas would likely have a lower variable operation and maintenance cost,
although the investment, fixed O&M, and fuel costs are the same. The nuclear power was
required to ramp at no more than twenty percent of maximum power per hour. It also has a
must run capacity of eighty percent. This was chosen based on the values from CAISO, with
some flexibility added to reflect more modern reactor design capabilities [8]. Hydroelectric
power capacity was constrained to five times the current installed hydro capacity in Puerto
Rico, and given a must run capacity of 13 percent to reflect requirements for downstream
water supply. Solar power was not included, as the data from Finland reflected very strong
seasonal solar variations, due to the longitude of Finland, that would not be reflective of
Puerto Rico.

Table 1: Parameters values used in grid optimization

Type Invest Fixed O&M Var. O&M Fuel Must Run Ramp
%Pmax
Units (k$/MW) (k$/MW) ($/MWh) ($/MWh) % Pmax hour
Nuclear 4000 [7] 84.5 [9] 0.6 [9] 8.4 [10] 80 % [8] 20% [8]
Nat.Gas 1200 [9] 14 [11] 6.8 [11] 11.825[11] 0 % [8] 330% [8]
Hydro 3000 [9] 29.23 [9] 3.16 [9] 0 [9] 13 % [8] 100% [8]

UC Berkeley, CE295
29

Part III: Grid Optimization Results

Using this optimization formulation previously described, we found the optimal mix of
controllable generators, specifically nuclear, hydro, and natural gas for a range of different
wind penetrations. We ran this for a wind penetration from zero to sixty percent over a data
set of 10 days See Figures 1, 2, 3.

Figure 1: Effect of low penetration of Wind generation on the overall generation for 10 day
period

UC Berkeley, CE295
30

Figure 2: Effect of moderate penetration of Wind generation on the overall generation for
10 day period

Figure 3: Effect of high penetration of Wind generation on the overall generation for 10 day
period

UC Berkeley, CE295
31

There are several important conclusions from the running the optimization problem for
a variety of data sets with a variety of parameters. The first is that the optimal result
depends strongly on the length of the data set. Short time periods usually don’t feature the
bounding ramp rates that long time periods have. Very large ramp rate requirements for a
short time period have a big impact on the final result. This shows why it may be beneficial
to have distributed energy storage to reduce variability. Initially, the optimal result was a
very large amount of hydro with little of anything else. This was due to the low costs with
high flexibility to allow hydro to load follow. A maximum capacity was added, as there are
large geographic restrictions on the amount of hydro that can be installed. There were also
some interesting trends to note with varying wind penetration. As shown in Figure 4, as wind
penetration increases, the required natural gas also increased to allow for more flexibility in
the grid. With increased wind penetration, more flexible generation is required and therefore
nuclear is not suitable for a grid with high penetrations of wind power due to its must run and
ramp rate constraints. It is also evident that with increased wind generation, CO2 emissions
actually increase substantially due to the increased reliance on natural gas power plants.
This was without wind curtailment, importing or exporting power, distributed storage, or
a significant penetration of wind generation. Further work will likely address some of these
challenges.

Figure 4: Optimal Generating Mix for Various Wind Penetrations, with Corresponding
Emissions

UC Berkeley, CE295
32

These results were used to determine the energy mix to be used for the resilience study
discussed later in this paper. A wind penetration of 30 percent was chosen for the grid
resiliency study. This keeps emissions relatively low while incorporating a more diverse grid.

Resiliency Study
Part I: Obtaining Grid Network Model
Our base constraints of the report are borrowed from ”HW3: Optimal Economic Dispatch
in Distribution Feeders with Renewables.” These constraints deal with the power in and
power out of nodes, as well as voltage drops and minimum voltage. Our study focuses on
adapting this framework to optimize for resilience in the face of natural disasters. This
involves changing the objective function to account for total power delivered rather than
cost.

Part II: Recasting Objective Function for Resilience

Our objective function measures our desire to meet load demand without incentivising
meeting negative demand and is given by:
X Q Q
P P
f (x) = max(ldemanded − li,consumed , 0) + max(ldemanded − li,consumed , 0) (15)
i

Using an epigraphic reformulation, in order to recast the objective function which contains
a ”max” into a familiar form, we introduce intermediary variables tPj and tQ j which represent
the ”Demand Gap” or the difference in power demanded and power delivered. The new
objective function and constraints incorporating these variables are:
12
X
f (x) = tQ P
j + tj (16)
j=0

tQ P
j ≥ 0, tj ≥ 0 ∀ j ∈ N odes (17)

tQ Q Q
j ≥ lj,demanded − lj,consumed ∀ j ∈ N odes (18)

P
tPj ≥ lj,demanded P
− lj,consumed ∀ j ∈ N odes (19)

Q
P
Next, we substitute lj,consumed and lj,consumed for ljP and ljQ from the optimization in
Homework 3 to get new power constraints:

X
P
Pij = (li,consumed − pi ) + rij Lij + Aij Pij ∀ j ∈ N odes i = ρ(j) (20)
k∈N odes

Q
X
Qij = (li,consumed − qi ) + rij Lij + Aij Qij ∀ j ∈ N odes i = ρ(j) (21)
k∈N odes

UC Berkeley, CE295
33

Part III: Adding Constraints for Node Disconnections

Here, we incorporated the necessary constraints to account for power line failure. We do
so by detaching node j from its parent node i so that the detached node now has one unit
voltage (allowing for separate networks), and no power or current flows between the node
and its parent. The constraints are as follows:

Pij = 0, Lij = 0 ∀ j = jdetached and i = ρ(j) (22)

Part IV: Accounting for Renewable Uncertainty

In order to account for variability in solar and wind generation, we added constraints to
make our optimization robust to the stochasticity inherent in each renewable generator:

ā ∗ [si , σi,1 , σi,2 ] + kE ∗ [si , σi,1 , σi,2 ]k2 ≤ 0 (23)

σi,1 , σi,2 ∈ [0, 1] (24)

Here, ā and E contain information on the mean and covariance of the power generation.
σi,1 and σi,2 are decision variables that are the portion of energy used from individual solar
panels. Figure 5 shows the change in overall power delivery after including the robustness
constraint and then increasing the variance.

Figure 5: Comparing Power Delivered for Different Renewable Generation Variances

UC Berkeley, CE295
34

Part V: Prioritizing Nodal Power Delivery

We considered that it is more important to deliver power to certain nodes. For example,
if a particular node were a grocery store, hospital, or other priority energy location, that
node should take precedence over nodes that contain, for instance, recreational centers (sorry,
Disneyland). We were able to account for this by adjusting our objective function to include
the following ”priority” term which weighs the Demand Gap by coefficients that are set by
us to be larger if a node’s power delivery is more important:
12
X
maximize: P riorityj ∗ (tQ P
j + tj ) (25)
j=0

Part VI: Final Integration and Results

Finally, after integrating all of our new constraints into the optimization formulation we
obtained the following optimization problem:
12
X
maximize: P riorityj ∗ (tPj + tQ
j ) (26)
j=0

subject to: tQ P
j ≥ 0, tj ≥ 0 ∀ j ∈ N odes (27)

tQ Q Q
j > lj,demanded − lj,consumed ∀ j ∈ N odes (28)

tPj > lj,demanded

P P
− lj,consumed ∀ j ∈ N odes (29)
X
P
Pij = (li,consumed − pi ) + rij Lij + Aij Pij ∀ j ∈ N odes i = ρ(j) (30)
k∈N odes

Q
X
Qij = (li,consumed − qi ) + rij Lij + Aij Qij ∀ j ∈ N odes i = ρ(j) (31)
k∈N odes

P Q
pj ≥ 0, qj ≥ 0, li,demanded ≥ 0, li,demanded ≥ 0 ∀ j ∈ N odes (32)

kpj , qj k2 = sj ∀ j ∈ N odes (33)

sj ≤ sj,max (34)

2 2
vmin ≤ Vj ≤ vmax (35)

2
Lij ≤ Iij,max (36)

ā ∗ [si , σi,1 , σi,2 ] + kE ∗ [si , σi,1 , σi,2 ]k2 ≤ 0 (37)

UC Berkeley, CE295
35

σi,1 , σi,2 ∈ [0, 1] (38)

Pij = 0, Lij = 0 ∀ j = jdetached and i = ρ(j) (39)

After implementing in Matlab, the total percentage of power delivered was calculated as
the geometric mean of the fraction of active and reactive power delivered. We calculated
these percentages using the following equation:

P
i=nodes Desired − Delivered
F ractionDelivered = 1 − P (40)
i=nodes Delivered

The following figure shows visually how our new node network could look different from
the network in Homework 3. We changed the energy sources to also include nuclear, wind,
and hydro and their respective optimal generation determined from the Generation Mix
Optimization. We also treated Node 12 like a hospital by giving it priority in the objective
function.

Figure 6: Homework 3 Network Adjusted

We wanted to see how much power would be able to be supplied to the entire network as a
result of cutting off each node from its parent node. We used the F ractionDelivered metric
to be able to easily visualize this. As you can see, the worst case scenarios (scenarios where

UC Berkeley, CE295
36

the network is far from providing adequate power to the entire network) involve breaking off
Node 6, Node 9, and Node 10 from their respective parent node.

Figure 7: Comparing Power Delivered

We then considered what would happen in the disaster scenario where the line between
Node 6 and its parent node became disconnected. The following figure depicts what the
corresponding satisfied power demand would be at each node. As you can see, the entirety
of Node 12’s power demand is satisfied since it was prioritized in the objective function and
the other connected nodes in the subnetwork would suffer. In order to prevent this disaster
scenario, the redesign would involve adjusting generator placement in such a way that one
renewable generator would not have to service 7 nodes, even in the worst case.

UC Berkeley, CE295
37

Figure 8: Comparing Power Delivered

Discussion
This kind of work is a step toward envisioning energy grids that are sustainable and also
resilient. Through this work, we have developed a methodology that one can use to focus on
the important factors of grid resiliency. Puerto Rico can use a model like ours to learn more
about their grid and its shortcomings, utilizing this information to defend against insufficient
power delivery. This information is crucial to grid redesign.
Power grids are complicated systems, and clean energy generation is inherently limited
by variability and availability. One of the best ways to solve this problem is through system-
atically determining the best implementation of power distribution. With limited resources
in system settings, it is essential to use optimization to make the best use of the resources
we have.
Through our work, we created a framework for assessing resiliency of a network, but
haven’t answered several key questions. For one, we only used our framework for one network
layout. This could be expanded by applying the resiliency study to several network layouts to
find the optimal network layout for resiliency. This future work would also involve developing
a method/metric for comparing networks in their performance under disasters. Future teams
might also think about incorporating constraints for failure of grid components as a result of

UC Berkeley, CE295
38

increased power sent to some nodes. As it stands, we allow more power to be delivered than
is demanded which could pose a problem if such power surges lead to component failure.
Another future work could be incorporating the monetary value of resiliency and solving the
original minimum-cost-optimization but with a multi objective function that accounts for
generation cost and the cost of grid vulnerability. A final addition that may be interesting is
looking at grid resiliency after incorporating energy storage. This would open a new problem
of incorporating energy storage that would require the use of optimal control.

Summary
In this study, an optimal energy generation mix for Puerto Rico was determined by min-
imizing the cost of generation given a variety of physical constraints such as ramp rates and
various levels of penetration of intermittent renewable energy sources. The results showed
that increasing the amount of wind energy penetration increases the required amount of
natural gas peaker power plants needed to meet demand, and therefore increases greenhouse
gas emissions. The need for peaker plants is due to the slow ramp rates of base load nuclear
power A generation mix was selected to conduct a grid resiliency study of a grid network.
Natural disasters were simulated by severing transmission between grid network nodes. The
results of resiliency study show that the implementation of smart grid control logic such as
prioritizing certain nodes (such as hospitals) in an emergency scenario can be beneficial.

References
[1] D. Ferris and P. Behr, “Microgrids could save puerto rico. but first, a fight.” 2017.

[2] C. De Jonghe, “Determining optimal electricity technology mix,” Applied Energy,

vol. 88, pp. 2231–2238, 2011.

[3] D. N. O. Egbue and P. Peterson, “2016 international conference on smart grid and clean
energy technologies,” 2016.

[4] P. M. M. Panteli, D.N. Trakas and N. Hatziargyriou, “Boosting the power grid resilience
to extreme weather events using defensive islanding,” IEEE Transactions on Smart
Grid, vol. 7, no. 6, pp. 2913–2922, Nov. 2016.

[5] Fingrid. [Online]. Available: https://data.fingrid.fi/open-data-forms/search/en/

?variable id=75

[6] O. P. S. D. sources. Wind and solar power time series. [Online]. Available: https:
//open-power-system-data.org/data-sources#8 Wind and solar power time series

[7] C. Surran, “Ge hitachi touts development of smaller, cheaper nuclear reactor,” Seeking
Alpha, Apr. 20, 2018.

[8] Caiso integrated resource plan 2017: Inputs and assumptions.

UC Berkeley, CE295
39

[9] Transparent cost database. [Online]. Available: https://openei.org/apps/TCDB/

#blank
[10] Average power plant operating expenses for major u.s. investor-owned electric utilities,
2006 through 2016. [Online]. Available: https://www.eia.gov/electricity/annual/html/
epa 08 04.html
[11] E. Commission, “Gcost estimates for thermal peaking plant,” June 2008.

Acknowledgments
Thank you to Dr. Scott Moura for his guidance and mentorship throughout this project.

About the authors

Dane de Wet and Shane Gallagher: Two of the group members, Dane de Wet and
Shane Gallagher, are Ph.D. students in the Department of Nuclear Engineering. Their
work involves studying thermal hydraulics in advanced nuclear reactors, which address
most of the concerns associated with old nuclear power generation. Advanced reactors
are being designed for passive safety, adaptability, and economic power generation.
Their research group has been designing a reactor called the fluoride salt-cooled high
temperature reactor (FHR). This type of reactor operates at much higher temperatures,
and therefore higher efficiencies, while being able to operate at near atmospheric pres-
sures in the primary loop, providing enhanced safety margins. For them, the project
will further the understanding of the role advanced nuclear power can play in providing
a resilient, economic, and clean energy grid.
Dylan Kato and Anoush Razavian: The other two students are undergraduates in the
CEE department and are primarily interested in making smarter energy systems. Dylan
has taken 2 previous courses in optimization and another in controls and is excited to
apply his previous knowledge as well as the knowledge gained in this course to tackle a
real-world problem. Dylan plans to start his Ph.D. in the fall and hopes the knowledge
gained from this project will strengthen his ability to model infrastructural systems.
Anoush is passionate about clean energy and smart city development. She hopes that
by taking this course, she can learn valuable skills that she can apply to have an
influential career in fields like clean-tech, microgrid management, and civil systems
engineering. Anoush will be working in Kimley-Horn’s Intelligent Transportation Sys-
tems group this fall, viewing it as a step toward being a front-runner in creating smarter
and more sustainable infrastructure. She hopes to ultimately obtain a Masters in Civil
Systems. Anoush has also taken one previous course in optimization and another in
cyber-physical systems.
Together, Anoush and Dylan built a Cyber-Physical system for residential-scale energy
monitoring and control that collects usage at the breaker level and actuates at the outlet
level.

UC Berkeley, CE295
40

UC Berkeley, CE295
41

Robust Optimal Sizing and Operation of a Microgrid

with Electric Vehicle Charging and Renewable Energy
Generation
Armando A. Domingos, Marc R. Hutton, Aaron Jagtianey, and Soomin Woo

Abstract
Electric vehicles (EVs) are increasing in popularity and access to charging infrastructure
will play an important role in their widespread adoption. Additional investment in exist-
ing commercial facilities will be required and pairing vehicle charging infrastructure with
microgrids (MGs) may help to mitigate energy consumption requirements from EVs. This
study developed a model for the economic optimization of a grid-connected commercial MG
that includes renewable energy generation, on-site energy storage, and EV charging with
Level 2 and 3 capabilities. A robust optimization method that considers the stochastic na-
ture of energy demanded and supplied was developed while satisfying constraints imposed
by the facility. Results from a Mixed-Integer Linear Program (MILP) and Mixed-Integer
Second-Order Cone Program (MISOCP) were evaluated to determine primary drivers for
infrastructure sizing and to evaluate the robustness of the optimization model for various
uncertainty levels of input data. The MISOCP was able to identify infeasible scenarios and
was able to plan the optimal scenario while being robust to all stochastic variables, making
it the preferred choice for planning purposes over the MILP.

1 Introduction
1.1 Motivation and Background
Electric Vehicles and Microgrids
With new electric vehicle (EV) registrations reaching record numbers in recent years,
the EV market is expected to reach forty to seventy million vehicles by 2025 [1]. To meet
increasing demand, additional charging infrastructure will be required in most commercial
facilities, where existing charging infrastructure is often limited to a small proportion of over-
all parking. Increasing the number of charging stations and accommodating a large number
of EVs charging simultaneously could also multiply the energy load requirements of com-
mercial facilities, which in turn could threaten the stability of local electrical infrastructure
and require expensive upgrades to local and grid level equipment [2].

UC Berkeley, CE295
42

Additionally, due to long charging times for current charging technology, service reliability
is crucial to alleviate EV owner range anxiety and to provide adequate charge for commuting
needs. Methods to optimize service will be critical to satisfy EV owners, who will expect
charging access at all times, even during periods of peak demand. While charging demand
could be met by deploying additional charging stations, minimizing system cost will be an
important consideration for facility owners when determining the number of chargers to
install.
To meet the increased energy demand from EV charging, facility managers could face
additional cost for upgrades to grid energy import infrastructure, or expensive tariffs with
peak demand charges or time-of-use (TOU) rates. Installing distributed generation (DG)
in the form of a microgrid (MG) that includes renewable energy generation (REG) and
storage allows for facility managers to meet increased demand locally. So called “smart-
grid” infrastructure has utilized local MGs to meet increased demand from a variety of
sources, including EVs, and to reduce peak system demand [3]. In addition to reducing
peak demand, MG components such as REG and energy storage systems (ESS) allow for
additional revenue streams for facility managers when combined with net energy metering
(NEM) schemes or community choice aggregation (CCA) arrangements.
Therefore, minimizing the cost of charging infrastructure upgrades and operational energy
demand while meeting high level of service is critical to commercial facility managers. Com-
bining traditional MG resources such as REG and ESS with EV charging provides synergistic
opportunities that could mitigate operational expenses from energy demand and reduce costs
for facility managers. Introducing revenue streams from REG export to the grid and EV
charging can help offset initial investments and promote additional EV charging facilities.

Stochastic Elements
To determine optimal sizing, the stochastic nature of system inputs must be considered.
Renewable energy is stochastic in monthly, diurnal, hourly (and shorter) timescales, due to
seasonal changes and local weather, but reliable on long time scales when stochastic elements
are aggregated into long-term trends. Similarly, building energy consumption is also reliable
on long time scales while maintaining stochasticity on shorter timescales, primarily due to
lighting and heating, ventilation and air-conditioning (HVAC) loads that can change due to
weather and individual preferences [4].
EV charging demand enjoins energy consumption in these traits. Because of individ-
ual differences between driver schedule and travel distance, battery state-of-charge (SOC),
and therefore overall charging times exhibit stochastic elements on shorter timescales while
maintaining reliability in the long-term [5]. Unlike grid connected building energy loads,
additional energy is not easily imported into the vehicle once the EV has left the charging
point, increasing the importance of high-level reliability in charging service.

Project Goal
Our project aim is to develop a robust optimization model of a mixed-integer second-order
cone program (MISOCP) from a white paper that integrates a MG and EV charging [6]. The
model should incorporate stochastic REG, stochastic building energy demand, and stochas-

UC Berkeley, CE295
43

tic EV charging demand while minimizing capital and operational cost and determining the
optimal number of chargers required to meet charging demand. Linear and non-linear meth-
ods of optimization should be explored, and both feasible and infeasible solutions should be
determined for optimization. Robustness should be the defining model feature, and bound-
ary cases and corner conditions should be explored to resolve any outstanding concerns in
this area.

1.2 Literature Review

Recently, there has been significant in-
terest in optimal planning for integrated EV
charging systems. The most comprehensive
systems include EV charging stations, non-
EV site loads, vehicle-to-grid (V2G) connec-
tions, REG such as photovoltaic (PV) pan-
els, and ESS. For example, a smart home
equipped with an EV charger, smart ap-
pliances, DG, and ESS was modeled us-
ing a mixed-integer linear program (MILP)
with the objective of minimizing operational
cost under different demand response (DR)
strategies [7]. The smart home system was Figure 1: Spider plot of weighted values in
reevaluated with respect to optimal PV and several categories for a typical sample of the
ESS sizing in [8], using a MILP with the ob- reviewed literature. The primary method that
jective function of minimizing both capital has been used to minimize cost for similar sys-
costs and operational costs. A smart home tems is a mixed integer linear program with
with bi-directional EV and ESS was consid- a combination of renewable energy, storage,
ered in [9], using a MILP to minimize the and vehicle charging, but none also included
cost of electricity. the minimization of capital and operational
The most common method for address- expenses. Furthermore, none considered ro-
ing optimal planning for integrated EV bustness as a model feature
charging systems is to utilize MILP in vary-
ing scenarios, from minimizing capital ex-
penses or operational expenses. Optimal operation is explored in [10], using a robust two-
stage stochastic problem which accounts for solution uncertainty and model uncertainty.
The problem is formulated as a MILP with the objective of minimizing operational cost
given stochastic power trading with the grid, stochastic EV charging demand, but does not
consider stochastic REG.
While incorporating stochastic elements into optimization problems is well understood,
few studies also incorporate robustness as a model feature, accounting for all stochastic
variables present in a MG with EV charging. As seen in Figure 1, we see that robustness as
a model feature is not well represented in existing literature, even as renewable energy and
EV charging plays a prominent role. It is likely that this is due to the difficulty of solving
SOCP optimization problems, which are a method for addressing robustness with stochastic
variables.

UC Berkeley, CE295
44

1.3 Focus of this Study

This study aims to develop a robust optimization model from a MISOCP to minimize
capital and operational expenses of a microgrid incorporating building load renewable en-
ergy, energy storage, and EV charging demand while accounting for the stochastic variables
associated with these components. Robustness across all stochastic variables is considered
the defining model feature.

2 Technical Description
2.1 Optimization Model
The model is formulated to optimize the sizing and energy management of a commercial
facility with REG, ESS, and EV chargers. The model contains two components: the first
contains the capital costs for REG, energy storage, and EV chargers. The second contains
operating costs for how the facility allocates the energy supply throughout the MG network.
Constraints for the model include physical system parameters such as grid import and
export limits, energy storage dynamics, EV charging dynamics, renewable energy scale limits,
battery scale limits, and the number of EV chargers.

Mixed Integer Linear Program

To start, a deterministic optimization model was formulated to verify the objective func-
tion variables, parameters, and constraints. After formulating the SOCP model, changes
were first tested with the deterministic model, allowing for increased flexibility due rela-
tively fast computational times. The MILP was also used in comparison to the robust
optimization model to determine how chance constraints influenced optimal sizing.

Objective Function
Our objective equation is to minimize capital cost, energy cost, demand charge, and
export revenue as follows:

min cb · b + cs · s + mc2 · n2 + mc3 · n3 (capital cost)

N
X
+ cI (k) · ∆t · [L̄(k) + Bc (k) + C2 (k) + C3 (k) + GE (k) (energy cost)
k=0
− s · S̄(k) + Sc (k) − Bd (k)]
+ cd Gd (demand charge)
N
X
− ∆t · [rE (k) · GE (k) + r2 (k) · C2 (k) + r3 (k) · C3 (k)] (export revenue) (1)
k=0

UC Berkeley, CE295
45

Constraints
Subject to these constraints:

0 ≤ Sc (k) (PV curtailment) (2)

1
E(k + 1) = E(k) + [ηc · Bc (k) − · Bd (k)] · ∆t (ESS dynamics) (3)
ηd
E(0) = b · Emax · 0.10 (ESS initial condition) (4)
E(24) ≥ b · Emax · 0.10 (ESS final condition) (5)
−1 0
C2 (24) ≤ Fzmin,2(24) · (αz ) (level 2 charger limit) (6)
−1 0
C3 (24) ≤ Fzmin,3(24) · (αz ) (level 3 charger limit) (7)
0 ≤ E(k) ≤ b · Emax (ESS energy limits) (8)
0 ≤ Bc (k) ≤ b · Bmax (ESS power limits) (9)
0 ≤ Bd (k) ≤ b · Bmax (ESS power limits) (10)
Z2 (k + 1) = Z2 (k) + [η · C2 (k)]∆t (delivered EV energy dynamics) (11)
Z3 (k + 1) = Z3 (k) + [η · C3 (k)]∆t (delivered EV energy dynamics) (12)
Z2 (0) = 0 (initial EV energy) (13)
Z3 (0) = 0 (initial EV energy) (14)
−1
Z2 (k) ≥ Fzmin,2(k) · (αz ) (EV energy lower bound (15)
−1
Z3 (k) ≥ Fzmin,3(k) · (αz ) (EV mobility constraint) (16)
0 ≤ C2 (k) ≤ n2 · C2,max (EVSE power limits) (17)
0 ≤ C3 (k) ≤ n3 · C3,max (EVSE power limits) (18)
q
σs2 (k) · s2 + σL2 (k) ≤ (minimum import power limit) (19)
1
· [L̄(k) − s · S̄(k)
Φ−1 (αG1 )
+Bc (k) + C2 (k) + C3 (k)
+GE (k) + Sc (k) − Bd (k)]
or equivalently
||s · σs (k) · σL (k)||2 ≤ RHS (minimum import power limit)
||s · σs (k) · σL (k)||2 ≤ (maximum import power limit) (20)
1
· [ s · S̄(k) − L̄(k) + GI,max
Φ−1 (αG2 )
−Bc (k) − C2 (k) − C3 (k)
−GE (k) − Sc (k) + Bd (k)]
0 ≤ GE (k) ≤ GE,max (grid export power limit) (21)
GI (k) ≤ GD (demand charge) (22)
or equivalently
[L̄(k) + Bc (k) + C2 (k) + C3 (k) + GE (k) (demand charge)

UC Berkeley, CE295
46

−s · S̄(k) + Sc (k) − Bd (k)] ≤ GD

smin ≤ s ≤ smax (solar scale limit) (23)
bmin ≤ b ≤ bmax battery scale limit) (24)
n2 ∈ Z (number of level 2 chargers) (25)
n3 ∈ Z (number of level 3 chargers) (26)
0 ≤ n2 ≤ η2,max (max number of level 2 chargers) (27)
0 ≤ n3 ≤ η3,max (max number of level 3 chargers) (28)

Variables
Primary optimization variables are b, s, n2 , n3 , Bc (k), C2 (k), C3 (k), GE (k), Sc (k), Bd (k),
Gd . Optimization variables in equations (1) through (28) are highlighted in blue. A full list
of optimization variables are included in Appendix A.

Data
Solar resource data were obtained from NREL’s PVWatts calculator, for Oakland Inter-
national Airport and mean values and standard deviations were calculated from hourly data
from 2016 [11]. Solar PV pricing was determined from national and regional averages from
2017 data [12]. Daily pricing was calculated from per kWh averages as shown in equation
(32).
EV related data were obtained from a BMW Group project document that included the
charging start and stop times with SOC levels for a limited fleet of BMW battery EVs and
plug-in hybrid EVs [13]. Mean and standard deviations were calculated from this data set.
Building load data were also obtained from the BMW Group project document. Building
load data included hourly demand in kWh for homes in the San Francisco Bay area [13].
Mean and standard deviations were calculated from this data set.
Capital costs and efficiency for EV charging infrastructure were determined from industry
standards, and conversion to daily costs were calculated [14] [15]. Battery costs were also
determined from industry standards, and daily price was calculated from per kWh averages
as shown in equation (32) [15].
A full list of parameters and values are included in Appendix B.

2.2 Refinement of the Model

Mixed Integer Programming
The general form of a mixed-integer linear program is as shown in equation 29 as follows:

minimize fT x
x
subject to: Ax ≤ b
(29)
Aeq x = beq
x∈Z

UC Berkeley, CE295
47

Integer programs are inherently non-convex, since they impose discrete values onto the
decision variables of the system. Discrete problems are non-convex because a convex set
cannot be formed. Solving a mixed-integer program as such is not possible with standard
CVX software due to its non-convexity. In the case of this project, the discrete states are
the number of Level 2 (L2) and Level 3 (L3) chargers installed in the commercial facility. To
solve the mixed integer program without using mixed integer solvers, the LP and SOCP were
solved iteratively for combinations of L2 and L3 chargers. The resulting total cost at optimal
were compared from each iteration to determine the number of chargers that resulted in the
least cost. Equation (30) illustrates the concept of converting the mixed-integer programs
into programs solvable in CVX, with n2 and n3 being the number of L2 and L3 chargers,
respectively. To limit computation times, the maximum number of L2 and L3 chargers were
set to 4, with a range from 0 to 4.

minimize fT x
x
subject to: Ax ≤ b
Aeq x = beq (30)
x∈R
n2 , n3 ∈ {1, 2, 3, 4}

Calculating Daily Costs

 t
1
c= P 1+r
365 × t
Where: c = Daily cost
(31)
t = Lifetime of equipment
P = Principal
r = Annual investment rate
The general formula for approximating daily costs taking into account annual interest
rate is shown in equation (31).
1
c= P (32)
365 × t
For simplicity, the interest rate was not considered in this project. This reduces equation
(31) into equation (32), which was used to calculate the per-day costs of installed equipment.
Refer to Appendix B for a list of assumed costs.

MILP
The Mixed Integer Linear Program was developed before the Mixed-Integer Second Or-
der Cone Program to verify the validity of the dynamics and parameters to be used in the
optimization program. The MILP uses the mean values of building load, solar generation,

UC Berkeley, CE295
48

and minimum energy delivered to the EVs (L(k), S(k), Z min (k), respectively) as determinis-
tic inputs from the data collected instead of treating them as random variables. Due to the
lack of chance constraints, equations (6)-(7), (15)-(16), and (19)-(20) were not considered in
the MILP.

SOCP
The Second Order Cone Program considered chance constraints on the building load,
solar generation, and minimum energy delivered to the EVs (L(k), S(k), Z min (k)), which
distinguishes it from the MILP. It is assumed that the hourly insolation and building load
have independent Gaussian distributions, with an average (L̄(k), S̄(k)) and standard devia-
tion calculated from the obtained data sources (σL (k), σS (k)).

3 Results
Utilizing iterations of multiple SOCP problems, our optimal number of L2 and L3 chargers
are 3 and 1, respectively, considering the uncertainty described by solar generation, building
energy demand, and EV charging demand as shown in Figure 2. To determine robustness,
we refined the model to reflect the realistic situations of operating a microgrid. These
refinements include corner cases such as simultaneous battery charging and discharging, and
boundary conditions including excessive discharging of battery at the final time step due to
finite time horizon of optimization, and over-speculation of vehicle charging demand.

Figure 2: Optimal Size of Level 2 and Level 3 EV Chargers with Real Data. L2 and L3
reflect the number of Level 2 and Level 3 chargers, respectively. Total system cost is $70 per
day.

UC Berkeley, CE295
49

3.1 Robust Optimal Sizing of the Microgrid

Figure 2 shows the result of SOCPs solved iteratively for combinations of L2 and L3
chargers. With the given data and assuming that L2 or L3 chargers will be used only for L2
and L3 charging demand, respectively, there are only three feasible combinations of sixteen
possible combinations. Total cost of $0 in Figure 2 represents infeasible combinations of
the solution, where the objective value is unbounded. Among the feasible combinations, the
optimal combination is the lowest total system cost per day, with the optimal combination
three L2 chargers and one L3 charger. Total system cost includes capital cost such as
solar panels, batteries, and EV chargers as well as operational costs such as building energy
demand. Total system cost of $70 per day results in $2.33 per day per unit in the 30-unit
commercial residential property, including charging costs for all EVs. At optimum, the solar
scale and battery scale factor were 20 and 9.6 units, respectively. This results in a solar PV
system size of 20 kW and battery size of 9.6 kW.

3.2 Robust Optimal Operation of the Microgrid

Figure 3: Optimal operation of a microgrid with EV charging using real data over 24 hours.

Figure 3 shows the result of the operation strategy at optimal size of the solar panels,
battery scales, and EV chargers. Note that the TOU price of electricity import from the grid
is artificially designed in this result. To simulate time-of-use tariffs, from 4PM to 10PM the
grid imported electricity rate is increased threefold to observe its impact on the operation
strategy. To minimize cost, the model results show that there is a dramatic decrease of grid
import power during higher electricity rates, with demand met by battery discharge. Figure
3 also shows that the battery does not exhibit simultaneous charging and discharging. Refer
to Section 3.5 for a detailed analysis of the behavior of battery discharge and the decrease
of grid import at the final hour. Refer to Section 3.6 for a detailed analysis of EV charging
behavior during peak electricity rates.

3.3 Robustness to Uncertainty in Energy Generation and Demand

UC Berkeley, CE295
50

The robustness of the proposed optimization problem to

uncertainty of energy generation and demand is shown by com-
parison to a optimization problem that does not consider the
uncertainty in these variables. Results show that the robust
optimization model recommends installation of three L2 charg-
ers, whereas the non-robust optimization problem recommends
only one. Referring to Figure 4, it can be seen that if the charg-
ing demand of L2 occurs one standard deviation larger than the
expected charging load, the non-robust optimization solution
of only one charger cannot meet the charging load. However
since the proposed optimization problem considers the uncer-
tainty of the load prediction, the maximum capacity of the L2
Figure 4: Robustness to
charging with three L2 chargers from the robust model can
Uncertainty in EV Charg-
facilitate the EV charging need.
ing Demand

3.4 Model Refinement

Refinement of the optimization model in operating condi-
tions was necessary to maintain robustness and to account for realistic situations that may
occur, such as corner cases and boundary conditions. The following scenarios as depicted in
Figure 2 and Figure 3 were induced with hypothetical data.

Simultaneous Battery Charge and Discharge

Referring to the constraints of
the optimization problem as pro-
posed, it was determined that a
”battery wear” or ”battery health”
constraint was required. In real
world scenarios it is undesirable to
induce simultaneous charging and
discharging as it degrades battery
life. However, while implementing
various scenarios of data input, i.e.
different scales of solar energy gen-
eration, EV charging demand, and
building energy demand, it was de-
termined that it is mathematically
possible to produce simultaneous
charging and discharging of bat-
Figure 5: A Example Corner Case of Simultaneous tery. This result is shown in Figure
Charging and Discharging of Battery. Blue regions indi- 5, where we observe both simulta-
cate desired behavior while red shaded regions indicate neous charging and discharging af-
undesirable battery behavior, i.e., simultaneous charg- ter 8AM.
ing and discharging.

UC Berkeley, CE295
51

To induce this behavior, the to-

tal energy demand in the MG is sig-
nificantly and unrealistically smaller compared to solar generation, while energy export to
the grid is maximized from solar energy generation. As the original problem was not con-
strained to consider the battery’s internal resistance, the optimization problem attempted
to expend surplus energy generation. Since this is a corner case derived from unrealistic
data, the solution proposed to this issue is to observe and analyze the input data to the
optimization problem and adjust the optimization scenario to prevent such a result.

Excessive Battery Discharge on Final Time-step

As the optimization problem considers the operation of the MG in a finite time horizon,
the model tries to maximize the profit without consideration to the future after the time
horizon. Therefore, the optimal result may produce a scenario in which the battery discharges
either to the grid or to EV chargers that increase the revenue at the final time step. In Figure
6, the battery discharges largely at the final time step, and the battery size at optimal is
the maximum level. This indicates that the model tries to maximize profit by use of the
discharge of battery unrealistically. As this result is short-sighted and unsustainable for the
operation of the MG, an additional constraint on the final SOC of the battery is necessary
to prohibit this behavior. This constraint is described in the equation (5), limiting battery
SOC at the final state to greater than or equal to 10 percent.
In Figure 3, results show that
final hour discharge of the bat-
tery still occurs and grid import
decreases even after implement-
ing the final SOC level constraint.
However, the scale of discharge
is smaller than that of Figure 6,
since the battery size at optimum
is much more realistic in Figure 3.

Over-speculation of Vehicle
Charging Demand
Due to conversion of the MIS-
OCP to multiple SOCPs another
boundary condition required re-
finement. Though the reduction in
Figure 6: Comparison between Bounded and Un- complexity of the problem is a large
bounded Final SOC of the Battery benefit to computation, iterating
the problem over a given number
of chargers creates over-speculation
of vehicle charging demand. While
the MISOCP considers the number of EV chargers as an optimization variable, the SOCP
considers the number of chargers as a fixed given parameter. Since the objective function of

UC Berkeley, CE295
52

the SOCP is an affine function, the constant cost of chargers does not play a role in deter-
mining the planning and operation of the MG at optimum. Therefore, it is in the interest
of each SOCP to maximize the profit by selling as much EV charging load as possible with
a given number of chargers.
This strategy becomes unrealistic when the robust optimization scheme was adopted.
Consider a non-robust optimization, where the expected charging load has no uncertainty.
The optimal result would be to produce the exact amount of the EV charging load. However,
a robust optimization problem considers the uncertainty in the prediction of the EV charging
load and only requires the MG to be prepared for the EV charging load higher than the
average load by a certain uncertainty level. This only presents the lower bound of the EV
charging load as a constraint.
The absence of the upper bound for EV charging load conflicts with the motive of the
SOCP within a given set of constant chargers to maximize revenue from the charging of
vehicles. Therefore the result at optimum plans for an EV charging load that is much larger
than the mean value. In other words, the EV charging load is bounded only by the physical
power limits of the given EV chargers. This over-speculation of the vehicle charging demand
only considers the minimum required EV charging load for EV mobility needs, but is not
bound by a realistic EV demand scenario, i.e., it is unlikely that EV charging demand will
be at such a high level at all time steps.
To address this boundary condition, additional constraints can be formulated to limit the
upper bound of EV charging demand for each SOCP and therefore limit the over-speculation
of the vehicle charging demand. This constraint is described in the equations (6) and (7),
where we limit demand to a threshold value. The result of this constraint is given in Figure
7, where L3 charging power and charging demand on average are presented. The model then
binds the EV charging power to be at maximum the 99th percentile EV charging demand
throughout the time series.
Comparisons between the re-
sults depicted in Figure 3 and Fig-
ure 7 point to additional obser-
vations about model performance.
In Figure 3, during peak electric-
ity rates, the charging power of
both L2 and L3 decrease while still
matching the minimum delivered
energy to the vehicles. During this
time period, the electricity rate in-
creases from $0.26 to $0.78 USD
per kWh, while the revenue from
the L2 and L3 charging are held
constant as $0.20 and $0.30 USD
per kWh, respectively. Therefore,
Figure 7: The Upper Bound of EV Charging Demand
during the time of low rates, the
Uncertainty
model maximizes revenue from the
MG as much as possible. However,
during the time of high electricity rates, the grid import price for charging the vehicles is

UC Berkeley, CE295
53

not cost-effective, and the optimal strategy is to maintain charging at the minimum required
level.

4 Discussion
Mixed-Integer Linear Program
Due to computation times both L2 and L3 chargers ranging between 1 to 4 is about 20
seconds on MATLAB, the MILP was an ideal test bed for verifying the fidelity of the code and
the parameters used. However, each iteration could only be used for a single determinate
scenario, which is insufficient for robust optimization. Various scenarios must be tested
by iterating through cases such as days with low solar irradiation. Thus, even though each
iteration is computationally light, the necessity for conducting various iterations for different
scenarios makes the overall computational penalty of the MILP high. Considering the case
of using this program for planning, the MILP fails to provide a single optimal case, as every
iteration produces varying optimal values for equipment sizing.
Several cases also show that the MILP is not the most practical optimization method for
planning a commercial MG. As long as the charging limits of L3 chargers are not active, the
LP assumes that none of the EVs will be charging on L2, since the revenue per kWh for L3
charging is higher. This assumption is not realistic since popular PHEV vehicles such as the
Chevrolet Volt cannot charge via L3 [16]. The MILP is also able to derive a feasible solution
even when solar generation drops to nearly zero, as long as grid power can be imported.
Additionally, the MILP fails to provide a feasible solution when the building load exceeds
four times the mean load, or when the EV charging demand exceeds thirteen times the mean
EV demand used. Both of these cases are better managed by the SOCP that uses chance
constraints instead of a deterministic model.

Mixed-Integer Second-Order Cone Program

The MISOCP requires far greater computational power and often requires several hours
for the calculations to be completed. However, it results in a global optimal planning scenario,
as it is robust to stochasticity in energy generation and demand on the order of 1 hour.
Therefore, the MISOCP is the preferred optimization program for the case of planning a
MG. Using the MISOCP for planning MGs can advance adaptation of renewable energy
resources while decreasing dependence on grid imports, as risk factors surrounding return
on investment are not often predictable. However, since the MISOCP provides a benchmark
of projected daily revenue or cost while taking into account stochastic uncertainties, it is
able to present a more robust case for implementing a commercial MG in comparison to the
MILP that is not able to consider uncertainties in energy demand and supply, and thus costs
and revenues.

UC Berkeley, CE295
54

Future Work
Excessive battery discharge during the final time-step was revealed to related to the model
selling power to the grid before the end of its 24 hour “lifetime” to maximize revenue for the
day. Although a new constraint was implemented to avoid the battery from fully discharging
during the final time-step, it did not address the cause of this phenomenon. Some cases in
the MILP did occur where the battery was not being used other than discharging during
the final time-step. With the current model, the initial energy stored in the battery does
not have an incurred cost to it. A possible area of investigation would be to analyze the
relationship between the revenue generated by the discharge and the TOU cost of energy
imported from the grid.
The current models assume a time-step of 1 hour. The data collected to test the MILP
and the MISOCP were in terms of hourly increments. In reality, a true MG system should
adapt to rapid changes in energy generation and demand, such as when an energy-intensive
appliance is turned on. These changes occur in smaller time-steps, between seconds to
minutes. Shorter time-scales may require modified behavior for some system components.
For example, the power capacity of the battery may have to increase for it to be able to
regulate the system by injecting power as necessary in short time increments.
The results from both the MISOCP and MILP show that the optimal sizing for the
number of solar panels installed was often the maximum number possible. Although this
model was able to calculate the optimal sizing from an economic perspective, it does not
take into account the stability of the system. For example, in a scenario where the number of
solar panels is high and cloud cover suddenly clears, the increased generation from solar will
result in a sharp spike in energy supply. Without a protective device that limits the power
flow from solar generation, the system may attempt to shed excess energy in its battery
as discussed above. Costs incurred by the addition of such protective devices should also
be considered in the case of planning a commercial MG. Furthermore, the validity of the
constraints must be confirmed through stability analysis instead of assuming an arbitrary
number of maximum installed equipment.
Finally, V2G technologies have been gaining interest from both the public and private
sector as a method to stabilize the electricity grid as renewable penetration increases [17].
This optimization program did not include V2G capabilities, such that EVs plugged into the
chargers were not able to sell their electricity back to the grid. Enabling V2G capabilities
in the optimization program may be beneficial for future commercial use as vehicle owners
are able to choose whether they would like their vehicle to export power to the MG, and
increase revenue potential therefore minimizing cost for the system.

5 Summary
In this project, we developed a MISOCP for robust optimal planning and operation of
a commercial EV charging facility. The model was developed using real capital cost data,
artificial TOU pricing scheme, real solar data, EV charging data, and building load data.
The robust model results in an optimal design consisting of three L2 chargers and one L3
charger, with a total system daily cost of $70.39 per day. By comparison, the non-robust

UC Berkeley, CE295
55

model calculates an optimal design of one L2 charger and four L3 chargers. From this
example, it is clear that the robust model is well-suited for a likely range of scenarios,
while the non-robust model optimizes for the rare case of unusually high charging demand.
Further model development could include constraints for battery health, finer time-steps,
bi-directional EV charging, and accounting for capital costs of MG stabilization equipment.

UC Berkeley, CE295
56

References
[1] Frost & Sullivan. (2017) Global electric vehicle market outlook, 2017. [Online].
Available: https://cds.frost.com/p/72699/#!/ppt/c?id=MCC9-01-00-00-00

[2] Electric Power Research Institute (EPRI), “Understanding the grid impacts of plug-in
electric vehicles (pev),” 2012.

[3] N. G. Paterakis, O. Erdinç, I. N. Pappi, A. G. Bakirtzis, and J. P. S. Catalão, “Coor-

dinated operation of a neighborhood of smart households comprising electric vehicles,
energy storage and distributed generation,” IEEE Transactions on Smart Grid, vol. 7,
no. 6, pp. 2736–2747, Nov 2016.

[4] J. Munkhammar, J. Widen, and J. Ryden, “On a probability distribution

model combining household power consumption, electric vehicle home-charging and
photovoltaic power production,” Applied Energy, vol. 142, pp. 135 – 143, 2015. [Online].
Available: http://www.sciencedirect.com/science/article/pii/S0306261914012884

[5] S. Gao, K. T. Chau, C. Liu, D. Wu, and C. C. Chan, “Integrated energy management of
plug-in electric vehicles in power grid with renewables,” IEEE Transactions on Vehicular
Technology, vol. 63, no. 7, pp. 3019–3027, Sept 2014.

[6] S. J. Moura, “White paper: Robust optimal sizing energy management for solar +
storage + ev chargers,” Department of Civil and Environmental Engineering, University
of California, Berkeley, Berkeley, CA, Tech. Rep., December 2017.

[7] O. Erdinc, “Economic impacts of small-scale own generating and storage units,
and electric vehicles under different demand response strategies for smart
households,” Applied Energy, vol. 126, pp. 142 – 150, 2014. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0306261914003535

[8] O. Erdinc, N. G. Paterakis, I. N. Pappi, A. G. Bakirtzis, and J. P. Catalão, “A new

perspective for sizing of distributed generation and energy storage for smart households
under demand response,” Applied Energy, vol. 143, pp. 26 – 37, 2015. [Online].
Available: http://www.sciencedirect.com/science/article/pii/S0306261915000318

[9] O. Erdinc, N. G. Paterakis, T. D. P. Mendes, A. G. Bakirtzis, and J. P. S. Catalão,

“Smart household operation considering bi-directional ev and ess utilization by real-
time pricing-based dr,” IEEE Transactions on Smart Grid, vol. 6, no. 3, pp. 1281–1291,
May 2015.

[10] S. Bahramara and H. Golpı̂ra, “Robust optimization of micro-grids operation problem

in the presence of electric vehicles,” Sustainable Cities and Society, vol. 37, pp. 388
– 395, 2018. [Online]. Available: http://www.sciencedirect.com/science/article/pii/
S2210670717303505

[11] NREL. (2018) Pv watts calculator. [Online]. Available: https://pvwatts.nrel.gov/

pvwatts.php

UC Berkeley, CE295
57

[12] R. Fu, D. Feldman, R. Margolis, M. Woodhouse, and K. Ardani, “U.s. solar photovoltaic
system cost benchmark: Q1 2017,” NREL, 2017.

[13] BMW Group, “White paper: Ev charging data,” Tech. Rep.

[14] M. Gupta, “U.s. front-of-the-meter energy storage system prices 2018-2022,” GTM Re-
search, 2018.

[15] S. Knupfer, J. Noffsinger, and S. Sahdev, “How battery storage can help charge the
electric-vehicle market,” McKinsey Center for Future Mobility, 2018.

[16] EVTown. (2015) Levels of charging. [Online]. Available: http://www.evtown.org/

about-ev-town/ev-charging/charging-levels.html

[17] K. M. Tan, V. K. Ramachandaramurthy, and J. Y. Yong, “Bidirectional battery charger

for electric vehicle,” in 2014 IEEE Innovative Smart Grid Technologies - Asia (ISGT
ASIA), May 2014, pp. 406–411.

UC Berkeley, CE295
58

6 Appendix
Appendix A: List of Variables

List of Variables
Variable Units Description
s, b [-] Scale factors for solar and ESS sizes
GI (k), GE (k) [kW] Power imported and exported from/to grid
GD [kW] Maximum power for demand charge
S(k) [kW] (r.v.) Power generated from solar
S c (k) [kW] Curtailed solar power
B d (k), B c (k) [kW] Power discharged from / charged into battery
C 2 (k), C 3 (k) [kW] Charging load of L2, L3 EVSEs
L(k) [kW] (r.v.) Power load of facility
E(k) [kW] Energy level of ESS
Z(k) [kWh] Cumulative energy delivered to EVs
Z min (k) [kWh] (r.v.) Minimum delivered energy to EVs

Appendix B: List of parameters used in the optimization programs

List of Parameters Used

Parameter Value Units Description
cb 0.0274 [USD/day] Battery marginal cost for 20 year payment
cs 0.0822 [USD/day] Solar marginal cost for a 20 year payment
mc2 0.137 [USD] Marginal cost of level 2 chargers
mc3 0.219 [USD] Marginal cost of level 3 chargers
ci 0.26 [USD/kWh] Price of grid-imported power
re 0.03 [USD/kW] Revenue of exported power to the grid
r2 0.20 [USD/kWh] Revenue for level 2 charging
r3 0.30 [USD/kWh] Revenue for level 2 charging
cd 0.30 [USD/kWh] Demand charge
dt 1 [hour] Timestep
ηc 98 [% ] ESS charge efficiency
ηd 98 [% ] ESS discharge efficiency
E0 5 [kWh] Initial ESS energy level
Emax 14 [kWh] Nominal ESS energy capacity
Bmax 7 [kW] Nominal ESS power capacity
η2 98 [%] L2 charge efficiency
η3 98 [%] L3 charge efficiency
C2,max 10 [kW] L2 EVSE power capacity
C3,max 30 [kW] L3 EVSE power capacity
GI,max 100 [kW] Maximum grid import power
GE,max 100 [kW] Maximum grid export power
smin 1 [] Solar scale limit

UC Berkeley, CE295
59

smax 20 [ ] Solar scale limit

bmin 1 [ ] Battery scale limit
bmax 20 [ ] Battery scale limit
αz 0.95 [ ] Probability of satisfying EV energy limits
αG,1 0.95 [ ] Probability of satisfying grid import limits
αG,2 0.95 [ ] Probability of satisfying grid import limits

Acknowledgments
Thank you mom and data, Professor Moura and Bertrand for their time and patience

About the authors

Armando A. Domingos is a M.S. candidate in the Civil and Environmental Engineering
department at University of California, Berkeley. He likes cars, and thinks the best
PHEV is the Porsche 918.

Marc Hutton is a M.S. candidate in the Energy, Civil Infrastructure, and Climate program
in the Civil and Environmental Engineering department at University of California,
Berkeley. He hopes this project will help him get a job at Tesla (or any job, really).

Aaron Jagtianey is a B.S. candidate in the Energy Engineering program at the University
of California, Berkeley. He is a fan of the high revving Honda Civic Type-R EVs.

Soomin Woo is a PhD candidate in the Civil and Environmental Engineering with focus
on transportation engineering at the University of California, Berkeley.

UC Berkeley, CE295
60

UC Berkeley, CE295
61

Part II

Electric Vehicles and

Transportation

UC Berkeley, CE295
62

The Application of Estimation Techniques to Autonomous

Vehicles Related Object Tracking Problems

Hanxiao Deng, Haoran Su, Kun Qian, Yue Hu

1 Abstract
Tracking objects over time is a major challenge for understanding the present location and environment
surrounding an autonomous vehicle. Kalman filters(with its extensions) and Particle filter are traditional
but very powerful estimation techniques in estimation problems. In this project, we will implement a
Kalman Filter and an Extended Kalman Filter in C++ environment to predict the location of other moving
vehicles on the road with noisy LiDAR and Radar measurements around the autonomous car. We will
also estimate the location of an autonomous vehicle with the measured location using Particle Filter. The
main focus of this project is to apply estimation techniques to autonomous vehicle related estimation and
tracking problems.

2 Introduction
2.1 Motivation and Background
Autonomous Vehicle is a heated topic recently. A key technique used is state estimation-to understand the
location of the vehicle itself and the environment around. Here, we want to solve two problems: Tracking
around objects with LiDAR and Radar on an autonomous car; Tracking the pose of an autonomous
vehicle with noisy location measurement data. We simplify the scenario that there are localization
measurements by LiDAR and Radar of only one moving car, as shown in figure 1. We describe how data
from LiDAR and Radar can be used for optimal estimation with Kalman filter. Sensor fusion of Lidar
and radar combines the advantages of both sensor types to increase the precision of estimation. In the
vehicle’s pose estimation problem, we suppose that the sensor will present us with the location data of a
vehicle in 2D sense.
The LIDAR will produce 3D measurement px, py, pz. But for the case of driving on the road, we could
simplify the pose of the tracked object as: px, py and one rotation. In other words, we could only use px, py
to indicate the position of the object, and one rotation to indicate the orientation of the object. But in real
world where you have very steep road, you have to consider z axis as well. Also in application like
airplane and drone, you definitely want to consider pz as well.

2.2 Relavent Literature

All Kalman filters have similar main steps: 1. Initialization, 2. Prior Prediction, 3. Measurement Update.
A Standard Kalman Filter (KF) can only handle linear equations [1]. Both the Extended Kalman Filter
(EKF) and the Unscented Kalman Filter (UKF) allow you to use non-linear equations. The work flow

UC Berkeley, CE295
63

(a) LiDAR Measurement

(b) Radar Measurement

Figure 1: Lidar and Radar Measurements

(a) Extended Kalman Filter Workflow (b) Extended Kalman Filter v.s. Kalman Filter

Figure 2: Kalman Filter and Extended Kalman Filter

for Extended Kalman Filter and its comparison with Kalman Filter is shown in figure 2. The difference
between EKF and UKF is how they handle non-linear equations: Extended Kalman Filter uses the
Jacobian matrix to linearize non-linear functions[2]; Unscented Kalman Filter, on the other hand, does not
need to linearize non-linear functions, instead, the unscented Kalman filter takes representative points
from a distribution and then employ unscented transform[3]. These tools come in handy when trying to
estimate the trajectory from LiDAR or RADAR data, where depending on the measurement we can both
encounter linear and nonlinear equations.
Another problem to solve is that how to extract location information about vehicles from cloud point
data collected by LiDAR. People made research on using machine learning techniques to extract vehicles’
location information from cloud point data[6]. We follow the track to extract vehicle’s location informa-
tion from cloud point data that LiDAR collected.
For the tracking of autonomous vehicles itself, we use the simplified bicycle model[7] to describe the
autonomous vehicle which is proved to be relatively precise and computationally efficient. We will set up
the state space functions based on the bicycle model and apply particle filters directly.

2.3 Focus of this study

1. Implement feasible estimators to track other vehicles in the environment with LiDAR and Radar
measurements;

2. Estimate the pose of vehicle based on the location measurement;

3. Improve algorithms to implement estimation to meet real time requirement;

UC Berkeley, CE295
64

4. Discuss and compare the precision and computation amount between different techniques;

3 Model Development
3.1 Model for Vehicles Tracking Problem
3.1.1 The goal of our state space model
We would like to depend on the state space to estimate the position of other moving cars. The data we
have is the measurement data by Lidar and Radar. Lidar will give us the position data of a moving
vehicles. We suppose that the Lidar data comes directly in the caritsean grid. Radar will provide us the
position and velocity data of a moving object, however, in the polar grid.
 
  ρ
x
z Lidar = , zradar =  θ 
y
ρ̇

3.1.2 Physics function of our system

In fact, the physics equation that we depend on to do the tracking problem is the Motion function.
Because here we have no knowledge about the acceleration information about the object that we are
tracking, thus we consider the acceleration as a noise. It may cause some problems because usually
Kalman Filter requires noise to be unbiased, however, here acceleration is sometimes biased.

ẋ = v

v̇ = ν
We would like to write this equation in a discrete form.

xk+1 = xk + ηvk + ν1

vk+1 = vk + ν2
in which xk is the location of the vehicle at time k, η is the time step size, vk is the speed at time k, ν is
the process noise of the model. Here the process noise mainly comes from our ignorance of acceleration
information.  
x
y
We consider the problem from 2D sense. So that the state of our system will be  v x  and the formulation


vy
of the state space model in continuous form will be
      
ẋ 0 0 1 0 x ν1
 ẏ  0 0 0 1  y  ν2 
 =  + 
v˙x  0 0 0 0 v x  ν3 
v˙y 0 0 0 0 vy ν4

If we write it in the discrete form, it would be(with the η be the time step)
      
x ( k + 1) 1 0 η 0 x (k) ν1
 y ( k + 1)  0 1 0 η  y(k)  ν2 
 v x ( k + 1)  = 0
    + 
0 1 0  v x (k) ν3 
v y ( k + 1) 0 0 0 1 vy (k ) ν4

UC Berkeley, CE295
65

So that we will have the A matrix:  

1 0 η 0
0 1 0 η
A=
0

0 1 0
0 0 0 1

3.1.3 The form of measurement equation

LiDAR data
As stated above, we suppose that Lidar will directly give us the location measurement at Cartisean grid.
The measurement equation will be
 
    x (k)  
z x (k) 1 0 0  y(k)  + wx
0  
=
zy (k ) 0 1 0 0 v x (k) wy
vy (k )
 
1 0 0 0
H=
0 1 0 0
in which z is the measurement data and w is the measurement noise.
To ensure that our estimation error will finally converge, here we want to see the observability or
detectability of this system.
For the simplicity of our work, we use PBH test. Because the λ A = 1, then

0 0 0
 
η
0 0 0 η
   
λA I − A 0 0 0 0
= 
H 0 0
 0 0
1 0 0 0
0 1 0 0
 
λA I − A
Rank( )=4
H
So this system is observable. In fact to ensure the convergence of Kalman filter error, we only need
detectable. Anyway, here observability will do a lot of good to our problem solving and also proves that
our formulation of this system is reasonable.

Radar data
The case for Radar measurement data is a little bit more complex. Because we need to do the transforma-
 
ρ
tion between polar and Cartisean. We suppose that the Radar will present us with the data ρ̇  and the

Φ
relationship with our state variables are q
ρ= x 2 + y2
y
Φ = arctan( )
x
q vy
ρ̇ = v2x + v2y cos (Φ − arctan )
vx

UC Berkeley, CE295
66

We can see that the measurement output of Radar data is nonlinear. Here we can use both EKF or UKF.
EKF depends on the linearization process to translate the problem to the Kalman Filter’s range. UKF
depends on the unscented transform to pass on the PDF of estimation object.
Here we will use EKF first, so we need to calculate the Jacobian of the nonlinear output function. The
Jacobian is calculated as follows.

∂ρ x
= p
∂x x 2 + y2
∂ρ y
= p
∂y x 2 + y2
∂Φ 1
= −p
∂x x 2 + y2
∂Φ x
= p 1
∂y x 2 + y2
∂ρ̇ q y vy 1
= v2x + v2y (− sin(arctan( ) − arctan( )))(− p )
∂x x vx x + y2
2

∂ρ̇ q y vy x
= v2x + v2y (− sin(arctan( ) − arctan( )))( p )
∂y x vx x + y2
2

∂ρ̇ q y vy vx vx y vy
= v2x + v2y (− sin(arctan( ) − arctan( )))(− q )+ q cos (arctan( ) − arctan )
∂v x x vx 2
v +v 2 2
v +v 2 x v x
x y x y

∂ρ̇ q y vy vx vy y vy
= v2x + v2y (− sin(arctan( ) − arctan( )))(− q )+ q cos (arctan( ) − arctan )
∂vy x vx v2 + v2 v2 + v2 x vx
x y x y

3.2 Model for Vehicle Pose Estimation

In studying the dynamics of autonomous vehicles, as stated before, we usually use the simplified bicycle
model to study the problem. The bicycle model consider the vehicle as a bicycle, and consider the location,
heading angle as states of the bicycle.
For the inputs of the system, we will have knowledge of the steering angle and rotating speed of wheel.
Comparing to the more complexed model, the simplified bicycle model doesn’t consider side slip of
wheels and thus can describe the system with less states.
The description of the bicycle model is as below. In which we will use x, y, θ as the state of the system. γ
and ω are known inputs.

(The Picture is taken from E-Bicycle website, [9])

ẋ (t) = v(t) cos(θ (t))

ẏ(t) = v(t) sin(θ (t))
v(t)
θ̇ (t) = tan(γ(t))
B
According to the physical relation, v(t) = rω B (t), ω B (t) = 5ω (t), t we will have v(t) = 5rω (t). Then the
equation will turn into the following form.

UC Berkeley, CE295
67

Figure 3: Bicycle Model

ẋ (t) = 5rω (t) cos(θ (t))

ẏ(t) = 5rω (t) sin(θ (t))
5rω (t)
θ̇ (t) = tan(γ(t))
B
For the observation part, we are going to measure the location of the center of the bicycle. The observation
functions are as follows.

1
xc (t) = x (t) + B cos(θ (t))
2
1
yc (t) = y(t) + B sin(θ (t))
2
In which γ(t) and ω (t) are known inputs.
We will consider the noise terms in the following part.

3.2.1 Discrete Model

We use the first order Euler approximation to turn the continuous system into a discrete one. The discrete
equations are as follows.

x (t + 1) = x (t) + 5rω (t) cos(θ (t))∆t

y(t + 1) = y(t) + 5rω (t) sin(θ (t))∆t
5r
θ (t + 1) = θ (t) + ω (t) tan(γ(t))∆t
B
The observation model will be the same as above.

3.2.2 Assumption on Noise

Let’s then consider the noise terms.

Process Noise
According to the provided information, process noise can come from several sources.

UC Berkeley, CE295
68

1. The uncertainty in the size of bicycle B and r, the uncertainty of B is given between ±10%, the
uncertainty of r is given between ±5%, however, for specific cars, the size is fixed.

2. The noise in the measurement of ω (t) and γ(t)

3. The time step may have uncertainty

We make the following assumptions on process noise.

1. ω (t) and γ(t) also follow Gaussian distribution with N (ω (t), 0.01) and N (γ(t), 0.01).

2. Comparing to other noise, the noise in time step size can be roughly ignored. Also, to solve
the problem that measurement time gap is not fixed, we can just do the process update without
measurement update if we don’t have measurement at that step.

Measurement Noise
Measurement noise mainly comes from the noise in measuring the vehicle’s location. We consider it as
additive noise. The system with noise term will be as follows.

1
xc (t) = x (t) + B cos(θ (t)) + w1
2
1
yc (t) = y(t) + B sin(θ (t)) + w2
2
We make the assumption that the distribution of measurement noise will be the same all the time. So we
will be able to calculate out the characteristics of measurement noise through callibration dataset.
Through the calculation, we can see that the distribution of measurement noise is as follows.
The variance of w1 and w2 are listed as below(we made the assumption that w1 and w2 are independent
with each other):
σw2 1 = 1, σw2 2 = 1.7

Noise Model for Particle Filter

x (t + 1) = x (t) + 5rω (t)(1 + vω (t) ) cos(θ (t))∆t

y(t + 1) = y(t) + 5rω (t)(1 + vω (t) ) sin(θ (t))∆t
5r
θ ( t + 1) = θ ( t ) + ω (t) tan(γ(t)(1 + vγ(t) ))∆t
B

1
xc (t) = x (t) + B cos(θ (t)) + w1
2
1
yc (t) = y(t) + B sin(θ (t)) + w2
2
in which vγ(t) ∼ N (0, 0.1), vω (t) ∼ N (0, 0.1), w1 ∼ N (0, 1), w2 ∼ N (0, 1.7).
We make the assumption on the distribution of initial state that x (0) ∼ N (0, 3), y(0) ∼ N (0, 3),
θ (0) dim N ( π4 , 0.3).

UC Berkeley, CE295
69

4 The Iterative Form of Estimators

In numerical implementation, we need to do the estimation in discrete case. The estimation methods
usually follow the Bayesian tracking thought in which three steps are initialization, prior update and
measurement update. Different methods based on different assumptions on noise and different methods
to describe the property of the distribution of the states. Detailed introductions are given below about
different estimation techniques. We refer to relevant studying materials about these methods.[8]

4.1 Kalman Filter update

In the lecture we introduced the continuous form of Kalman filter. However, when doing computation,
we should use the discrete form which is much easier to understand.
We use some auxiliary variables here.
x m (0) = x (0)
x p ( k ) = A ( k − 1) x m ( k − 1) + ν ( k − 1)
zm (k ) = H (k ) x p (k ) + w(k )
f xm (k) = f x p (k)|zm (k)
The philosophy behind this is two steps of update.
The first update is called prior update which is denoted by x p . In this update step, we get our prior
knowledge about the state with xm (k − 1) and our process function(our state space model).
The xm denote the measurement update of x which is defined above through its PDf. And from the
iterative view, xm is in fact the estimated x with the adjustment from measured data z.
The desired estimated state here is in fact to be extracted from the PDF of xm .
Because of the finiteness of our inner storage space, we cannot store all the information about PDF.
We will only pass the expected value and variance to the next iterative. The following is the iterative
algorithm:

init: xbm (0) = x0

xbp (k ) = A(k − 1) xbm (k − 1)
Pp (k ) = A(k − 1) Pm (k − 1) A> (k − 1) + Q(k − 1)
K (k ) = Pp (k) H > (k)( H (k) Pp (k) H > (k) + R(k))( − 1)
xbm (k ) = xbp (k) + K (k)(z(k ) − H (k) xbp (k ))
Pm (k) = ( I − K (k ) H (k)) Pp (k − 1)( I − K (k) H (k))> + K (k) R(k)K > (k)

To make the process more intuitive, I implemented it in Python as follows:

4.2 EKF difference

The only difference that EKF may have with KF is that EKF deals with nonlinear functions. However,
EKF is not a magic but only an approximation with Taylor 1st order expansion. So here, in order to deal
with the nonlinearity that Radar measurement brings to us, we are going to linearize the measurement
data around every time period.

UC Berkeley, CE295
70

In fact, as I have listed the Jacobian of the nonlinear function, I just need to fill the linearized H matrix
with the value that I have got.

√ 2x 2 √ 2y 2
 
 ∂ρ ∂ρ  0 0
0 0  x +y x +y 
 ∂x
∂ ρ̇
∂y
∂ ρ̇ ∂ ρ̇ ∂ ρ̇   
H (k) =   =
 
 ∂x ∂y ∂vx ∂vy  ∂ ρ̇ ∂ ρ̇ ∂ ρ̇ ∂ ρ̇ 
∂Φ ∂Φ
 ∂x ∂y ∂v x ∂vy 
0 0
− √ 21 2 √ x21 2
 
∂x ∂y 0 0
x +y x +y

in which
∂ρ̇ q y vy 1
= v2x + v2y (− sin(arctan( ) − arctan( )))(− p )
∂x x vx x 2 + y2
∂ρ̇ q y vy x
= v2x + v2y (− sin(arctan( ) − arctan( )))( p )
∂y x vx x 2 + y2
∂ρ̇ q y vy vx vx y vy
= v2x + v2y (− sin(arctan( ) − arctan( )))(− q )+ q cos (arctan( ) − arctan )
∂v x x vx v2 + v2 v2 + v2 x vx
x y x y

∂ρ̇ q y vy vx vy y vy
= v2x + v2y (− sin(arctan( ) − arctan( )))(− q )+ q cos (arctan( ) − arctan )
∂vy x vx v2 + v2 v2 + v2 x vx
x y x y

4.3 Particle Filter

Particle Filter is in fact similar to Kalman Filter inside. They are all Bayesian tracking problem based
Bayes rule. However, Particle Filter employ Monte Carlo Method to describe the distribution. Also
particle filter is not strict on the requirement of noise distribution. Usually the more particles we have, the
better the estimation will be but also more expensive in computation. We pick up roughly 1000 particles
to characterize the distribution of states of autonomous vehicle’s location information.

Initialization
n (0) based on f ( x (0)). These are initial particles.
Draw N samples xm

Prior update
x np = qk−1 ( xmn ( k − 1), vn ( k − 1)), for n = 1, 2, ..., N
n
in which v (k − 1) will be N particles sampled from f (v(k − 1)), the function qk will be the process
function listed above.

Measurement update
Scale each particle by measurement likelihood:

z(k)| x np (k)), for n = 1, 2, ..., N

β n = α f z(k)| x(k) (b

where α is the normalization constant chosen such that ∑nN=1 β n = 1.

n ( k ) from the scaled distribution.
Re-sample N particles xm

UC Berkeley, CE295
71

5 Estimation Results
5.1 Car Tracking with LiDAR and Radar Result

(a) Single frame point cloud from Velodyne Lidar (b) Detected objects in red bounding box

Figure 4: Vehicle Detection plots

From Figure.3, we present the raw point cloud data generated from Velodyne Lidar sensor in the real
road. Then, we perform the standard connected component algorithm to detect objects, as showed in
the red bounding boxes. After getting object point cloud from each bounding box, we need to classify
these object categories based on 28 features extracted from point cloud, which include 4 global features
and 24 local features. Each object labeled with categories can be associated with the objects in last frame,
which is called data association, based on Kalman Filter prediction and Hungarian algorithm. After
data association, we can use the measurement data - the center of the bounding box as object position to
correct the prediction, which is called measurement update. Figure.4 shows a frame tracking of front car.
With the extracted measurement data from LiDAR and Radar, we track other vehicles with estimation
technique and physics model. The estimation result is shown as below in Figure6. From the estimation
result we can see that the red circles and blue circles are measurement points, green triangles are estimated
positions. The estimated route of vehicle is smooth. However, cases we consider here are relatively linear.
The extended Kalman Filter may have a poor performance in nonlinear cases.

5.2 Vehicle Pose Estimation Result

With Particle Filter, we present our estimation result below in Figure 7.

In the following figure, we plot the measurements points and the estimation results of the location of
the autonomous vehicle in the same plot. Red crossings are measurements at some time steps. The blue
line is the moving route drawn from our estimation.
Because we don’t have knowledge of real states, we can only present the measurement points and
compare it with our estimation to see if it’s reasonable. We know that measurements come with noise
so it’s not very accurate. In some very noisy cases, we cannot tell the vehicle’s moving track from only
the measurement. With the help of physical model and the estimation technique, we present a relatively
smooth and accurate moving track of the vehicle.
To decide the precision of the estimator, we will compare the estimation result at the last step with the
true state and see the error. Because for particle filter, the estimation results will be different every time.

UC Berkeley, CE295
72

Figure 5: Tracked object (green) by Kalman filter

Figure 6: Estimation Results From LiDAR and Radar

So the error is not fixed after every run but will be similar.
A sample group of errors under several cases are listed below in the following table.

Case No. Error X(m) Error Y(m) Error θ(rad)

1 -0.212 -0.311 0.022
2 0.296 0.801 0.385
3 0.322 0.640 0.215
4 -1.219 0.805 -0.292

UC Berkeley, CE295
73

(a) Case 1 (b) Case 2

(c) Case 3 (d) Case 4

Figure 7: Vehicle Tracking plots

UC Berkeley, CE295
74

6 Summary and Future work

6.1 Summary
1. In the first part, we employed Kalman filter and Extended Kalman filter to track other vehicles with
the measurement data from both LiDAR and Radar. Machine learning method is used to extract
object’s location information from cloud point data. And then Kalman filter is used to handle
LiDAR measurement data and Extended Kalman Filter is used to handle Radar measurement
because of the nonlinear measurement equation of Radar. Through our observation, we found
that the estimation precision is greatly improved after combining measurements from two sources
together. Also, because we made the assumption that the cases we focus on is close to linear cases,
our estimation turns out to be good.

2. In the second part, particle filter is employed to estimate the pose of an autonomous vehicle itself.
Particle Filter is another choice for estimation problem combining the thought of Monte Carlo
Method and tracking method together. This usually gives better estimation result but at the cost of
much higher computation amount. For a simplified autonomous vehicle model, it’s tractable to
employ particle filter to estimate because it doesn’t have too many states. Thus we can estimate
states under many very nonlinear cases with a reasonable number of particles. The key issues in
implementing a particle filter include preventing the number of different particles from converging
and giving a right assumption on the distribution of the noise.

3. When choosing appropriate estimation technique to use, we need to consider the assumption of
noise of the system. It’s always a trade off between estimation precision and computation amount.
So we need to choose a suitable estimation method according to our estimation requirements.
Usually Kalman filter(with its extensions) come with low computation cost but lower precision.
Particle filter is with better precision and higher computation amount.

6.2 Future Work

1. In the first problem–tracking other vehicles around, we can try to extend the situation to rela-
tively nonlinear cases and use some other methods to see the performance of different estimation
techniques under these nonlinear cases.

2. In the second problem–estimation of vehicle’s pose, there are some parameters in the system which
should be a fixed number but we don’t know for certain. We can list these parameters in the state
to try to identify them. It may help to improve the accuracy of the estimation.

7 Acknowledgement
Thanks Professor Moura for excellent instruction for CE295, really benefit from this course a lot!
Thanks Bertrand for serving as GSI and holding helpful sections!
Thanks those who devoted to developing CE295 into a charming course!
Time in Berkeley passes by so quickly, no matter it’s a happy or a not that satisfying experience to us, we
cannot deny that we all learn a lot here. Goodbye Berkeley, will miss the campus and friends here.:)
Fiat Lux!

UC Berkeley, CE295
 75

References
[1] Kalman Filtering: Kalman RE. A New Approach to Linear Filtering and Prediction Problems. ASME.
J. Basic Eng. 1960;82(1):35-45. doi:10.1115/1.3662552.

[2] Extended Kalman Filtering: Frühwirth, R. Application of Kalman filtering to track and vertex
fitting. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers,
Detectors and Associated Equipment 262, 444–450 (1987).

[3] Unscented Kalman Filtering: Simon J. Julier, Jeffrey K. Uhlmann, "New extension of the Kalman filter
to nonlinear systems", Proc. SPIE 3068, Signal Processing, Sensor Fusion, and Target Recognition VI,
(28 July 1997); doi:

[4] Particle Filtering: Tutorial on Particle Filters for Online Nonlinear/NonGaussian Bayesian
Tracking. Bayesian Bounds for Parameter Estimation and Nonlinear Filtering/Tracking (2009).
doi:10.1109/9780470544198.ch73

[5] Modelling: A Tutorial on Particle Filters for Online Nonlinear/NonGaussian Bayesian Track-
ing. Bayesian Bounds for Parameter Estimation and Nonlinear Filtering/Tracking (2009).
doi:10.1109/9780470544198.ch73

[6] LiDAR based object perception: M.Himmelsbach, A. Muller, T.Luttel and H.-J.Wunsche, LIDAR-
based 3D Object Perception,Proceedings of 1st international workshop on cognition for technical
systems, volume 1. 2008

[7] Bicycle Model: P. Falcone, F. Borrelli, J. Asgari, Predictive Active Steering Control for Autonomous
Vehicle Systems, IEEE Transactions on Control Systems Technology, Volume. 15, Issue. 3, May 2007

[8] Mark Mueller, Advanced Control(2) Lecture Notes, 2018

[9] Bicycle picture: http://www.ebicycles.com/bicycle-tools/measure-frame

8 Appendix: Code Implementation in C++

8.1 Simulation of Moving Object
We simulate the moving object by dividing the program into initialization, prediction and update parts:
The initialization part goes as follows:
1 void T r a c k e r : : ProcessMeasurement ( geometry_msgs : : PoseArray& msg )
2 {
3 /∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗
4 ∗ Initialization
5 ∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗/
6 if (! is_initialized_ )
7 {
8 // f i r s t measurement
9 k f _ . x_ = VectorXd ( 4 ) ;
10 k f _ . x_ << 1 , 1 , 1 , 1 ; //avoid no v a l u e s i n x_
11
12 // we need t o s e l e c t t h e convoy l e a d e r v e h i c l e here
13 // we assume t h a t t h e s t r a i g h t f r o n t and c l o s e d one i s t h e l e a d e r v e h i c l e
14 geometry_msgs : : P o i n t p o i n t ;
15 p o i n t . x = n u m e r i c _ l i m i t s < f l o a t > : : max ( ) ;
16 p o i n t . y = n u m e r i c _ l i m i t s < f l o a t > : : max ( ) ;
17 geometry_msgs : : P o i n t o r i g i n ;
18 origin . x = 0 . 0 ;
19 origin . y = 0 . 0 ;

UC Berkeley, CE295
 76

20 f l o a t d i s t a n c e = e u c l i d e a n _ d i s t a n c e ( point , o r i g i n ) ;
21 i n t s i z e = msg . poses . s i z e ( ) ;
22 f o r ( i n t i = 0 ; i < s i z e ; ++ i )
23 {
24 geometry_msgs : : P o i n t c a r _ p o s e = msg . poses [ i ] . p o s i t i o n ;
25 i f ( ( c a r _ p o s e . x > 0 . 0 ) && ( abs ( c a r _ p o s e . y ) < 6 . 0 ) && ( e u c l i d e a n _ d i s t a n c e ( o r i g i n ,
car_pose ) < distance ) )
26 {
27 point = car_pose ;
28 distance = euclidean_distance ( origin , car_pose ) ;
29 }
30 }
31
32 // no l e a d e r v e h i c l e a s s o c i a t e d
33 i f ( d i s t a n c e > 1000 )
34 {
35 ROS_INFO ( " Convoy l e a d e r a s s o c i a t i o n f a i l e d a t t h i s frame\n " ) ;
36 return ;
37 }
38 k f _ . x_ << p o i n t . x , p o i n t . y , 0 , 0 ; // x , y , vx , vy
39 previous_timestamp_ = msg . header . stamp . t o S e c ( ) ; // s e t c u r r e n t time stamp
40 // done i n i t i a l i z i n g , no need t o p r e d i c t or update
41 i s _ i n i t i a l i z e d _ = true ;
42 return ;
43 }

The prediction parts go as follows:

1 // compute t h e time e l a p s e d between t h e c u r r e n t and pr e viou s measurements
2 f l o a t dt = msg . header . stamp . t o S e c ( ) − previous_timestamp_ ; // i n seconds
3 previous_timestamp_ = msg . header . stamp . t o S e c ( ) ;
4
5 f l o a t dt_2 = dt ∗ dt ;
6 f l o a t dt_3 = dt_2 ∗ dt ;
7 f l o a t dt_4 = dt_3 ∗ dt ;
8
9 // Modify t h e F matrix so t h a t t h e time i s i n t e g r a t e d , from 1 t o s p e c i f i c dt ( s e e l i n e 5 1 )
10 k f _ . F_ ( 0 , 2 ) = dt ;
11 k f _ . F_ ( 1 , 3 ) = dt ;
12
13 // s e t model n o i s e s ( here i s t h e a c c e l e r a t i o n )
14 f l o a t noise_ax = 9 ;
15 f l o a t noise_ay = 9 ;
16
17 // s e t t h e p r o c e s s c o v a r i a n c e matrix Q (Q = X ∗ a ∗ X_T )
18 k f _ . Q_ = MatrixXd ( 4 , 4 ) ;
19 k f _ . Q_ << dt_4 /4∗ noise_ax , 0, dt_3 /2∗ noise_ax , 0,
20 0, dt_4 /4∗ noise_ay , 0 , dt_3 /2∗ noise_ay ,
21 dt_3 /2∗ noise_ax , 0, dt_2 ∗ noise_ax , 0,
22 0, dt_3 /2∗ noise_ay , 0 , dt_2 ∗ n o i s e _ a y ;
23
24 kf_ . Predict ( ) ;

The update part goes as follows:

1 // Lidar updates
2 k f _ . H_ = H_lidar_ ;
3 k f _ . R_ = R _ l i d a r _ ;
4
5 // we need t o s e l e c t t h e convoy l e a d e r v e h i c l e here
6 // by gated n e a r e s t neighbor data a s s o c i a t i o n but r e a s o n a b l e d i s t a n c e t o avoid a s s o c i a t e
other v i h i c l e
7 geometry_msgs : : P o i n t p o i n t ;
8 p o i n t . x = n u m e r i c _ l i m i t s < f l o a t > : : max ( ) ;
9 p o i n t . y = n u m e r i c _ l i m i t s < f l o a t > : : max ( ) ;

UC Berkeley, CE295
 77

10
11 geometry_msgs : : P o i n t car_pred ;
12 car_pred . x = k f _ . x_ [ 0 ] ;
13 car_pred . y = k f _ . x_ [ 1 ] ;
14 f l o a t d i s t a n c e = e u c l i d e a n _ d i s t a n c e ( point , car_pred ) ;
15 i n t s i z e = msg . poses . s i z e ( ) ;
16 f o r ( i n t i = 0 ; i < s i z e ; ++ i )
17 {
18 geometry_msgs : : P o i n t c a r _ p o s e = msg . poses [ i ] . p o s i t i o n ;
19 i f ( ( c a r _ p o s e . x > 0 . 0 ) && ( e u c l i d e a n _ d i s t a n c e ( car_pred , c a r _ p o s e ) < d i s t a n c e ) )
20 {
21 point = car_pose ;
22 d i s t a n c e = e u c l i d e a n _ d i s t a n c e ( car_pred , c a r _ p o s e ) ;
23 }
24 }

8.2 Estimation Code

Then we would employ the extended Kalman Filter to estimate the therotical position of the object:
1 void K a l m a n F i l t e r : : I n i t ( VectorXd &x_in , MatrixXd &P_in , MatrixXd &F_in , MatrixXd &H_in ,
MatrixXd &R_in , MatrixXd &Q_in )
2 {
3 x_ = x _ i n ;
4 P_ = P_in ;
5 F_ = F_in ;
6 H_ = H_in ;
7 R_ = R_in ;
8 Q_ = Q_in ;
9 }
10
11 void K a l m a n F i l t e r : : P r e d i c t ( )
12 {
13 /∗∗
14 ∗ p r e d i c t t h e s t a t e x_k|k−1 and s t a t e c o v a r i a n c e matrix P_k|k−1
15 ∗/
16 x_ = F_ ∗ x_ ;
17 MatrixXd F _ t = F_ . t r a n s p o s e ( ) ;
18 P_ = F_ ∗ P_ ∗ F _ t + Q_ ;
19 }
20
21 void K a l m a n F i l t e r : : Update ( c o n s t VectorXd &z ) {
22 /∗∗
23 ∗ measurement update t o c o r r e c t t h e p r e d i c t i o n
24 ∗ z i s t h e t r u e measurement
25 ∗/
26 MatrixXd z_pred = H_ ∗ x_ ;
27 MatrixXd S_ = H_ ∗ P_ ∗ H_ . t r a n s p o s e ( ) +R_ ;
28 MatrixXd K_ = P_ ∗ H_ . t r a n s p o s e ( ) ∗ S_ . i n v e r s e ( ) ;
29
30 // e s t i m a t e
31 x_ = x_ + K_ ∗ ( z − z_pred ) ;
32 P_ = P_ − K_ ∗ (H_ ∗ P_ ∗ H_ . t r a n s p o s e ( ) + R_ ) ∗ K_ . t r a n s p o s e ( ) ;
33 }

The following is the code for particle filter in tracking autonomous vehicle which is implemented with
python.

1 import numpy as np
2 import s c i p y as sp
3 #NO OTHER IMPORTS ALLOWED ( However , you ’ r e allowed t o import e . g . s c i p y . l i n a l g )

UC Berkeley, CE295
 78

4
5 def e s t I n i t i a l i z e ( ) :
6 # F i l l i n whatever i n i t i a l i z a t i o n you ’ d l i k e here . This f u n c t i o n g e n e r a t e s
7 # t h e i n t e r n a l s t a t e o f t h e e s t i m a t o r a t time 0 . You may do whatever you
8 # l i k e here , but you must r e t u r n something t h a t i s i n t h e format as may be
9 # used by your run ( ) f u n c t i o n .
10 #
11
12 #we make t h e i n t e r a l s t a t e a l i s t , with t h e f i r s t t h r e e elements t h e p o s i t i o n
13 # x , y ; t h e angle t h e t a ; and our f a v o u r i t e c o l o r .
14
15 n p a r t i c l e s = 1000
16 x _ i n i t = np . random . normal ( 0 , 3 , ( n p a r t i c l e s ) )
17 y _ i n i t = np . random . normal ( 0 , 3 , ( n p a r t i c l e s ) )
18 t h e t a _ i n i t = np . random . normal ( np . p i / 4 , 0 . 5 , ( n p a r t i c l e s ) )
19 p a r t i c l e s = np . column_stack ( [ x _ i n i t , y _ i n i t , t h e t a _ i n i t ] )
20 w_mu = np . a r r a y ( [ [ 0 ] , [ 0 ] ] )
21 w_sig = np . a r r a y ( [ [ 1 , 0 ] , [ 0 , 1 . 7 ] ] )
22
23 c o l o r = ’ green ’
24 # note t h a t t h e r e i s ∗ a b s o l u t e l y no p r e s c r i b e d format ∗ f o r t h i s i n t e r n a l s t a t e .
25 # You can put i n i t whatever you l i k e . Probably , you ’ l l want t o keep t h e p o s i t i o n
26 # and angle , and probably you ’ l l remove t h e c o l o r .
27 internalState = [ nparticles ,
28 particles ,
29 color ,
30 w_mu,
31 w_sig
32 ]
33 return i n t e r n a l S t a t e
34
35 import numpy as np
36 import s c i p y as sp
37 #NO OTHER IMPORTS ALLOWED ( However , you ’ r e allowed t o import e . g . s c i p y . l i n a l g )
38
39 def estRun ( time , dt , i n t e r n a l S t a t e I n , s t e e r i n g A n g l e , pedalSpeed , measurement ) :
40 # In t h i s f u n c t i o n you implement your e s t i m a t o r . The f u n c t i o n arguments
41 # are :
42 # time : c u r r e n t time i n [ s ]
43 # dt : c u r r e n t time s t e p [ s ]
44 # i n t e r n a l S t a t e I n : t h e e s t i m a t o r i n t e r n a l s t a t e , d e f i n i t i o n up t o you .
45 # s t e e r i n g A n g l e : t h e s t e e r i n g angle o f t h e bike , gamma, [ rad ]
46 # pedalSpeed : t h e r o t a t i o n a l speed o f t h e pedal , omega , [ rad/s ]
47 # measurement : t h e p o s i t i o n measurement v a l i d a t t h e c u r r e n t time s t e p
48 #
49 # Note : t h e measurement i s a 2D v e c t o r , o f x−y p o s i t i o n measurement .
50 # The measurement s e n s o r may f a i l t o r e t u r n data , i n which c a s e t h e
51 # measurement i s given as NaN ( not a number ) .
52 #
53 # The f u n c t i o n has f o u r outputs :
54 # e s t _ x : your c u r r e n t b e s t e s t i m a t e f o r t h e b i c y c l e ’ s x−p o s i t i o n
55 # e s t _ y : your c u r r e n t b e s t e s t i m a t e f o r t h e b i c y c l e ’ s y−p o s i t i o n
56 # e s t _ t h e t a : your c u r r e n t b e s t e s t i m a t e f o r t h e b i c y c l e ’ s r o t a t i o n t h e t a
57 # i n t e r n a l S t a t e : t h e e s t i m a t o r ’ s i n t e r n a l s t a t e , i n a format t h a t can be understood by
t h e next c a l l t o t h i s f u n c t i o n
58
59 # Example code only , you ’ l l want t o h e a v i l y modify t h i s .
60 # t h i s i n t e r n a l s t a t e needs t o correspond t o your i n i t f u n c t i o n :
61 p r i n t time
62 nparticles = internalStateIn [0]
63 particles = internalStateIn [1]
64 color = internalStateIn [2]
65 w_mu = i n t e r n a l S t a t e I n [ 3 ]

UC Berkeley, CE295
 79

66 w_sig = i n t e r n a l S t a t e I n [ 4 ]
67
68 ( x_m , y_m) = measurement
69 i f np . i s n a n ( x_m ) :
70 # ( here t o be f i x e d about dt )
71 r = 0.425
72 B = 0.8
73 v_r = np . random . normal ( 0 , 0 . 0 5 , ( n p a r t i c l e s , 1 ) )
74 v_w = np . random . normal ( 0 , 0 , ( n p a r t i c l e s , 1 ) )
75 v_b = np . random . normal ( 0 , 0 . 1 , ( n p a r t i c l e s , 1 ) )
76 v_gamma = np . random . normal ( 0 , 0 , ( n p a r t i c l e s , 1 ) )
77 x_p = p a r t i c l e s [ : , 0 ] [ np . newaxis ]
78 x_p = x_p . T
79 y_p = p a r t i c l e s [ : , 1 ] [ np . newaxis ]
80 y_p = y_p . T
81 t h e t a _ p = p a r t i c l e s [ : , 2 ] [ np . newaxis ]
82 theta_p = theta_p . T
83 x_new = x_p + 5 . 0 ∗ r ∗ pedalSpeed ∗ dt ∗ ( 1 . 0 + v_r ) ∗ ( np . cos ( t h e t a _ p ) ) ∗ ( 1 . 0 + v_w
)
84 y_new = y_p + 5 . 0 ∗ r ∗ pedalSpeed ∗ dt ∗ ( 1 . 0 + v_r ) ∗ ( np . s i n ( t h e t a _ p ) ) ∗ ( 1 . 0 + v_w
)
85 theta_new = t h e t a _ p + 5 . 0 ∗ r / B ∗ pedalSpeed ∗ dt ∗ ( 1 . 0 + v_r ) ∗ ( np . tan (
s t e e r i n g A n g l e ∗ ( 1 . 0 + v_gamma ) ) ) ∗ ( 1 . 0 + v_w ) / ( 1 . 0 + v_b )
86 p a r t i c l e s = np . column_stack ( [ x_new , y_new , theta_new ] )
87 else :
88 # ( here t o be f i x e d about dt )
89 r = 0.425
90 B = 0.8
91 v_r = np . random . normal ( 0 , 0 . 0 5 , ( n p a r t i c l e s , 1 ) )
92 v_w = np . random . normal ( 0 , 0 , ( n p a r t i c l e s , 1 ) )
93 v_b = np . random . normal ( 0 , 0 . 1 , ( n p a r t i c l e s , 1 ) )
94 v_gamma = np . random . normal ( 0 , 0 , ( n p a r t i c l e s , 1 ) )
95 x_p = p a r t i c l e s [ : , 0 ] [ np . newaxis ]
96 x_p = x_p . T
97 y_p = p a r t i c l e s [ : , 1 ] [ np . newaxis ]
98 y_p = y_p . T
99 t h e t a _ p = p a r t i c l e s [ : , 2 ] [ np . newaxis ]
100 theta_p = theta_p . T
101 x_new = x_p + 5 . 0 ∗ r ∗ pedalSpeed ∗ dt ∗ ( 1 . 0 + v_r ) ∗ ( np . cos ( t h e t a _ p ) ) ∗ ( 1 . 0 + v_w
)
102 y_new = y_p + 5 . 0 ∗ r ∗ pedalSpeed ∗ dt ∗ ( 1 . 0 + v_r ) ∗ ( np . s i n ( t h e t a _ p ) ) ∗ ( 1 . 0 + v_w
)
103 theta_new = t h e t a _ p + 5 . 0 ∗ r / B ∗ pedalSpeed ∗ dt ∗ ( 1 . 0 + v_r ) ∗ ( np . tan (
s t e e r i n g A n g l e ∗ ( 1 . 0 + v_gamma ) ) ) ∗ (
104 1 . 0 + v_w ) / ( 1 . 0 + v_b )
105 p a r t i c l e s = np . column_stack ( [ x_new , y_new , theta_new ] )
106
107 f _ z = np . z e r o s ( ( l e n ( p a r t i c l e s [ : , 0 ] ) , 1 ) )
108
109 w_1 = x_m − x_new − 1 . 0 / 2 . 0 ∗ B ∗ np . cos ( theta_new )
110 w_2 = y_m − y_new − 1 . 0 / 2 . 0 ∗ B ∗ np . s i n ( theta_new )
111 f _ x 1 = 1 . 0 / ( 2 . 0 ∗ np . p i ∗ 1 . 0 8 9 ∗ 1 . 0 8 9 ) ∗ np . exp ( − (w_1 − 0 . 0 ) ∗ ( w_1 − 0 . 0 ) / 2 . 0 /
(1.089 ∗ 1.089) )
112 f _ x 2 = 1 . 0 / ( 2 . 0 ∗ np . p i ∗ 2 . 9 9 ∗ 2 . 9 9 ) ∗ np . exp ( − (w_2 − 0 . 0 ) ∗ ( w_2 − 0 . 0 ) / 2 . 0 /
(2.99 ∗ 2.99) )
113 f_z = f_x1 ∗ f_x2
114 # t o be f i x e d here
115 alpha = 1 . 0 / np . sum ( f _ z [ : ] )
116 f _ z = f _ z ∗ alpha
117 c d f = np . cumsum( f _ z )
118
119 # resample d i s t r i b u t i o n samples
120 old_p = p a r t i c l e s . copy ( )

UC Berkeley, CE295
 80

121 p a r t i c l e s = np . a r r a y ( [ old_p [ np . argwhere ( c d f > np . random . uniform ( ) ) [ 0 , 0 ] ] f o r i i n

range ( l e n ( f _ z ) ) ] )
122 # adding n o i s e t o re −sampled p a r t i c l e s t o avoid i t from converging ( keep t h e amount o f
different particles )
123 v_xy = np . random . normal ( 0 , 0 . 3 , ( n p a r t i c l e s , 2 ) )
124 v_phi = np . random . normal ( 0 , 0 . 1 , ( n p a r t i c l e s , 1 ) )
125 v_sample = np . c o n c a t e n a t e ( ( v_xy , v_phi ) , a x i s =1)
126 p a r t i c l e s = p a r t i c l e s + v_sample
127
128 x = np . mean ( p a r t i c l e s [ : , 0 ] )
129 y = np . mean ( p a r t i c l e s [ : , 1 ] )
130 t h e t a = np . mean ( p a r t i c l e s [ : , 2 ] )
131 return x , y , theta , i n t e r n a l S t a t e
132
133 import numpy as np
134 import m a t p l o t l i b . pyplot as p l t
135 from estRun import estRun
136 from e s t I n i t i a l i z e import e s t I n i t i a l i z e
137
138 # provide t h e index o f t h e e x p e r i m e n t a l run you would l i k e t o use .
139 # Note t h a t using " 0 " means t h a t you w i l l load t h e measurement c a l i b r a t i o n data .
140 experimentalRun = 04
141
142 p r i n t ( ’ Loading t h e data f i l e # ’ , experimentalRun )
143 experimentalData = np . genfromtxt ( ’ data/run_ { 0 : 0 3 d } . csv ’ . format ( experimentalRun ) , d e l i m i t e r
= ’ , ’)
144
145
146 p r i n t ( ’ Running t h e i n i t i a l i z a t i o n ’ )
147 internalState = estInitialize ()
148
149 numDataPoints = experimentalData . shape [ 0 ]
150
151 #Here we w i l l s t o r e t h e e s t i m a t e d p o s i t i o n and o r i e n t a t i o n , f o r l a t e r p l o t t i n g :
152 e s t i m a t e d P o s i t i o n _ x = np . z e r o s ( [ numDataPoints , ] )
153 e s t i m a t e d P o s i t i o n _ y = np . z e r o s ( [ numDataPoints , ] )
154 estimatedAngle = np . z e r o s ( [ numDataPoints , ] )
155
156 p r i n t ( ’ Running t h e system ’ )
157 dt = experimentalData [ 1 , 0 ] − experimentalData [ 0 , 0 ]
158 f o r k i n range ( numDataPoints ) :
159 t = experimentalData [ k , 0 ]
160 gamma = experimentalData [ k , 1 ]
161 omega = experimentalData [ k , 2 ]
162 measx = experimentalData [ k , 3 ]
163 measy = experimentalData [ k , 4 ]
164
165 # run t h e e s t i m a t o r :
166 x , y , t h e t a , i n t e r n a l S t a t e = estRun ( t , dt , i n t e r n a l S t a t e , gamma, omega , ( measx , measy ) )
167
168 # keep t r a c k :
169 estimatedPosition_x [k] = x
170 estimatedPosition_y [k] = y
171 estimatedAngle [ k ] = t h e t a
172
173
174 p r i n t ( ’ Done running ’ )
175 #make sure t h e angle i s i n [− pi , p i ]
176 estimatedAngle = np . mod( estimatedAngle+np . pi , 2 ∗ np . p i )−np . p i
177
178 posErr_x = e s t i m a t e d P o s i t i o n _ x − experimentalData [ : , 5 ]
179 posErr_y = e s t i m a t e d P o s i t i o n _ y − experimentalData [ : , 6 ]
180 angErr = np . mod( estimatedAngle − experimentalData [ : , 7 ] + np . pi , 2 ∗ np . p i )−np . p i

UC Berkeley, CE295
 81

181
182 print ( ’ Final error : ’)
183 print ( ’ pos x = ’ , posErr_x [ − 1 ] , ’m’ )
184 print ( ’ pos y = ’ , posErr_y [ − 1 ] , ’m’ )
185 print ( ’ angle = ’ , angErr [ − 1 ] , ’ rad ’ )
186
187 ax = np . sum ( np . abs ( posErr_x ) ) /numDataPoints
188 ay = np . sum ( np . abs ( posErr_y ) ) /numDataPoints
189 ath = np . sum ( np . abs ( angErr ) ) /numDataPoints
190 s c o r e = ax + ay + ath
191 i f not np . i s n a n ( s c o r e ) :
192 # t h i s i s f o r e v a l u a t i o n by t h e i n s t r u c t o r s
193 p r i n t ( ’ average e r r o r : ’ )
194
195 print ( ’ pos x = ’ , ax , ’m’ )
196 print ( ’ pos y = ’ , ay , ’m’ )
197 print ( ’ angle = ’ , ath , ’ rad ’ )
198
199 # our s c a l a r s c o r e .
200 p r i n t ( ’ average s c o r e : ’ , s c o r e )
201
202
203 p r i n t ( ’ Generating p l o t s ’ )
204
205 figTopView , axTopView = p l t . s u b p l o t s ( 1 , 1 )
206 axTopView . p l o t ( experimentalData [ : , 3 ] , experimentalData [ : , 4 ] , ’ rx ’ , l a b e l = ’Meas ’ )
207 axTopView . p l o t ( e s t i m a t e d P o s i t i o n _ x , e s t i m a t e d P o s i t i o n _ y , ’ b − ’ , l a b e l = ’ e s t ’ )
208 axTopView . p l o t ( experimentalData [ : , 5 ] , experimentalData [ : , 6 ] , ’ k : . ’ , l a b e l = ’ t ru e ’ )
209 axTopView . legend ( )
210 axTopView . s e t _ x l a b e l ( ’ x−p o s i t i o n [m] ’ )
211 axTopView . s e t _ y l a b e l ( ’ y−p o s i t i o n [m] ’ )
212
213 f i g H i s t , a x H i s t = p l t . s u b p l o t s ( 5 , 1 , s h a r e x=True )
214 a x H i s t [ 0 ] . p l o t ( experimentalData [ : , 0 ] , experimentalData [ : , 5 ] , ’ k : . ’ , l a b e l = ’ t ru e ’ )
215 a x H i s t [ 0 ] . p l o t ( experimentalData [ : , 0 ] , experimentalData [ : , 3 ] , ’ rx ’ , l a b e l = ’Meas ’ )
216 a x H i s t [ 0 ] . p l o t ( experimentalData [ : , 0 ] , e s t i m a t e d P o s i t i o n _ x , ’ b − ’ , l a b e l = ’ e s t ’ )
217
218 a x H i s t [ 1 ] . p l o t ( experimentalData [ : , 0 ] , experimentalData [ : , 6 ] , ’ k : . ’ , l a b e l = ’ t ru e ’ )
219 a x H i s t [ 1 ] . p l o t ( experimentalData [ : , 0 ] , experimentalData [ : , 4 ] , ’ rx ’ , l a b e l = ’Meas ’ )
220 a x H i s t [ 1 ] . p l o t ( experimentalData [ : , 0 ] , e s t i m a t e d P o s i t i o n _ y , ’ b − ’ , l a b e l = ’ e s t ’ )
221
222 a x H i s t [ 2 ] . p l o t ( experimentalData [ : , 0 ] , experimentalData [ : , 7 ] , ’ k : . ’ , l a b e l = ’ t ru e ’ )
223 a x H i s t [ 2 ] . p l o t ( experimentalData [ : , 0 ] , estimatedAngle , ’ b − ’ , l a b e l = ’ e s t ’ )
224
225 a x H i s t [ 3 ] . p l o t ( experimentalData [ : , 0 ] , experimentalData [ : , 1 ] , ’ g − ’ , l a b e l = ’m’ )
226 a x H i s t [ 4 ] . p l o t ( experimentalData [ : , 0 ] , experimentalData [ : , 2 ] , ’ g − ’ , l a b e l = ’m’ )
227
228 a x H i s t [ 0 ] . legend ( )
229
230 a x H i s t [ − 1 ] . s e t _ x l a b e l ( ’ Time [ s ] ’ )
231 a x H i s t [ 0 ] . s e t _ y l a b e l ( ’ P o s i t i o n x [m] ’ )
232 a x H i s t [ 1 ] . s e t _ y l a b e l ( ’ P o s i t i o n y [m] ’ )
233 a x H i s t [ 2 ] . s e t _ y l a b e l ( ’ Angle t h e t a [ rad ] ’ )
234 a x H i s t [ 3 ] . s e t _ y l a b e l ( ’ S t e e r i n g angle gamma [ rad ] ’ )
235 a x H i s t [ 4 ] . s e t _ y l a b e l ( ’ Pedal speed omega [ rad/s ] ’ )
236
237 p r i n t ( ’ Done ’ )
238 p l t . show ( )

UC Berkeley, CE295
82

UC Berkeley, CE295
83

Nanjing Metro traveler entrance modeling and

minimization of system-wide wait time by optimizing
train distribution
Moon Ki Jung, Carlin Liao, Henry Teng, Johnny Wu

Abstract
Subway systems face two competing demands: from the passenger’s perspective minimiz-
ing travel time is paramount, while the operator is most concerned with reducing cost. In
this paper, we will implement and refine our methods on the Nanjing Metro system. Using
trip data harvested by smartcards, we create a viable mathematical model of system-wide
passenger wait time for the metro’s governing board to best optimize its limited resources
(trains) to meet both passenger and operational needs. Using the distribution of passengers’
station entrances between trains also developed in this paper, our optimization problem also
finds the ideal allocation of trains along multiple subway lines at any time of day for real
or simulated demand schedules in order to minimize total passenger wait times with the
resources available.

Introduction
Motivation and Background
As demonstrated in the first lecture of CE 295, many cities are growing rapidly, and
Nanjing is no different. The burden of effectively mobilizing that ever-growing population,
the tenth largest in China despite being close to the size of New York City, falls upon
transit systems like the Nanjing Metro, which served 977.4 million passengers in 2017. To
optimize the timetable of a metro system, we can take either the operator’s perspective or
the user’s perspective to find widely different consequences. From the operator’s point of
view, an ideal system would use the least amount of resources while still getting travelers to
their destinations regardless of how long that takes, while for a passenger, ideal would mean
getting from origin to destination in the least amount of time regardless of operations cost.
Therefore it is necessary to consider how to balance these two sets of priorities. Similar
to an approach taken from literature [1], this may be accomplished by minimizing the total
travel time while imposing operator constraints. In deciding which metro system to analyze,
our group was able to obtain two months’ worth of smart-card data from the Nanjing Metro
system. This smartcard data contains entries of every swipe-in and swipe-out of passengers

UC Berkeley, CE295
84

at every station on every core line in the Metro extant during the months of March and
April 2015. With such a high fidelity data set, our group decided to investigate how we
could model the system-wide wait time of a metro system and then how that wait time
could be minimized while satisfying operational constraints.
Considering the large size of both the Nanjing Metro and our data set, our group antic-
ipated challenges with processing and extracting useful information from our data set that
would best describe the passenger demands at stations in the system. In addition, transit
systems are inherently difficult to model and analyze, due to the variability in behavior of
their passengers. Many operational standards and protocols for the Metro, such as speed
limits, line constraints, train car specs, and even a fixed timetable were not made available
to our group, which required us to make several reasonable assumptions in developing a
mathematical model for our problem.
Originally our group had intended to implement our analysis on a dataset sourced from
the Shenzhen Metro by way of the Tsinghua-Berkeley Shenzhen Institute. Unfortunately,
due to complications with access and confidentiality, we were not able to secure the dataset
in time. Instead we were able to secure a near-identically featured dataset from the Nanjing
Metro system courtesy of Dounan Tang in Prof. Sengupta’s group which allowed us to
complete our analysis as planned.

Focus of this Study

Our project will develop a mathematical model of the system-wide wait time across the
Nanjing Metro, which we will use to optimize the distribution of trains across the system.
In doing so, we perform exploratory data analysis to investigate various assumptions of our
mathematical model, including the distribution of passenger entrances at origin stations.

Literature review
Thanks to the usage of smart card technology for public transit applications, there exists
an abundance of data upon which a myriad of analyses can be conducted. Many studies, [2],
focus on discerning passenger route choice and behavioral aspects from this data. However,
our team was more interested on using reducing wait-time across a system—a topic outside
the scope of these papers. Nonetheless, the methods of data processing and background
on transportation planning topics presented in these studies were invaluable in our project.
Select papers [1] [3] [4] investigate optimizing timetables for various subway systems, and it
is from these papers that we understood the context of the problem we wished to investigate,
namely developing improved timetables that lead to reduced passenger wait times. However,
these papers failed to account for the effects of limited resources on the operator’s side,
namely the number of trains available for dispatch. This then became our team’s ultimate
trajectory for the project.
However, with this question posed, many additional issues arose, specifically in the as-
sumptions we could make of our system, and how we could formulate our model. To this
end, we reached out to papers for techniques used by other researchers. For instance, we
adopted how a study [1] focused on only a subset of a system (select lines) in order to sim-
plify analysis. Select formulations (e.g. subdividing total travel time into its constituent

UC Berkeley, CE295
85

time components) and simplifications (e.g. uniform passenger arrival) used in papers [4] [5]
also gave inspiration to many of our assumptions and the chosen form of our model. The ex-
act details of our assumptions, formulation, and implementation are discussed in subsequent
sections of this report.

Technical Description
Due to the substantial scope of this project, the technical description consists of (1) a
description of our data set, (2) key assumptions & exploratory data analysis of our data set,
(3) the optimization problem formulation, (4) code implementation, and (5) optimization
results.

Dataset
The source dataset is a record of all smartcard trips made in the Nanjing subway system
from January through April 2015, which for this project we subsetted to the last week of
April. Originally provided as .mdb files with simplified Chinese headers, we used Microsoft
Access and Python to convert the tables from proprietary Microsoft databases to csv files
for easier manipulation and translated the table headers by hand.
The data is a collection of all tap-ins (entrances) and tap-outs (exits) by smartcards
in the subway system over the aforementioned four months in time. Note that this model
will thus not cover travelers using paper fares and as such assumes that trips made with
smartcards are an accurate representation of all trips in the system.
The headers of the dataset after translation are:

Card Id Card Id Long Card Type Fare Time Entry Or Exit Line Id
StationId Device Id

(We made an effort to match the initial translations used by Tang in the CEE 290I
homework, explaining the inconsistency in underscores versus camel case.)
Please refer to the section on Code Implementation for additional details on our data
processing.
Note that data .mdb and .csv files have not been included with this report due to their
large sizes. Contact us if you would like the original files to replicate our results.

Key Assumptions & Exploratory Data Analysis

For the purposes of our train distribution optimization problem, we have made three key
assumptions:

1. All trains have enough capacity to meet demand

2. Headways are constant on all lines in both directions during the time interval of our
analysis

UC Berkeley, CE295
86

3. Travelers enter stations to take the train in the same way regardless of how long or
short the headway between trains is (that is to say, they will always arrive in a uniform
distribution between trains, with the only variable affected by the headway being the
total number of people who arrive in each period)

None of these assumptions are made spuriously. The first arises from trains being the
limiting factor to better service, rather than train cars, so the analysis we make optimizes for
train frequency and not train length. The second comes from knowledge of scheduled train
frequency in the system, and the last from a simplifying assumption common in transporta-
tion planning where passengers are assumed to arrive in a uniform distribution [5]. Even so,
given our dataset, we have the capability to examine the last two assumptions, which we
will do so in our exploratory data analysis.
For the purposes of this analysis, we have chosen to only examine terminal stations.
Due to the nature of the dataset it is not immediately clear what direction train a traveler
intends to take except at terminal stations where there is only one option. At other stations,
inferring travelers’ route direction from their exit station would be necessary. This is done
during the optimization portion of this report, but is not necessary for the examination of
headway consistency and traveler entrance distribution.

Headway Verification
In order to verify the scheduled three-minute interval between trains on the Line 1’s
northern portion, we examine a histogram of travelers exiting the line’s northern terminus
over small time intervals. If trains were arriving every three minutes, the histogram should
exhibit corresponding spikes in exit counts approximately every 3 minutes as well as most
travelers exit the station as soon as the train arrives. In fact, this periodicity is exactly what
we see in Figure 1, empirically verifying that the headway on this line is three minutes long.

Figure 1: Histogram of travelers exiting the Line 1 north terminus over a weekday

Traveler Station Entrance Distribution

In line with the histogram of station exits, we also examined the same for station en-
trances. As shown in Figure 2, the periodicity is not as evident as with station exits, although
there still do seem to be spikes in entrance count that could suggest some travelers are adjust-
ing their arrival times to correspond with train arrivals, which would disprove the uniform
assumption.

UC Berkeley, CE295
87

Figure 2: Histogram of travelers entering the Line 1 north terminus over a weekday

In order to evaluate this, we collapsed the entire day of arrivals into a single three-minute
(180-second) period to correspond to their arrival times relative to the time until the next
train, and normalized the resulting histogram of arrivals per relative second so all bars would
sum to one (as in a probability distribution). Attempts to fit the resulting graph with known
probability distributions (seen in Figure 3) showed that the uniform distribution gave the
lowest sum squared error.

Figure 3: Normalized histogram of station entrances per second since last train and corre-
sponding fits

Normalization of the data over each 3-minute interval yielded very similar results (as in
Figure 4), so we conclude that the uniform distribution of station entrances is a realistic
assumption for Nanjing travelers. This conclusion parallels those of Luethi, who further
expands on passenger arrival modelling, [5].

UC Berkeley, CE295
88

Figure 4: Normalized cumulative distribution of station entrances per each 3-minute period
and corresponding fits

Train Distribution Optimization Problem Formulation

In developing our optimization problem, we will divide our formulation into two levels:
station-scale and multi-line scale. In doing so, we address the following points:

• Arrival behavior of passengers at a station in-between trains

• Relationship between headway on a line and the number of trains on a line
• Feedback of changing headway on the arrival of passengers at the station

We also carry forth the same three key assumptions discussed in the previous section.

Station-Scale Formulation
In order to calculate the total wait time for one duration of time between train departure
and arrival, we must account for all passengers on a platform, each of their arrival times, and
each passenger’s corresponding wait time. We first note that patrons are assumed to arrive
at the station uniformly over time as discussed in the prior section. We define a passenger
arrival function which will describe this rate of passenger arrivals. We then define a wait
function, a linearly decreasing function, in which the first passenger on the platform waits
the entire headway of the train, while the last passenger waits the least amount of time.
We define headway to be the span of time between trains (which in turn becomes the time
between train arrivals at a given station).
To obtain the total wait time on a single platform, we then multiply our passenger arrival
function by the wait function, and scale by the total number of passengers on the platform
during the time interval. This formulation is visually expressed in the following diagram.

UC Berkeley, CE295
89

Figure 5: Station-scale formulation

We define the passenger arrival function to be aHW , and the wait function to be wHW ,
with subscript HW denoting these two functions’ dependency on headway.
It is interesting to note that our passenger arrival function is equivalent to probability
density function of passenger arrivals, defined by a uniform distribution.
Our station-scale formulation is then defined as:

aHW ∗ wHW ∗ n (1)

where n is defined to be the total number of people on that platform when the incoming
train arrives. This value is extracted from our smart card data and is described Code
Implementation, an upcoming subsection of this report.
To illustrate how passenger arrivals change at a station over time, we provide the following
diagram. Note that we discretize time into increments equal to the headway on that line
(recall we assume that headways are constant along a line).

Figure 6: Program flow chart

Note how the assumption that train cars have infinite capacity allows us to reset our
cumulative number of passengers on a platform whenever a train arrives at the platform.
We index each increment of time by i ∈ 0, .., T where T is the final time duration we
are interested in. To find the total wait time across all time increments i, we merely add T
instances of Equation 1.

Multi-Line Formulation
Expanding Equation 1 beyond to multiple lines and stations, we note that aHW and wHW
do not change along a line, as we assume people will arrive at all stations along a line in the

UC Berkeley, CE295
90

same fashion, and these two functions are merely based on headway, which is assumed to be
constant along a line. In addition, we consider the bi-directionality of each line. in allocating
trains, both directions of a line share the same trains. We remedy this by adding another
index d ∈ {1, 2}, respectively denoting the forwards or backwards direction. Therefore, our
total wait time for a given line l, direction d, across all stations j, for all time increments i,
can be written as:
X X X X
TWT = aHWl ∗ wHWl ∗ nl,d,j,i (2)
l∈{1,...,L} d∈{1,2} j∈{1,...,Jl } i∈{0,...,T }

As defined for a system:

Figure 7: Schematic of multi-line system

See that aHWl and wHWl do not differ with direction of train, on one line. Special care
must be taken to calculate the appropriate nl,d,j,i term in the expansion, as we must organize
our data with respect to line, direction of journey, station, and time. With T W T fully
defined, we seek to minimize total wait time across the entire system with reference to HWl ,
which we understand influences every term in our cost function T W T . Again, we will discuss
how our n term is affected by headway in Code Implementation.
Our entire objective, however, is to optimally distribute a total number of trains. We
therefore must formulate a relationship between HWl and the number of trains on a line.
For simplicity, we use the relation:
sl
HWl = (3)
ntrainsl ∗ vtrainl
We divide the total length of a line by the speed of the trains on the line and the number of
trains on that line. We have assumed that all trains travel at the same speed on a line. vtrainl
is approximated from a timetable of first train arrivals at each station. Therefore, instead
of optimizing across HWl , we are actually optimizing across ntrainsl . With this relationship,
we can complete our optimization problem:

UC Berkeley, CE295
91

X X X X
minimize TWT = aHWl ∗ wHWl ∗ nl,d,j,i
ntrains−l
l∈{1,...,L} d∈{1,2} j∈{1,...,Jl } i∈{0,...,T }
sl
subject to HWl = , l ∈ {1, ..., L} Headway- # Trains Relation
ntrainsl ∗ vtrainl
X
ntrainl ≤ ntraintotal , l ∈ {1, ..., L} Total # of trains
l∈{1,..,L}

ntrainsl > 0, l ∈ {1, ..., L} Non-negative, non-zero # of trains on any line

HWlmin ≤ HWl ≤ HWlmax , l ∈ {1, ..., L} Min. & max. wait times on a line
ntrainsl ∈ Z Integer # of trains

nl,d,j,i is from our smartcard data, sl is found from a map of the system, vtrainl is calculated
from a timetable, ntraintotal is the total number of trains we make available to the system,
and HWlmin & HWlmax are limits on the headway that any line will experience, constraints
imposed by us to effectively enforce speed limits and a minimum level of service, respectively.
We note that this optimization problem is an integer problem, as our decision variables,
ntrainsl are restricted to whole numbers of trains. To solve this problem, we use a brute-force
search method in which we will calculate T W T for all possible combinations of ntrainsl across
l lines, and select the combination that arrives at the minimum T W T .

Remark on Feedback of headway on arrival of passengers

We must finally consider that if we change the value of ntrainsl , HWl will also change
(by way of Equation 4), which will then affect every term in T W T . In accounting for this
dependency, we make our third key assumption (refer to the section, Key Assumptions
& Exploratory Data Analysis).
This is a crucial assumption as we have no means of modeling passenger behavior with
merely our smartcard data (which does not have any information on the corresponding
headways during those times).
With this assumption, we will merely discretize our data evenly with the new HWl , and
calculate aHWl , wHWl and nl,d,j,i based on the new increments of data. This idea is detailed
further in Code Implementation.

Code Implementation
To improve understanding of our code implementation, please refer to the Program
Flowchart, Figure ?? in the Appendix.

Data Processing
The CSV data provided to us came in the form of rows of metro card swipe in/out in
one day. The columns consisted of the a short passenger ID, a long Passenger ID (includes
demographic information), card type, fare, time stamp of swipe in/out, entry (1) or exit (2),
line, station, and device ID. In order to determine T W T , we had to write a program that

UC Berkeley, CE295
92

would calculate our nl,d,j,i on each station, on a specific line, in a specific direction, and at a
specific time period. The inputs to the program would be the headways of the lines we were
interested in (in this case, line 1 and line 2).

Extraction of Origin-Destination (OD) Pairs

The objective of the program is to calculate the total wait time of passengers on each line,
given HWl inputs to those lines (1 and 2 for example). As we were only focused on trains
operating on lines 1 and 2, we exclude entries of passengers entering/exiting stations on lines
1 and 2 to either transfer into a different line or transferred into lines 1 and 2 from a different
line. This specification was made because focusing on passengers travelling exclusively on
line 1 or line 2 would allow us to readily determine the direction they would be commuting
on that line by comparing an OD pair’s station of entry and exit.
All data entries were ordered by time of card swipe entry/exit. Then, the data entries
were filtered and parsed by line and entry or exit. We then created OD pairs by matching
the occurrences of a unique passenger ID in the data. If a pair was successfully found, then
the station ID at entry and exit were compared to determine which direction on that line
the passenger was travelling in.

Computation of T W T Cost Function

Our program at this point generates a (2-by-N) cell array, matching the N number of
stations on a specific line. The 2 rows indicate the direction of travel along that line. Each
cell would contain all the entries of passengers (ordered by time) who entered that station
and made a OD pair in direction 1 or 2. Given an input HWl (let’s say 6 mins on Line 1),
the code iterates across each of the (2-by-N) cells, and calculates the number of entries that
arrive in successive 6 minute increments in each cell until a time horizon T , thus finding each
nl,d,j,i . Assuming a uniform passenger arrival distribution, the code multiples each nl,d,j,i by
HWl
2
, the average wait time at that station. Summing all n ∗ HW2
l
terms at each station, and
then summing across all stations gives us the wait times on a line. This method is repeated
for other lines, and then the wait times are summed across all lines, obtaining T W T

Exhaustive Search

Figure 8: Exhaustive Search Method: Creating viable combinations of train distribution

along two train lines

UC Berkeley, CE295
93

Recall the program calculating T W T does so for inputs HWl on desired lines l. Imple-
menting the exhaustive search method, we generate all combinations of N trains available
for dispatch across k lines, which are directly related to HW by the headway to number
of trains equality (see Multi-Line Formulation). Combinations that violate any constraints
are ejected,Pand a T W T is generated for each viable HWk . P It is important to note that
we enforce ntrainsl ,2 = ntrainstotal instead of the inequality ntrainsl ≤ ntrainstotal . This
is because having the maximum number of trains in the system would always result in the
minimal total time wait in the system, as more trains in the system always equated to lower
wait times for passengers.

Optimization Results and Demand Simulation

We can now execute our problem using the Nanjing smartcard data.
For all subsequent analyses, let us consider ntraintotal = 50, to be distributed between
lines 1 and 2. To visualize these two lines spatially, please refer to a map of the Nanjing
Metro system, included in the Appendix.
We recall that T W T can be calculated across any time horizon T . We choose to optimize
train distribution across a daily horizon and a 3-hour horizon.

Daily Analysis
In this analysis, we compute T W T across an entire day. We visualize the passenger
demand across a Sunday and Monday with the following plots for lines 1 and 2.

(a) Monday, Line 1 (b) Monday, Line 2

(c) Sunday, Line 1 (d) Sunday, Line 2

Figure 9: Passenger demand, Monday & Sunday, Lines 1 & 2, North to South

UC Berkeley, CE295
94

On both lines, the Monday passenger demands exhibits a clear workday pattern, with
two peaks corresponding to morning and evening rush hours, while the Sunday passenger
demands lack any clear pattern. With this in mind, we present our optimization results:

Table 1: Results of Daily Analysis, ntraintotal = 50

Monday Sunday
ntrainL1 ntrainL2 T W T (min) ntrainL1 ntrainL2 T W T (min)
25 25 796,269 25 25 481,183
26 24 791,539 26 24 479,626
27 23 789,331 27 23 479,605
28 22 789,654 28 22 481,140
29 21 792,577 29 21 484,4285
30 20 798,220 30 20 489,128

The bolded entries correspond to the optimal train distribution. We note they are equiv-
alent, despite the passenger demands on both lines between Monday and Sunday being very
different. We therefore move to calculate T W T across 3-hour chunks in order to capture
more variation in the passenger demand through a day.

3-Hour Analysis
On the same passenger demands displayed in Figure 9, we present the results:

Table 2: Results of 3-Hour Analysis, ntraintotal = 50

Monday Sunday
Time Block [HR] ntrainsL1 ntrainsL2 Time Block [HR] ntrainsL1 ntrainsL2
0-3 25 25 0-3 27 23
3-6 28 22 3-6 28 22
6-9 28 22 6-9 28 22
9-12 28 22 9-12 26 24
12-15 28 22 12-15 26 24
15-18 28 22 15-18 26 24
18-21 28 22 18-21 28 22
21-24 28 22 21-24 27 23

The train allocations are varied through the day, as we consider smaller chunks of time,
allowing the optimization problem to capture fluctuations in passenger demand. This is
particularly apparent on Sunday, in which there lacks a regular pattern in passenger demand
through the day.

Artificial Demand Simulation

We now consider an artificial demand, in which we purposefully impose a large demand
at a given station, and determine the best allocation of trains to meet that demand. In

UC Berkeley, CE295
95

this situation, we impose a load of 10,000 total passengers entering Station 15 on Sunday
between 5 and 8 PM on Line 1. We visualize this load in the following comparison.

(a) Sunday, Line 1 (b) Sunday, Line 1, Artificial Demand

Figure 10: Passenger demand, Sunday, Lines 1 & 2, North to South

The 3-hour analysis on these two different passenger demand schemes results in the
following train allocations.

Table 3: Results of 3-Hour Analysis, ntraintotal = 50, Artificial Demand

Sunday Simulated Sunday
Time Block [0-24] HR ntrainsL1 ntrainsL2 Time Block [0-24] HR ntrainsL1 ntrainsL2
0-3 27 23 0-3 27 23
3-6 28 22 3-6 28 22
6-9 28 22 6-9 28 22
9-12 26 24 9-12 26 24
12-15 26 24 12-15 26 24
15-18 26 24 15-18 29 21
18-21 28 22 18-21 28 22
21-24 27 23 21-24 27 23

We see that the optimization problem shifted three trains from Line 2 to Line 1 during
the 3 hours that the load was applied on Line 1. It is illuminating to see how this reallocation
affects the wait times on stations along lines 1 and 2 during that 15-18 timeslot. We compute
those total wait times per station, and present the results below:

UC Berkeley, CE295
96

(a) Stations on Line 1 (b) Stations on Line 2

Figure 11: Passenger Demand, Sunday, Lines 1 & 2, North to South

We note that the additional trains reallocated to Line 1 on the “simulated” Sunday, wait
times across Line 1 are overall lower than a “regular” Sunday, except for Station 15 (where
the 10,000 extra passengers were injected). On Line 2, which received fewer trains than
a “regular” Sunday, wait times were slightly higher across the board on the “Simulated”
Sunday. These results are to be expected due to the reallocation of trains.

Discussion
Our project results indicate that our mathematical model can respond to the fluctuating
system demands that exist in a modern public transit system. Our formulation is flexible
enough to analyze the problem at any time horizon desired, and is therefore powerful, as it
allows one to allocate trains given that they know the anticipated passenger demand, acting
within our model assumptions. Indeed, our project touches on an aspect of a growing issue
in the transportation field; how to best allocate limited resources to meet demand while
reducing costs.
There exist many ways to extend and improve this project. By happenstance, during the
time period of data given, a new metro line opened. This inspired authors Max Jung, Carlin
Liao, and others in Prof. Joan Walker’s CE 264 to study the effects this opening has on
passenger behavior. Such a behavioral aspect could also improve the accuracy of our model;
we currently assume that passengers will arrive at stations in the same fashion regardless
of headway, for instance. Following this thought, longer and variable headways may affect
passengers’ decision making in when they arrive at a station. An extension to the Traveler
Station Entrance Distribution section of our analysis is explored by author Carlin Liao in
Prof. Roy Dong’s IEOR 290.
This project also illuminated the limitations of optimization and modelling techniques
in engineering. The more accurate a optimization problem becomes, the more complex the
model becomes, thereby limiting available solution methods. Our model ultimately required

UC Berkeley, CE295
97

an exhaustive search method. For a larger sized problem (perhaps if we considered more
lines), such a method would become too computationally expensive. The computational
aspect of our project is being investigated by authors Max Jung and Henry Teng in Prof.
Raja Sengupta’s CE 290I.
Another improvement would be to improve our model of passenger wait time. As learned
through literature review, we must add terms accounting for passengers left on the platform
due to limited train capacity. Also, as our methods are not specific to the Nanjing Metro
system, further studies could also replicate our procedure on similar smartcard data from
BART, or even the Shenzhen metro dataset we had originally anticipated.

Summary
Our project aimed at combining the operator’s and passenger’s perspective in allocating
trains across lines of a metro system. In particular, we hoped to reduce operator costs (the
number of trains deployed), while also reducing passenger wait time. This was accomplished
by formulating a cost function that accounted for the wait time of passengers across multiple
lines of a subway system, and related the number of trains on a line to the headway on that
line. This formulation allowed us to optimize the cost function given additional operational
and demand constraints to determine the best distribution of trains to reach a minimum
system-wide passenger wait time. To ensure the validity of our problem, exploratory data
analysis was done to study our problem assumptions, namely passenger arrival and head-
way assumptions. Our optimization problem was solved given various passenger demand
schedules (from Nanjing Metro smartcard data), and we were able to successfully account
for anticipated passenger demands with our model by allocating trains from lesser demand
lines to higher demand lines during select hours of the day.

References
[1] P. Shang, R. Li, and L. Yang, “Optimization of urban single-line metro timetable for
total passenger travel time under dynamic passenger demand,” Procedia Engineering,
vol. 137, 2016. [Online]. Available: https://www.sciencedirect.com/science/article/pii/
S1877705816002587
[2] J. Zhao, F. Zhang, L. Tu, C. Xu, D. Shen, C. Tian, X.-Y. Li, and Z. Li, “Estimation
of passenger route choice pattern using smart card data for complex metro systems,”
IEEE Transactions on Intelligent Transportation Systems, vol. 18, no. 4, 2017. [Online].
Available: https://arxiv.org/pdf/1605.08390
[3] L. Sun, J. G. Jin, d.-H. Lee, K. W. Axhuasen, and A. Erath, “Demand-driven timetable
design for metro services,” Transportation Research Part C: Emerging Technologies,
vol. 46, 2016. [Online]. Available: https://www.sciencedirect.com/science/article/pii/
S0968090X1400182X
[4] Y. Wang, B. De Schutter, T. van de Boom, B. Ning, and T. Tang,
“Real-time scheduling for single lines in urban rail transit systems,” IEEE

UC Berkeley, CE295
98

International Conference on Intelligent Rail Transportation, 2013. [Online]. Available:

https://www.sciencedirect.com/science/article/pii/S1877705816002587

[5] M. Luethi, U. Weidmann, and A. Nash, “Passenger arrival rates at public transport
stations,” Transportation Research Board 86th Annual Meeting, 2007. [Online]. Available:
https://trid.trb.org/view/801220

Acknowledgments
Thanks to Dounan Tang and by extension the Nanjing Metro Group Company for pro-
viding us the dataset for use in our analysis; Kai Zheng and Zhe Zhou of Tsinghua University
for providing our initial inspiration with the Shenzhen system; and Hualiang Teng of the
University of Nevada, Las Vegas for initial suggestions on our project.
And, of course, a very special thanks to Scott Moura, Bertrand Travacca, and the rest of
CEE 295 for guiding and supporting our project, as well for awarding us the People’s Choice
Award for best overall project this year.

About the authors

Carlin Liao is graduating with his M.S. in Civil Engineering (Civil Systems) from UC
Berkeley and will be pursuing his Ph.D. in Transportation at UT Austin.

Johnny Wu is graduating with his B.S. in Energy Engineering from UC Berkeley.

Henry Teng will be completing his M.S. in Civil Engineering (Civil Systems) at UC Berke-
ley with Prof. Khalid Mosalam.

Moon Ki “Max” Jung is graduating with his M.S. in Civil Engineering (Civil Systems)
from UC Berkeley. He completed research with Prof. Stephen Glaser.

UC Berkeley, CE295
99

Appendix
MATLAB scripts for the train distribution optimization and Jupyter notebooks for the
station entrance analysis have not been included with this report. If you would like the
former, please contact the authors, while the latter can be found at
https://github.com/CarlinLiao/smartcard-analysis

Figure 12: Map of the Nanjing Metro system, present day

UC Berkeley, CE295
100

UC Berkeley, CE295
101

Minimizing Ridesplitting Vehicle Miles Traveled

Caroline Neaves, Kaylee Homolka, Riley James, Shanay Kapadia

Abstract
To achieve emission reduction goals, it is necessary to understand the vehicle miles trav-
eled (VMT) and emissions impacts of ridesourcing. Recent studies suggest that ridesourcing
yields an increase in VMT and therefore an increase in emissions. However, ridesplitting
may be a way to preserve the advantages of ridesourcing while negating the harmful envi-
ronmental effects. This study aims to quantify the difference in VMT between ridesplitting
and ridesourcing. To achieve this, we compared data from ridesplitting trips in Berkeley,
CA, to an optimization program built to model rides in the same area. The results show
that ridesplitting saves an average of 0.23 miles per vehicle miles travelled compared to single
rides. This value could be increased by 35% if drivers followed an optimal route to minimize
emissions.

Introduction
Motivation and Background
To keep global warming between 2 and 3 degrees Celsius, we need to stabilize the atmo-
spheric concentration of carbon dioxide at 450 parts per million. To accomplish this goal,
drastic cuts are needed in every sector of the economy. According to the United States
Environmental Protection Agency, transportation accounts for 28% of US greenhouse gas
(GHG) emissions, and 40% of GHG emissions in California. Currently, emissions projec-
tions for transportation are expected to grow. At the same time, the transportation sector
is undergoing significant changes. Transportation services such as on-demand ridesourcing
are becoming increasingly popular all over the world.
Since the introduction of Uber in 2009, ridesourcing services have increased significantly.
Ridesourcing allows users many benefits. The convenience of door-to-door transportation is
very enticing in many situations. A study done by A. Henao (2017) found that both car
owners and non-drivers use ridesourcing for various reasons. People that own cars choose
to use a ridesourcing service when they are out of town, when they are going to/from the
airport, and for social trips.The primary reasons drivers may favor ridesourcing for social
trips are difficulty of finding parking or to avoid drinking and driving (Henao, 2017). People
that do not own cars regularly use ridesourcing to commute to work or school as a result of
lack of satisfactory public transportation or time constraints (Henao, 2017).

UC Berkeley, CE295
102

On the other hand, the adoption of ride-hailing services may be harmful from an en-
vironmental perspective. The environmental effect of ridesourcing depends on the type of
transportation passengers would have chosen if ridesourcing were not available. Studies show
that this is very dependent on location. For example, Hampshire et al. (2016) completed a
study in Austin, Texas where driving is very prevalent. If ridesourcing were not available,
the study revealed that 45% would have driven. Conversely, Rayle et al. (2014) conducted
a study in San Francisco, California where driving is far less common. The San Francisco
study showed that only 7% of the trips would have been taken by driving had ridesourcing
been unavailable. Instead, this study shows that 30% of ridesourcing trips are substitut-
ing for public transit and 9% walking or biking. In locations such as Austin, ridesourcing
may not have significant environmental repercussion at this time. However, in locations
such as San Francisco, where ridesourcing trips are replacing much lower emission forms of
transportation, there must be an increase in emissions.
Another major concern related to the popularity of ridesourcing is induced demand.
Induced demand is a phenomenon where more of a good is consumed after supply increases.
In the case of ridesourcing, this means that passengers choose to take rides they otherwise
would not have made because of the availability of these services. Rayle et al. (2014)
reported that 8% of respondents said that, had ridesourcing not been available, they would
not have made the trip at all. Similarly, Henao (2016), reported that in Denver and Boulder,
Colorado, there was a 12% induced demand effect. While these values may seem small,
they are not inconsequential. Any trip from induced demand results in additional emissions,
which is very concerning from a climate change perspective.
An important metric of ridesourcing are out-of-service miles, also referred to as deadhead-
ing. Deadheading is the miles driven by ridesourcing drivers while driving to the passenger
pickup point, waiting for a passenger request, or seeking passengers in a higher demand
area (Shaheen et al., 2018) (Ngo, 2015). These additional miles driven are potentially in-
creasing fuel consumption, emissions, and congestion. While there are not many studies on
deadheading, the few results show very significant numbers. Henao (2016), found that, on
average, ridesourcing drivers travel an additional 69.0 miles in deadheading for every 100
miles they are with passengers. A study completed in San Francisco found that 20 percent
of ridesourcing VMT are deadheading miles (SFCTA, 2017). While there are not many
studies examining deadheading, the results show very significant numbers that should not
be overlooked.
There have been very few studies that attempt to quantify the overall Vehicles Miles
Traveled (VMT) impacts of ridesourcing services. Presently, VMT change is difficult to
measure because ridesourcing companies are reluctant to share data and there are potential
modal shift implications (Shaheen et al., 2018). However, as a result of mode substitution,
induced demand, and deadheading, it may be reasonable to assume that ridesourcing in-
creases overall VMT in most locations. Schaller (2017) provided a preliminary calculation
of 3.5 percent increase in citywide VMT. If left unmitigated, congestion and emissions are
expected to grow as a result of this service (Clewlow et al.)
In 2014, some ridesourcing services began offering ridesplitting options. Ridesplitting
involves matching passengers traveling in the same direction in order to share the ride and
the cost. The primary motivation for users to choose this option over a private ride is the
reduced cost. However, the social and environmental benefits may be reason enough for some

UC Berkeley, CE295
103

riders. Conversely, there are many factors that discourage the use of ridesplitting. Excess
travel-time is considered the main deterrent. If a rider is in a hurry, they may be willing to pay
the extra few dollars to get to their destination as quickly as possible. Additionally, personal
preferences such as comfort travelling with strangers may prevent users from choosing the
ride sharing option. Ridesplitting options are currently only offered in some locations.
In this study, we focus on the ridesplitting options offered by Lyft and Uber. Lyft Line
and UberPOOL operate very similarly. A passenger requests a Line or Pool for up to two
people for a discounted price. The ride may or may not be matched with other riders
travelling in the same direction. Uber Express Pool is another ridesplitting options offered
by Uber. For an Express pool, a passenger requests a shared ride for a discounted cost
and is instructed to walk a short distance to a pickup point. At the end of the shared ride,
passengers are dropped off a short distance away from their destination. Essentially, Express
Pool is a ridesplitting service with more efficient pickups and dropoffs.
Ridesplitting systems pose a potential strategy for alleviating the harmful environmental
effects while preserving the societal benefits. The intent of the rideshare system is to bring
together travelers with similar itineraries and time schedules on short-notice (Agatz, Niels,
et al.). Studies suggest ridesplitting could reduce the number of cars on the road, while
simultaneously increasing the utilization of available seat capacity. This solution has the po-
tential to significantly reduce congestion, vehicle emissions, fuel use, auto reliance and travel
costs (Shaheen et.al, 2018). Ultimately, increasing occupancy rates can lead to improved air
quality and reduced carbon dioxide emissions - a major contributor to global climate change.
There are currently few studies of the impacts of new service models such as ridesplitting.
Dynamic ridesourcing is an automated system that facilitates drivers and riders to share
trips on very short notice. Optimization of ridesplitting can focus on three specific objectives:
1. Minimizing system-wide vehicles-miles, reducing pollution and congestion 2. Minimizing
the system-wide travel time, an important convenience consideration 3. Maximize the num-
ber of customers, maximizing revenues for private ridesourcing providers (Agatz, Niels, et
al., 2012)

Relevant Literature
Given that ride-sharing companies have only risen to prominence over the last several
years, research on the environmental impacts of ride-sharing is limited. In addition, many
of these studies are small-scale and region specific. While there are not many aggregate
studies on how ride-sharing can reduce emissions, several case studies of specific cities have
been done. For example, a study of ride sharing use in Beijing found that using ride-sharing
instead of a personal vehicle could lead to an overall reduction in CO 2 emissions from the
transportation sector, along with reductions in emissions of other air pollutants like NOX
and SOX (Yu et al., 2017). Furthermore, a study of vehicle use in the city of Changsha,
China found that ride-sharing could reduce CO 2 emissions by up to 4 tons a day when
compared emissions from using a personal vehicle (Jalali et al., 2017). These studies may
not be applicable to other regions with different characteristics, and therefore more research
is necessary, especially in US cities.
Additionally, several case studies have used dynamic ride-sharing models to investigate
the carbon emissions impacts of ride-sharing systems. One such study from Sichuan Uni-

UC Berkeley, CE295
104

versity used an integer programming model to maximize greenhouse gas reductions from
ride-sharing, and used this model to calculate both overall emissions savings and poten-
tial increases in savings from expanding ride sharing participation. The study found that
ride-sharing could reduce emissions in two main ways: reducing the total number of vehicle
miles traveled by sharing rides between passengers with overlapping routes, and decreasing
the overall number of cars on the road, which decreased traffic and congestion (Ma et al.,
2016). Similarly, another study using a mixed integer programming model to minimize total
vehicle miles traveled in ride-sharing populations found that reductions in total VMT can be
increased by increasing driver flexibility and the density of participants (Armant & Brown,
2014).
In contrast, some studies have found that the increased use of ride-sharing has not nec-
essarily been effective at reducing the environmental impacts from driving. For example,
a study investigating ride-sharing patterns in California found that Uber and Lyft vehicles
are on average less fuel efficient than taxis, and therefore may not result in an overall re-
duction in carbon emissions (Wagner 2017). In addition, a case study of Paris found that
while increasing vehicle occupancy through ride sharing had the potential to significantly
reduce transportation CO2 emissions, various rebound effects could lead to decreases in these
emissions savings in the long run (Yin et al., 2017).

Focus of this Study

This study aims to quantify the difference in VMT as well as associated emissions as a
result of ridesplitting as opposed to individual ridesourcing.

Key Contributions
Due to the conflicting findings of existing studies, further research on the environmental
impacts of ride-sharing is needed. This study provides a case study of the VMT and emissions
associated with ridesourcing versus ridesplitting for Berkeley, CA.

Data Identification and Sourcing

Data is collected on Uber Pool, Uber Express Pool, and Lyft Line rides using the re-
searchers own ride histories and the ride histories of other UC Berkeley Students. For each
ride, the cost of the trip, the company and type of trip, the time and date, and the number of
passengers in the riders party is recorded. The main riders pick-up and drop off points along
with total trip distance and duration is measured. Additionally, the number of additional
passengers and their pick up and drop off points is recorded, if applicable. This is used to
calculate the overlap between the main riders route and the routes of additional passengers.
Finally, in the case of Uber Express Pool, the distance the rider walked to get to the Express
pick-up point is measureds. There are several limitations associated with this data. First,
this study uses ridesharing data from only UC Berkeley students, meaning our results cannot
necessarily be applied to rideshare customers as a whole. In addition, Uber and Lyft do not
provide detailed data on the routes of additional passengers, so overlap between the main
rider and additional passengers has to be estimated instead of directly measured. To collect

UC Berkeley, CE295
105

this data, it was necessary to note the approximate address of additional passengers pick up
and drop off points as the ride was in progress. If respondents did not record these data, the
ride data was unusable. Additionally, our current model is designed to analyze rides only in
a section of Berkeley. As a result, many of our collected data points were unusable, leaving
our sample size at 22 rides. Finally, we do not have any driver data, so we are not able to
calculate differences of deadheading between individual ridesourcing rides and ridesplitting.
For future research, we would aim to collect additional data points, and possibly collect more
precise driver and rider data directly from Uber or Lyft.

Methodological Approach
To minimize VMT and emissions, an optimization program was implemented over a
set of nodes located at cross streets in Berkeley. Each of these nodes has an associated
longitude, latitude, and elevation. The grid boundaries were chosen to include common
Berkeley destinations, including UC campus, grocery stores, and restaurants. The grid was
simplified to include only intersections that were roughly one block (0.1 mi) apart or more.
One ways streets and dead ends were accounted for when calculating optimal routes. An
adjacency matrix was built to convert the grid into a logical array. For a given node, adjacent
nodes were marked with a 1, and all other nodes left as 0. The grid of Berkeley was simplified
to include 174 nodes, each at an intersection, as seen in Figure 1. Data is collected from
group members and classmates based on their recent Uber trip history. Only rides beginning
and ending within our defined grid were considered. The goal is to obtain the VMT saved
by taking an Uber Pool vs. an individual ride. Respondents recorded their own pickup and
dropoff points, as well as those of any riders on their trip.

UC Berkeley, CE295
106

Figure 1: Grid of cross streets in Berkeley, CA

The distance between the nodes is found using the Haversine formula. Details on how
the Haversine formula works can be found in Appendix A.
Using this grid and optimization techniques, the route that results in the least vehicle
miles traveled was found. In addition, dynamic programming was used to find the shortest
route between pickup and drop of points, and integer linear programming to find the optimal
path between them. Dijkstra’s algorithm, an implementation of dynamic programming, was
used to find the distance between nodes. A brief explanation of how this algorithm works
can be found in Appendix B.
The formulation for our integer linear program is as follows (Nicholas et al., 2016):
P
M inimize : d= ij Cij Xij (1)

Subject to : Bi +tij −Bj ≤ M (1−Xij ) (2)

Bj −Bi ≤ M Xij −tij (3)

UC Berkeley, CE295
107

Bi ≤ Bi+2 (4)
(
1 if traveling f rom node i to node j
Xij = (5)
0 otherwise
B1 = 0 (6)
P
j Xij = 1 f or i ∈ driver, origin (7)
P
j Xij = −1f orj ∈ origin, destination (8)

where d is the shortest distance, Cij is the cost, or distance, to go from node i to node j,
Xij is the decision variable which is equal to 1 if the optimal route passes from node i to j,
Bi is the time of arrival at node i, tij is the time to travel from node i to node j, and M is a
large number used in the ”big M method,” with Bi and Xij being the optimization variables.
The objective function, in equation 1, minimizes the cost-to-go between two nodes over
the full duration of an origin-destination trip, where Xij is assigned a value of 1 if the link
between nodes i and j is taken, and 0 if it is not, as shown in equation 5.
The function intlinprog was used to solve this in MATLAB. Integer linear programs
require the inputs in the form (MATLAB 8.0, 2018):


 x are integers

Ax ≤ b
minx f T x subject to (9)


 A eq x = beq
lb ≤ x ≤ ub


Using the formulation shown in equation 9, x1 to x25 correspond to X11 to X55 and
x26 to x30 correspond to B1 to B5 . The A matrix and b vector are created to encode
equations 2, 3, and 4, while the Aeq matrix and beq vector encode equations 6, 7, and 8. The
integer requirement of the decision variable X is implemented by the integer linear program
formulation.
To find the minimum VMT for the trip, first Dijkstras algorithm is used to find the
shortest path between the driver node, the origin nodes, and the destination nodes. For
the ridesplitting case, the above integer linear program is used to find the shortest route,
while visiting the nodes in the optimal order. The path is then recorded and plotted, along
with the associated VMT. For the single ride case, the path found by Dijkstras algorithm
between the riders origin and destination node is the optimal path, and is plotted along with
the optimal ridesplitting path.

Analysis
For this analysis, the primary variable of interest is vehicle miles traveled (VMT). Using
Dijkstras algorithm and integer linear programming, the minimal ride distance is found for
both ridesplitting and single ride cases. The results from the optimization program are used
to calculate the difference in ridesplitting and single rides. This difference is calculated using

UC Berkeley, CE295
108

equation 10.

∆dsingle = (driderA + driderB ) − d (10)

where ∆dsingle is the difference in distance for ridesplitting and single rides, and driderA
and driderB are the optimal distances for two riders taking separate rides. Also calculated
from the results of the optimization program is the difference between the optimal route and
the actual route that was taken. The miles saved for single and split rides are calculated
using the following formulas:

∆dactual = dactual −d (11)

∆dactual,single = (driderA +driderB −dactual ) (12)

P P
ractual = ∆dactual / dactual (13)
P P
ractual,single = ∆dactual,single / (driderA +driderB ) (14)
P P
rsingle = ∆dsingle / (driderA + driderB ) (15)

where dactual is the actual distance taken on the ridesplitting trip, d is the calculated
optimal distance for the split ride, and ∆dactual is the difference between these values. That
is, equation 11 calculates the miles that would be saved by taking the optimal route for
a split ride. ∆dactual,single is the difference between the optimal and actual distances for
a single rider, calculated in equation 12. ractual is the savings for ridesplitting passengers
taking the optimal route compared to the actual route, calculated in equation 13. Equation
14 calculates the savings between the split and single rider trips (ractual,single ), and equation
15 calculates rsingle , which is the savings for a single rider taking the optimal route. When
calculating these savings for Uber Express Pool trips the walking distance was included in
this difference.
In addition to the VMT differences between trips, the emissions saved by taking a shared
trip is also calculated. The saved trip emissions, net , is found using equation 16:

net = (dnet )(Egas )/F E (16)

where Egas is the the CO2 emissions associated with a burning a gallon of gasoline. The
two vehicles being analyzed in this case will be the Toyota Camry with a fuel efficiency (F E)
of 29 miles per gallon and the Toyota Prius with an F E of 54 miles per gallon (Toyota 2018).
These two vehicles were chosen for emissions analysis because they were the most common
vehicles used in our data set. According to the Energy Information Administration, there are
approximately 19.6 pounds, or 8.89 kilograms, of CO2 produced when one gallon of gasoline
is burned (EIA 2017). Therefore, using equation 16, Egas = 8.89kg/gallon.

UC Berkeley, CE295
109

Results and Discussion

To find the difference in distance for ridesplitting vs. single rides, the shortest distance
for each single ride is first found. Then, the difference in distance is calculated between
the sum of the distances of the two individual trips and the ridesplitting distance. Table 1
summarizes the optimal trip distance for a shared ride, along with the distance traveled if
each trip had been made up of two single rides (A and B) instead of a split ride. This table
also gives the combined distance of the two individual trips, and the actual distance of the
ridesplitting trip.

Table 1: Results of path optimization for each of the collected trips

UC Berkeley, CE295
110

By comparing the Optimal Distance traveled when the two rides are combined into
ridesplitting with the Total Distance traveled when each ride is conducted separately, it is
clear that ridesplitting reduced the overall VMT. However, the actual distance of the route
taken by the ridesplitting driver is often longer than the optimal route. In some cases,
this even results in the actual ridesplitting distance being larger than the sum of the two
individual distances for rides A and B. This occurs four times in our data set for Trips 1, 9,
13, and 20.
Figure 2 below shows a visual of the possible paths for Trip 15 traveling within the
Berkeley grid. The individual routes for trips A and B are shown in blue and red, respectively.
The optimal ridesplitting route is shown in black.

Figure 2: Plot of Trip 15. Ridespliting route plotted in black, with the single trips plotted
in blue and red

As shown in Figure 2, ridesplitting reduces total VMT when compared to the total
distance of each individual ride. The sum of the distances of individual rides A and B is
3.74 miles while the distance of the split ride is 2.40 miles, resulting in a total VMT savings
of 1.34 miles.

UC Berkeley, CE295
111

Table 2 gives the average savings in VMT from using ridesplitting instead of single rides,
for both the optimal ridesplitting difference and the actual ridesplitting difference. Thus
savings are also normalized per mile traveled. Furthermore, the table gives the potential
VMT savings from drivers following the optimal ridesplitting route instead of the actual
route taken. Finally, the corresponding CO2 emissions savings per mile traveled are given
for two different vehicles, the Toyota Camry and the Toyota Prius.

Table 2: Significant miles saved parameters

The results show that on average, 0.63 miles were saved from using ridesplitting instead
of taking two single rides, calculated using the actual ridesplitting distance instead of the
optimal distance. This corresponds to a per mile savings of 0.23 miles. Additionally, if
ridesplitting drivers followed the optimal route, these average savings could be increased by
over 35% to 0.31 miles saved per mile traveled. These optimal savings correspond to a CO2
emissions savings of 0.096 kg per mile traveled for the Toyota Camry, and a savings of 0.051
kg per mile traveled for the Toyota Prius.

Summary and Future Work

This paper supports the claim that ridesplitting reduces overall VMT and emissions
from ridesourcing since the results show that ridesplitting saves an average of 0.23 miles
per vehicle mile travelled compared to single rides. Additionally, ridesplitting fills empty
seats in cars and potentially reducing congestion. In order to maximize these benefits,
public policy should be implemented to incentivize ridesplitting. While ridesplitting already
offers a significant discount for riders, policy could dictate cheaper split rides, or conversely,
more expensive individual rides. This pricing technique to encourage ridesplitting could be
especially impactful during peak hours. Similarly, ridesourcing could be priced to incentivize
ridesplitting to public transportation connections, such as train stations, bus stations, or
airports. The emissions savings of ridesplitting could be further improved by implementing
ecorouting. Ecorouting is a navigation concept that finds a route requiring the least amount
of fuel or producing the least amount of emissions (Boriboonsomsin et. al, 2012).
Additional insight could further enhance this study. We limited the study to Berkeley,
which varies greatly from other parts of the country with respect to both geography and

UC Berkeley, CE295
112

demographics. This may have biased our data towards students and their behavior. It was
also not possible to get data about other members in the split ride, so we only took into
account the portion of the trip that overlapped with additional riders. Understanding the
chain of ridership as riders get picked up and dropped off within the same vehicle and route
could provide valuable data on behavior and miles saved. We also did not have the actual
algorithm used by Uber or Lyft for routing, which takes into account additional factors
such as congestion, speed limits, and u-turns. Finally, while the number of grid points was
simplified for this study, the node density could be improved to include every intersection
and crossing in Berkeley.
Additionally, further study should be done on the effects of elevation and deadheading on
emissions. Especially in a city marked by topographical changes such as Berkeley, vehicles
may have significant emission changes if rides are mostly uphill or downhill. Furthermore,
emissions associated with both deadheading and induced demand are likely significant and
should be accounted for in additional research.

Appendices
A: The Haversine Formula
This formula utilizes latitudes and longitudes to calculate the distance between two points
on the surface of a sphere.
The Haversine function is given as:

hav(θ) = sin2 (θ/2)

Using this, the distance between two points on a sphere is calculated using the following
formula:
p
dhav = 2rarcsin( hav(φ2 − φ1 ) + cos(φ1 )cos(φ2 )hav(λ2 − λ1 ))

where dhav is the distance between the two selected points on a sphere, (φi , λi ) are the
latitude-longitude pairs of the points, and r is the radius of the earth.

B: Dijkstra’s Algorithm
Given a source node, this algorithm calculates the shortest path to each other node in the
grid. The algorithm works in three main steps: initialization, calculation, and evaluation.
All nodes are initially set as unvisited, and a start node is chosen. The start node is given
a distance value of 0, as it is the current node, and all other nodes are given distance values
of infinity. In the calculation step, the cumulative distance between the current node and all
of its children are evaluated. These calculated distance values are compared to the present
values, and the smaller ones are assigned. The current node is then considered visited, and
the node with the smallest distance value is then selected and the calculation step is redone.

UC Berkeley, CE295
113

This is done until all of the nodes have been visited. This allows for the optimal shortest
path between any given origin and destination nodes to be calculated.

C: Division of Work
This project was divided among the authors as follows:

Caroline Neaves: Literature review, data collection, grid and adjacency matrix creation

Kaylee Homolka: Shortest path algorithm, mixed integer linear program algorithm assis-
tance, LaTeX implementation

Riley James: Shortest path algorithm, mixed integer linear program algorithm, data col-
lection

Shanay Kapadia: Data collection, grid and adjacency matrix creation

All members participated in the writing of the report and the discussion of the results.

Acknowledgments
We would like to thank our fellow classmates for providing survey data for us to conduct
this study, as well as Scott Moura and Bertrand Travacca for their guidance and support
throughout the process.

About the authors

Caroline Neaves Graduated from John Hopkins University with a degree in Environmental
Engineering. Loves Papa Johns pizza, long walks on the beach, and gardening.

Kaylee Homolka Graduated from the University of Nebraska-Lincoln with a degree in

mathematics. Fueled by coffee, dogs, and Radiohead.

Riley James Graduated from Occidental College with a major in physics. Passionate about
calligraphy and Big Brother.

Shanay Kapadia Graduated from Rice with a degree in chemical engineering. Enjoys
baseball, birdwatching, and battery storage.

References
Agatz, N., Erera, A., Savelsbergh, M., & Wang, X. (2012). Optimization for dynamic
ride-sharing: A review. European Journal of Operational Research, 223(2), 295-303.

UC Berkeley, CE295
114

Boriboonsomsin, K., Barth, M. J., Zhu, W., & Vu, A. (2012). Eco-routing navigation
system based on multisource historical and real-time traffic information. IEEE Transactions
on Intelligent Transportation Systems, 13(4), 1694-1704.

Clewlow, R. R., & Mishra, G. S. (2017). Disruptive transportation: the adoption, utiliza-
tion, and impacts of ride-hailing in the United States (Vol. 7). Research Report-UCD-ITS-
RR-17.

Hampshire, R. C., Simek, C., Fabusuyi, T., Di, X., & Chen, X. (2017). Measuring the
impact of an unanticipated suspension of ride-sourcing in Austin, Texas.

Hawkins, A. J. (2018, April 02). BMW and Daimler are putting their differences aside
to beat Uber. Retrieved from https : //www.theverge.com/2018/4/2/17188374/bmw −
daimler − merger − car2go − reachnow − mobility

Henao, A. (2017). Impacts of Ridesourcing-Lyft and Uber-on Transportation Including

VMT, Mode Replacement, Parking, and Travel Behavior. University of Colorado at Denver.

MATLAB 8.0 and Statistics Toolbox 8.1, The MathWorks, Inc., Natick, Massachusetts,
United States.

Ngo, V. D. (2015). Transportation network companies and the ridesourcing industry: a

review of impacts and emerging regulatory frameworks for Uber (Doctoral dissertation, Uni-
versity of British Columbia).

Nicolas, L., & Moura, S. J. (2016). Optimal Routing and Charging of Electric Ride-
Pooling Vehicles in Urban Networks.

Rayle, L., Dai, D., Chan, N., Cervero, R., & Shaheen, S. (2016). Just a better taxi? A
survey-based comparison of taxis, transit, and ridesourcing services in San Francisco. Trans-
port Policy, 45, 168-178.

San Francisco County Transportation Authority (SFCTA). (2017). TNCs Today: A

Profile of San Francisco Transportation Network Company Activity. Retrieved from http :
//www.sf cta.org

Schaller, B. (2017). Unsustainable? The Growth of App-Based Ride Services and Traffic,
Travel and the Future of New York City. Schaller Consulting.

Shaheen, S., & Cohen, A. (2018). Impacts of Shared Mobility: Pooling. Shared Mobility
Policy Briefs: Definitions, Impacts, and Recommendations, 12.

Shaheen, S., Totte, H., & Stocker, A. (2018). Future of Mobility White Paper.

Toyota Motor Sales (2018). 2018. Toyota Camry Full Specs and 2018 Prius Full Specs.

UC Berkeley, CE295
115

Retrieved from https : //www.toyota.com/prius/f eatures/mpg/1221/1223/1224

US Energy Information Administration. (2017). How Much Carbon Dioxide is Produced

from Burning Gasoline and Diesel Fuel? US Department of Energy. Retrieved from https :
//www.eia.gov/tools/f aqs/f aq.php?id = 307&t = 11

UC Berkeley, CE295
116

UC Berkeley, CE295
117

Electric Vehicle Charging Station Controller

Salma Benslimane, Ilias Atigui, Zuwa Oriakhi, Mathilde Badoual

Abstract
In this project, we aim to minimize the Electric Vehicle (EV) owner’s electricity cost
from charging of his/her vehicle at a charging station, while ensuring that the EV state of
charge is within acceptable bounds to meet their satisfaction. In developing this system, we
consider varying electricity prices as well as the availability of the vehicle to be charged.
To tackle the problem of stochasticity, we develop a charging station controller that
employs Stochastic Dynamic Programming and, more precisely, we model the vehicle avail-
ability as a Markov Decision Process. We contrast our results against those produced
from using other established optimization/control methods.

Introduction
Motivation and Background
A growing public awareness about clean energy, combined with continued innovation in
battery storage, suggests that the wide-scale adoption of Electric Vehicles (EVs) will be a
likely scenario in the near future. Defining the role that EVs will play in society is a challenge
that needs to be addressed on several fronts such as infrastructure compatibility, electricity
generation and distribution constraints, effects on mobility, etc., as well as the opportunities
that EVs present to the power system in the form of distributed storage.
Indeed, energy management strategies are the key to the successful integration of EVs
both from a storage and load perspective. It is logical to assume that the adoption of EVs
will increase the peak demand and will present added strain on the system. Effective demand
response can decrease this peak by shifting the load demand from EVs to non-peak times
which would significantly reduce consumer energy cost. Thus, our project will focus on
control of a fleet of EVs.
This work is particularly important at a time where California’s government wants 5
million zero-emission vehicles on the road by 2030 (’CEC: California EV chargers will add
1 GW of peak demand by 2025’, Utility Dive, Gavin Bade). Nonetheless, using statistical
data from the US Department of Energy, we can show that the current grid can only handle
those electric cars so long as they actually represent less than 25% of the total number of
cars on the road in a given area (’NREL Finds Future EV Charging Demand Will Require
Coordination Between Utilities Car Owners’, CleanTechnica, Steve Hanley).

UC Berkeley, CE295
118

The main issue of controlling EVs is to tackle EV presence and prices uncertainty. Then
we will consider the stochastic approach using random variables for those values. This
problem has been addressed in various papers cited in the Relevant Literature section.

Focus of this Study

Our goal is to develop an online controller to charge a fleet of EVs. The EVs should
be in the charging state depending on the energy prices while guaranteeing the consumers
satisfaction.
The first model will consider that the vehicles never leave and that the controller has
to balance between minimizing costs owing to electricity prices and guaranteeing consumer
satisfaction. Then, we will use a stochastic control to compute the optimal energy allocation
taking into account the cars presence. Finally, we run a ”stupid” model that charges when
the car arrives. To make a relevant comparison, we will run a Monte Carlo simulation for
the three models. We will assume that our models know the electricity prices during the
day.
Here is a graph summarizing the system we are analyzing (Figure 1).

Figure 1: Description of the system’s model

Literature Review
A number of modeling frameworks and optimization methods have been used to tackle
the issues linked to the development of an online controller which would be used in order to
charge a fleet of electric vehicles.
Keeping in mind that one of our main objectives is to minimize the electricity cost of
the different charging stations: Caroline Le Floch et al. develop in [?] a model predictive
control (MPC) formulation in order to minimize the electricity cost of a plug-in electric
vehicles fleet under several distribution grid constraints; this has been crucial to get a better
understanding of how a controller can actually be implemented in our system.
Motivated by the need to keep safe and efficient infrastructure, as well as the urgency to
provide sustainable energy, Xiaohua Wu et al. worked in [?] on a smart home designed to

UC Berkeley, CE295
119

minimize the total cost depending on a specific tariff which varies with the time of the day.
Furthermore, this paper also wants to satisfy the home power demand (which corresponds
in a way to the satisfaction function we tried to establish, the comparison has been really
helpful to get some insights of the problem).
In a context where the transition between fossil fuels and renewable energy is taking
shape, Emil B. Iversen et al. address in [?] the burning issue of the involvement of different
actors related to the energy sectors. And there is definitely a stochastic nature in these
driving patterns. Therefore, it shows us tha tusing a Markov decision process can definitely
be relevant in order to model the loading of an EV fleet.

Key Contributions
Thus, our team explored the computational complexity difference between convex opti-
mization and DP as well as compared a ’simple model’ to a more complex one, taking into
account stochasticity of cars’ presences.

Technical Description
We will first present the model of the EV charging station, the data we used for this
study, followed by the different controllers we designed and finally the comparison between
those controllers.

1 Description and Nomenclature

We define the variables and characteristics of our problem as follow:
- N : number of nodes (chargers) stations
- H : time horizon
- c[k] electricity prices at time k, in [ kW$ h ]
- xi [k] State of Charge (SOC) of car at node i at time k, in [kW h]
- ui [k] energy allocated to node i at time k, in [kW h] - also the control
- zi [k] ∈ {0, 1} presence of car at node i at time k
- xinit
i initial state of charge of car at node i
Let’s consider a charging station for electrical vehicles containing N chargers or nodes
- each one representing a plug-in.
We consider a time horizon of H with t ∈ [0, H − 1] and a discrete time step k ∈
{0, ..., H}
Each vehicle present at a plug, i, at time, k, has an initial SOC and needs to be charged.
At a certain time, the car will leave and we would need to keep track of the presence of cars,
so we will define a vector containing the car presence at each plug i.

UC Berkeley, CE295
120

2 Data Used
Electicity Price
We gathered electricity prices data from the California Independent System Operator
(CAISO). Our data is between 2013 and 2015, for each hour of the day and for each corre-
sponding month.
The following graph is shows the electricity prices during 24h for different days in 2013.
We can see that the prices are following a regular shape every day, thus we will consider the
average price at each hour as the data known by our controller.

Figure 2: Prices data for one day from Caiso Prices

Electric Vehicles Energy Consumption and Presence

From PECAN Street data we collected the electricity consumption for five EVs at each
hour of the day for each day of the month in 2014. This allowed us to deduce the presence
of an EV during the day.
Our project is focused on residential charging station but PECAN Street data is collected
from personal chargers, nevertheless, we can imagine the use case of a Smart Grid city such as
Ecoblock Project in Oakland [reference] where residents are charging shared electric vehicles.
Since it is a residential area, the charging profile might be similar to a personal charging
station.
The figure below shows the vehicles charging patterns for Sunday as follows (Appendix
A for the entire data) :

UC Berkeley, CE295
121

Figure 3: EV Data

By analyzing the shape of the car presence , we can deduce that it depends on the hour
of the day but also on the previous state (charging or not). Then we can reasonably model
this presence with a random variable following a Markov Chain model. If we denote Zi [k]
the random variable associated to the presence of the car, we can write the Markov Chain
model as:

∀i, P [Zi [k] = m|Zi [k − 1] = n] = pm,n [k]

With m and n in {0, 1}. 0 is unplugged and 1 is plugged. The following plot represents
the probability to switch modes.

Figure 4: Probability of switching modes for a station

And here is an extract of the probability matrix for the first hour of the day and for the
third one.

UC Berkeley, CE295
122

Figure 5: 3D representation of the probability of switching modes for a station

We can see that the plot above is consistent since the probability P of ’plugged to plugged’
tends to 0 around 5-6AM. That’s the time when most people go to work using their car.
Besides, we can see that during the night, between 2-3AM, P r(’plugged to plugged’) is
the highest (around 0.9) since the electricity price isn’t too high compared to the price at
the beginning of the night. Furthermore, we can highlight that although there are more
attractive prices during the day, the cars still need to be charged during the night to allow
people to commute in the morning.
We also computed the Markov Chain random variable for the prices in Appendix 1.

3 Satisfaction function
Each car plugged to the station is going to get charged up to a given SOC before leaving
the station.
In order to control the level to which each vehicle is being charged, we constraint our problem
by adding a satisfaction function to the cost.

UC Berkeley, CE295
123

We define the satisfaction function as follows - representing an exponential decay:

1 − eα·(1−x)
g(x) = (1)
1 − eα
with α a constant.

Figure 6: Plot of the satisfaction function vs the State of charge

4 Model 1: Simple Controller (Deterministic with con-

stant car presence)
Optimization formulation
In this model, we assume that the price c[k] is known. The controller is not taking into
account the cars presence - meaning that the presence zi [k] = 1 ∀k, i.
We would like to minimize the cost of charging the EVs with a penalty if the car is very
empty, so our cost function would be:
H−1
XN −1
X
min (c[k] · ui [k] + g(xi [k])) (2)
u,x
k=0 i=0

Under the constraints:

xi [k + 1] = xi [k] + ui [k] (3a)

0 ≤ ui [k] ≤ 1 − xi [k] (3b)
N
X −1
ui [k] ≤ Umax (3c)
i=0
xi [0] = xinit
i (3d)
xi [H] = 1 (3e)

UC Berkeley, CE295
124

The minimization of the cost is subject to system and vehicle-imposed constraints. The
first (Equ. 3a) describes the equation of battery charging (x being the SOC of the cars
plugged at each node). Then, the equations (Equ. 3b and 3c respectively) constrain the
control input to ensure that the SOC doesn’t exceed 1 and that the sum of the power drawn
over a time step across all chargers does not exceed the maximum permissible power that
can be drawn from the grid/transformer. Also, we assign the initial SOC and at last, we
would like to fully charge the electrical vehicles at the last time step, constraint described
by (Equ. 3e).

Results and Comparison

Optimization using CVX:
The problem defined in 3.1 is a linear problem that can be numerically solved and esti-
mated using CVX (Appendix 1).

We run the simulation for 3 chargers in a station and plot the SOC of the cars, the
cumulative costs as well as the energy provided at every node (Figure 8).

Optimization using Dynamic Programming:

Taking into account that the charging level for an electric vehicle is discrete, and that
we would like to include more complexity in our model (stochasticity), a logical step to be
taken after using linear programming to solve the problem is to use dynamic programming.
For that matter, we write the principle of optimality as follow:
 
Vk (xk ) = min c[k] · u[k] + g(x[k]) + Vk+1 (xk+1 ) (4)
u[k]

with
VN (xN ) = g(xN ) (5)
We implement the following equation numerically for all the chargers of the station, and
we obtain the plots in Figure 8.

Comparison of the two methods

The results of both methods are very close, the energy allocated to the cars in each node
has the same shape, we can see that for the linear programming using cvx, the algorithm
charges the station in order of the lowest SOC giving more energy to the station 1 - in our
case. However, we can see more random charging in the algorithm that uses dynamic pro-
gramming as it found a charging optimum that respects the constraints and the optimality
principle.

We also compare the cost and computational time cost of the two methods and find that:

UC Berkeley, CE295
125

Figure 7: Plots of the simulation results for the simple model using linear programming (plot
a on the left) and using dynamic programming (plot b on the right)
Cost in USD Computational Cost in seconds
Linear Optimization 63.9 0.27
Dynamic Programming 63.7 28

Table 1: Comparison between the two numerical methods in cost and cimputational cost

5 Model 2: Deterministic with changing car presence

Optimization formulation
As stated before, in reality the presence of cars in the station would vary throughout the
time horizon. For this reason, we consider an intermediate model in which the presence of
the cars is taken into account.
The only equation that would change from the previous model would be:
 
xi [k + 1] = zi [k + 1] · xi [k] + ·ui [k] (6)

With zi [k] the presence of a car at node i at time k. The optimization program can be
written as follow:
H−1
XN X−1
min c[k] · ui [k] (7)
u,x
k=0 i=0

UC Berkeley, CE295
126

Under the constraints:

xi [k + 1] = zi [k + 1] · (xi [k] + ui [k]) (8a)

0 ≤ ui [k] ≤ 1 − xi [k] (8b)
N
X −1
ui [k] ≤ Umax (8c)
i=0
xi [0] = xinit
i (8d)
xi [H] = 1 (8e)

Results and Comparison

Using the same logic as previously, we solve this problem numerically using cvx and using
the same DP algorithm as before and including the presence in the simulation.
We choose the following presence of our cars, in two nodes:

Figure 8: Presence of the cars at two different stations

We plot the results of the simulation for both methods for this second model:

UC Berkeley, CE295
127

Figure 9: Plots of the simulation results for the second model (includes the presence) using
CVX (plot a on the left) and then using dynamical programming (plot b on the right)

The advantage of dynamic programming in this case, is that it allows the use of a sat-
isfaction function in the optimization function, instead of the hard constraints that the
SOC needs to be full at the time where the car leaves.
We still witness a difference in computational cost between the use of cvx and dynamic
programming, which increases when we increase the number of node or chargers in the
station. (this is one of the reasons we choose to plot the simulation with 2 nodes with 2 cars
each in the station).
Here is a table summarizing the difference of costs between the two methods (which does
not imply the superiority of a method based on the numerical results only, because each
has specific advantages - for example the possibility to add non linearity to the dynamic
programming).

Cost in USD Computational Cost in seconds

Linear Optimization 123 0.3
Dynamic Programming 110 4.8

Table 2: Comparison between the two numerical methods in cost and computational cost

UC Berkeley, CE295
128

6 Model 3: Stochastic Model Predictive Control

Optimization formulation
The objective function can be expressed as:
H−1
XN −1
X
min E( c[k] · ui (zi [k]) + g(xi [k], zi [k])) ∀i, k (9)
u,x
k=0 i=0

Under the constraints:

zi [k] ∈ 0, 1 (10a)
xi [k] = 0 for zi [k] = 0 (10b)
xi [k + 1] = xi [k] + ui [k] for zi [k] = 1, zi [k + 1] = 1 (10c)
xi [k + 1] = xinit
i [k + 1] for zi [k] = 0, zi [k + 1] = 1 (10d)
ui [k] = 0 for zi [k] = 0 (10e)
g(xi )[k] = 0 for zi [k] = 0 (10f)
0 ≤ ui [k] ≤ 1 − xi [k] (10g)
N
X −1
ui [k] ≤ Umax (10h)
i=1
xi [0] = xinit
i [0] (10i)

pm,n [k] = P r zi [k] = m | zi [k − 1] = n], ∀m, n ∈ 0, 1, k = 0, . . . , N − 1 (10j)

From the above, the State of Charge xi [k] depends on the presence zi [k] and (10b) and
(10c) outline this dependency. When the vehicle returns (presence transition from 0 to 1)
it is given an arbitrary state of charge denoted by xinit i [k] as constraint (10d) shows. There
can be no control input when the vehicle is not present which (10e) accounts for. All other
constraints assume the same physical meanings as outlined in previous models.
Value Function
Let Vk (x(k)) denote the minimum cost from time step k to terminal time step N, given
the present State of Charge x(k) which depends on the vehicle presence z(k).
Bellman’s Principle of Optimality
In developing our controller, we consider the expected value function at every time step
and by solving backward in the time horizon, we determine the optimal control inputs, u.
In this case, the state variable x, the control input u, and the satisfaction function g are all
dependent on the presence, z(k) which is modeled as a Markov process. Thus, the Expected
value function at time [k + 1] can be determined for the different possible values of z, x and
weighted by the probability of transitioning to these states.

Vk (x(z[k])) = min c[k] · ui (z[k]) + g(x[k], z[k]) + E[Vk+1 (x(z[k + 1]))]
u[k]

X
Vk (x(z[k])) = min c[k] · ui (z[k]) + g(x[k], z[k] = Z m ) +

pm,n [k + 1]Vk+1 (x(z[k + 1]))
u
m,n∈0,1

UC Berkeley, CE295
129

The boundary condition is the same as in the previous model:

VN (xN ) = g(xN )

Results and Comparison

We first plot the results for one day simulation of the model 1 using DP and model 3
using SDP with the same presence scenario:

Figure 10: Plots of the simulation results for the first model using DP (left) and the third
model using SDP (right)

We can see that the two models are performing well, but to be able to really compare
the two controllers we have to run a Monte Carlo simulation. We did it using 2000 scenarios
from Pecan Street cars presence profiles.
In addition to the comparison between the two controllers, we computed a ”naive” con-
troller which charges directly when the car arrives. The results are presented in the following
figure:

UC Berkeley, CE295
130

Figure 11: Monte Carlo simulation

We can see that our stochastic model is much better that the naive model but only a
little better that the simple model. Indeed, the simple model is already taking into account
the satisfaction of the customer as well as the electricity prices. And both controllers are
better that the naive one which is not taking the prices into account.
Also, our stochastic controller adds a constraint (the presence of the car) which is de-
creasing a lot the price but increases the satisfaction cost. Indeed, the stochastic controller
can make the car waiting more before being charged since it knows if the car will probably
leave or not.

Executive Summary
This project aimed to develop a robust controller for a residential EV charging station.
The controller minimizes the EV owner’s charging costs while ensuring a level of satisfaction
which is dependent on the vehicle’s State of Charge at any time.
Our results show that Linear Programming is a reasonable controller and has a short
computation time. But it has two main problems: it cannot handle discrete charging levels
and cannot take the presence of cars into account. Dynamic Programming on the other
hand is a better controller because it can give discrete charging levels but takes a lot of
computation time.
Because prices are correlated to the demand/offer ratio, controlling the use of energy
depending on prices has a direct effect to reduce the ”duck curve” . It is the difference

UC Berkeley, CE295
131

between the low demand and high generation during the middle of the day and with the
high demand low generation at the end of the day.
As we saw using stochastic dynamic programming might be a good solution but the curse
of dimensionality and the changing environment might lead to the use of Markov Decision
Process. It uses Machine Learning approach with a closed loop system to improve the
controller depending on a measurement of its result.

UC Berkeley, CE295
132

[L]APPENDIX

Appendix

Figure 12: Markov Chain Model for electricity prices

UC Berkeley, CE295
133

Figure 13: Markov Chain Model for electricity prices

UC Berkeley, CE295
134

Acknowledgments
We would like to thank our Professor Scott Moura of eCal at UC Berkeley for his
guidance throughout the project and in the Energy Systems and Control course in general.
Specifically, we appreciate his suggestions on best methods with which to approach our
problem which allowed us to maximize the skills we gained from the course as well as develop
a deeper understanding of systems design techniques.
Also, we would like to thank our Graduate Student Instructor, Bertrand Travacca for
the time he took to meet with us regularly and discuss our progress in the project. He has
been a great guide and was especially helpful in helping us formulate the problem which
ensured the relevance of our focus of study.
[L]

About the Authors

Ilias Atigui is a Master of Engineering student in the Civil & Environmental Engineering
Department at UC Berkeley completing his degree in May 2018. During his undergrad
in France he experienced an overdose of theoretical Mathematics & Physics which con-
vinced him to apply at UC Berkeley in order to get some different kind of Engineering
experiences.

Mathilde Badoual is a Master of Science student in the Civil & Environmental Engineer-
ing Department at UC Berkeley. She has a bachelor in Electrical Engineering and
Computer Science. She has a focus on control of energy systems and integration of
renewables. Starting a PhD program at eCAL on this topic.

Salma Benslimane is a Master of Engineering student in the Civil & Environmental Engi-
neering Department at UC Berkeley completing her degree in May 2018. After studying
in the Preparatory Classes she decided to dive deeper into Aeronautical and Aerospace
domain in which she received her bachelor, before applying at Cal to focus on Systems
Engineering.

Zuwa Oriakhi is a Master of Engineering student in the Civil & Environmental Engineering
Department at UC Berkeley completing his degree in May 2018. He is specializing in
Systems Engineering with a focus on Energy systems and has a Bachelor’s Degree
in Electrical Power Systems Engineering. He is passionate about building sustainable
energy infrastructure and facilitating the grid-level integration of distributed resources.

UC Berkeley, CE295
135

UC Berkeley, CE295
136

Part III

Building Energy
Management

UC Berkeley, CE295
137

Predicting Building Electric Loads under Climate

Change
Megan Dawe, Samuel Fernandes, Jin Pan, Julia Szinai

Abstract
Accurately predicting buildings’ energy usage serves a variety of purposes, including util-
ity load forecasting, building energy management, and energy efficiency savings estimation.
In addition, because climate change is predicted to cause temperature increases, understand-
ing the building energy response to warming will aid planning authorities and utilities in the
development of climate adaptation strategies. There is a large body of literature about
methods to predict commercial building energy use. Most of these methods use outside air
temperature and building characteristics such as square footage and occupancy for their pre-
dictions. However, because data on building characteristics is often not publicly available,
it is useful to have relatively accurate models of energy usage that rely only on outside air
temperature measurements.
High resolution hourly smart meter data, and recent advances in data science and machine
learning have produced new statistical forecasting methods that can be applied to building
energy prediction. Using a sample portfolio of commercial buildings, the objective of this
study is twofold:

1. To test and choose the best predictive model of hourly energy usage for each build-
ing with weather data inputs among a variety of publicly available machine learning
methods;

2. To use the best predictive model for each building to forecast future energy usage under
a variety of climate warming scenarios.

1 Introduction
1.1 Motivation and Background
Commercial and residential buildings currently account for approximately 40 percent of
the total energy consumption in the U.S. and emit approximately one third of greenhouse
gas emissions [1]. Decreasing energy consumption, specifically electricity, through improved
efficiency in buildings can avoid construction of new power plants, reduce grid infrastructure
costs, and lower carbon emissions — in addition to saving customers money on energy bills
[2]. In parallel to these climate change mitigation efforts, adaptation to expected changes in

UC Berkeley, CE295
138

climate and weather patterns provides another motivation to accurately predict electricity
loads. Accurate predictive modeling tools under climate warming will be necessary to target
buildings for energy efficiency improvements and assist in grid-level resource planning and
decision-making. Particularly for grid resource planning, if predictions of energy usage in
general are not correct and underestimate loads, utilities may have challenges with under-
procured generation resources and reliability.
There is a large body of literature about predicting building energy usage, and historically
most studies have only used temperature data and traditional regression analysis. These
studies have typically predicted monthly or annual energy usage. However, with an increasing
amount of available hourly smart meter data from buildings and advanced computational
and analytical capabilities, there is potential for improved prediction accuracy, even with just
publicly available temperature data. Improved data-driven building energy usage modeling,
especially at the hourly time resolution, can provide more dynamic capabilities for building
energy management and building-to-grid interactions such as load-shifting for renewable
integration, among other applications. Hourly energy usage predictions can also improve
the forecast of peak load to help grid planners anticipate peak capacity needs.

1.2 Focus of this Study

The focus of the study is to identify robust models from currently available open-source
statistical methods to predict the electric load of a set of commercial buildings, using outdoor
air temperature data as an input parameter. Then, using the most accurate prediction
method for each building and temperature data downscaled from global circulation models,
this analysis forecasts future electric loads under feasible climate warming scenarios, which
can enable better utility planning both for building energy management and building-to-grid
interactions.

1.3 Literature Review

Numerous efforts have been made to develop, test, and compare methods that predict
building energy use, including engineering (e.g. simulation), statistical, and machine learning
models. The engineering methods use physical principles to calculate thermal dynamics and
energy behavior on the whole building level or for sub-level components [3]. Several elaborate
simulation tools have been developed based on the engineering methods including DOE-2,
EnergyPlus, BLAST and ESP-r. They generate accurate results effectively, but require
details of building and environmental parameters as inputs, which are usually difficult to
obtain. Although some simplified methods have been put forward, their application is limited
due to such simplification [3].
Statistical regression methods correlate energy consumption with influencing variables
[4, 5, 6]. These empirical methods are developed based on the historical energy data and are
widely used to predict the future energy usage over simplified variables, including climatic
variables [3]. Bauer and Scartezzini proposed a simplified method to calculate heating and
cooling load simultaneously [5]; Dhar et al. take outdoor dry-bulb temperature as the only
weather variable to calculate the heating and cooling load in commercial buildings [7]; Lei
and Hu further showed that a single variable linear model is sufficient and practical to

UC Berkeley, CE295
139

model the energy use in hot and cold weather conditions [8]. Autoregressive Integrated
Moving Average (ARIMA) model was first derived on the previous data and was then used
to predict the next-day load profile [9]. Overall, statistical regression methods are one of the
most commonly used method in the building-level energy prediction and forecasting.
The most common machine learning methods cited in the literature include Artificial
Neural Networks (ANN), Support Vector Regression (SVR), autoregressive models, random
forest, decision trees, and k-Nearest Neighbors (k-NN) [10]. There are also several proprietary
models developed by companies that provide energy prediction and energy management
services to building owners. These include those developed by Buildings Alive Pty. Ltd.,
Gridiu, Inc., Lucid Design Group, and Performance Systems Development of New York,
LLC. In addition, this study evaluates a publicly available model relying on random forests,
MAVE, which was developed at the University of California, Berkeley [11].
Artificial Neural Networks (ANN) are very widely used in the application of building
energy usage prediction, especially for hourly prediction. They are effective at solving non-
linear problems and thus can be applied to some very complex conditions with a variety
of variables. Moreover, ANN are demonstrated to have higher performance than engineer-
ing methods in estimating appliance, space cooling, and lighting energy consumption [12].
However, ANN usually requires data pre-processing techniques to reduce noise or useless
variables [3].
Support vector machines (SVM) are highly effective in solving nonlinear problems with
only small quantities of training data required. The performance of SVMs can be better than
the traditional neural networks, especially for hourly prediction [13]. However, the training
process of SVM becomes extremely slow when the training data size is large [14].
Another publicly available statistical model, Mave, which was developed by the Center
for the Build Environment, is a tool based on the machine learning methods for automated
Measurement and Verification (M&V) [11]. M&V is a process used to quantify energy savings
from an energy efficiency measure, typically by estimating the counterfactual energy usage
of the building, but-for the implementation of the energy efficiency measure, compared to
the actual post-implementation usage [15]. M&V 2.0 techniques use machine learning to
improve the processing speed and accuracy of the counterfactual estimation. Mave uses a
random forest model to estimate the counterfactual energy usage.
Based on the findings from Burger and Moura, even among the sophisticated machine
learning methods described above, no one model is the “best” for all buildings or scenarios
[16], and some buildings are inherently difficult to predict because of changing business
processes, occupancy, thermostat setpoints, and so on [17]. Model selection depends on the
scale of the study or project goal. For instance, studies that focus on individual building
energy load predictions with substantial building details may determine certain models are
more appropriate than those with the goal of portfolio-level predictions and limited building-
level data, such as experienced by utility forecasting efforts. At the building-level scale, the
motivations for energy prediction are for fault detection diagnosis (FDD) [10] and to provide
baseline estimates to assess energy efficiency intervention savings [18]. At the grid or utility
level, the motivation is to assist in load forecasting and utility-scale planning, such as energy
efficiency and demand-side management programs [16]. A major difference between the
investigations between these two groups is the availability of data and the resources (e.g.
time and funding) available for computation. For large-scale energy demand forecasting,

UC Berkeley, CE295
140

efficient and low-cost methods are preferred.

Studies use various methods to assess and compare goodness of fit for models. Common
metrics include root mean squared error (RMSE) [3] (and the normalized version called the
coefficient of variation RMSE or CV(RMSE)) and mean average percentage error (MAPE)
[16]. The choice of parameters used to fit the model and predict energy use vary between
studies and are related to data availability and project goals, as discussed above. Most
commonly, models are fit using measurements of electricity demand at some time interval,
date, time, and at least one weather variable [10]. Training and testing periods also differ by
project based on prediction goals. Most studies have approximately two years’ of electricity
demand data and use between 12 and 18 months as training, and the rest used for testing
[10]. Granderson et al. [18] found that “when the training period was shortened from twelve
months to nine, and then to six, there was a gradual degradation in predictive accuracy.”
Electricity consumption data can be available at 15-minute, hourly, or daily intervals, de-
pending on the prediction goal; that is, daily energy load cannot be used to provide accurate
hourly predictions, but hourly data can be used to predict daily loads.
For this project, we rely on examples in the literature that most accurately predict
building energy use with the least number of commonly available features. The following are
key studies to support the analysis methodology:

Table 1: Review of Building Energy Prediction Studies

Prediction Parameters Train Data Test Data Model Performance
Metric
Study 1 Hourly (1) Electric Buildings: 8 Buildings: (1) Ridge MAPE
[16] load demand (2) commercial, same as train (2) SVR
Time (3) Elec- 24 residential. Energy Data: (3) Decision
tric Demand Energy Data: 6 months Tree (4)
& time (4) 18 months for batch, k-NN
Electric de- for batch, 21 months
mand, time, 3 months for moving
& outside air for moving horizon
temp horizon
Study 2 Daily load (1) Outside air Buildings: 537 Buildings: 10 models NMBE;
[18] temp (2) Date commercial. same as train CV(RMSE)
(3) Time Energy Data: Energy Data:
12 months 12 months
Study 3 Daily load (1) Day of Buildings: 1 Buildings: Gaussian Training
[10] year (2) Week commercial same as train Process accuracy &
of year (3) Energy Data: Energy Data: Validation
Weekday (4) 21 months 3 months accuracy
Occupancy
(5) Heating
degrees

Lastly, we rely on some of the techniques used from the econometric analysis of Auffham-
mer and Aroonruengsawat [19], which forecast the building load impacts of climate change
under various emissions scenarios, as the basis of the forecasting portion of this analysis.

UC Berkeley, CE295
141

Auffhammer and Aroonruengsawat fit coefficients to historical temperatures and other pre-
dictors and use new temperature data from downscaled global circulation models to forecast
future energy consumption in different California regions.

1.4 Key contributions

This study contributes to the literature by comparing among a unique ensemble of ma-
chine learning methods, and applying the most appropriate method to a long-term forecast
of building loads with new temperature forecasts that reflect climate warming scenarios.

2 Technical Description
2.1 Data Selection and Pre-Processing
Building Load Data Selection: A common challenge for building energy load fore-
casting is the lack of publicly available high resolution data, such as on building character-
istics, square footage, and hours of operation or occupancy. We selected a dataset that had
multiple buildings and high time resolution because the analysis objective is to develop a
model selection and hourly forecasting methodology that can be adapted for a variety of
buildings and employed by a utility planner with only access to building load and temper-
ature data. We obtained a dataset from the Efficiency Valuation Organization (EVO) that
includes 15-minute electricity consumption and outdoor air temperature over a two year
period (2015 and 2016) for 277 commercial buildings in Fort Collins, Colorado [20]. The
dataset does not provide any additional information about the buildings. The electricity
consumption data is aggregated to hourly electricity consumption for this study.
Building Load Data Pre-Processing: We conducted a number of data-cleaning
steps on the building data. First, of the total 277 commercial buildings, we selected a
random subset of 50 buildings to use as our data for this analysis to maintain a reasonable
computational time for model fit and forecasting. Second, we removed 4 buildings from the
set of 50 because of many instances of missing data. Next, to account for load fluctuations
unrelated to weather, we added predictors for day of week, month of year, and hour of
day. We converted these categorical date-time variables into “dummies” through “one-hot
encoding”, removing the first binary variable of each category to prevent linearly dependent
columns [21].
In addition, we applied scaling factors to the temperature and load data for each building
based on the maximum values for each variable in the training data such that all variables
had values between 0 and 1. Scaling is particularly important for regularization so that the
shrinkage term λ will treat all variables on the same scale and not penalize variables with
larger units [21]. We record the scaling parameters used for each building and variable in
order to later inverse the scaling when calculating the predicted values and forecast values
into original units. The scaling parameters are calculated on the training data and are then
applied to the test and forecast values.
Climate Data: In order to forecast building energy load under future climate scenar-
ios, we downloaded maximum and minimum daily temperatures that have been statistically

UC Berkeley, CE295
142

downscaled using the Bias Corrected Constructed Analog (BCCA) method [22] for the Col-
orado region from a number of Global Circulation Models that are part of the most recently
available Coupled Model Intercomparison Project (CMIP5) ensemble for the IPCC Fifth Cli-
mate Assessment. The results of the climate models are archived and made available through
a consortium of organizations including USGS, Bureau of Reclamation, Army Corps of En-
gineers, and Scripps Institute of Oceanography, among others [23]. The available data is for
four scenarios that reflect different levels of emissions pathways and resulting climate forcing,
which are called Representative Concentration Pathways (RCPs) ranging from the most ag-
gressive climate mitigation scenario of RCP 2.5, to the least mitigation, or business-as-usual
(BAU), RCP 8.5. There are also two mid-case scenarios, RCP 4.5 and RCP 6.0 [24].
Climate Data Pre-Processing: Our query of the CMIP5 climate models yielded
outputs from 132 different models and model variants, for each year queried. We selected
outputs for the mid-century time period, 2045 - 2065 when warming effects are expected to
first be noticeable [24], and which is still within a reasonable lifetime of a building existing
in 2015, which is our training year. The results are 20 time series of maximum and minimum
daily temperature for all climate models. First, in order to account for variability between
climate models, we averaged the daily outputs for each year across all the models. Next, we
converted the daily temperature extremes to hourly temperatures, because our forecasting
step required hourly time resolution data to match test and training data. We use the
imposed offset/diurnal temperature conversion methodology broadly outlined by Guan, L.
[25]. We follow a specific functional form for daytime and nighttime hourly temperatures
devised by Linvill [26], simplified from Parton and Logan [27], that only requires the latitude
and time of day.
For daytime hours, we converted to hourly data with a truncated sine function:

(π ∗ t)
T (t) = (Tmax − Tmin ) ∗ sin[ ] + Tmin (1)
(DL + 4)

where T (t) is temperature at hour t after sunrise, Tmax and Tmin are daily temperature
extremes, and DL is day length in hours.
For nighttime hours, we converted to hourly data with an exponential decay function:

(Ts − Tmin )
T (t) = Ts − [ ] ∗ ln(t) (2)
(24 − DL)

where T (t) is temperature at hour t after sunset, Ts is the sunset temperature from (1),
Tmax and Tmin are daily temperature extremes, and DL is day length in hours. We used an
R package called ”chill R”, which has created some functions to automate this conversion,
including the calculation of the day length from the latitude.
To test the accuracy of this method, we construct hourly temperatures from observed
2015 temperatures and these estimated temperatures are only off from the observed data by
1.5 to 2 degrees Celsius, on average, each month. For the purposes of building simulation,
we found this to be an acceptable accuracy.
Finally, for our temperature forecasts, to smooth out any outlying year, we average the
resulting hourly temperatures for 2045-2065 to produce one year of hourly temperature data
for each RCP, representing mid-century warming scenarios under different emissions futures

UC Berkeley, CE295
143

for Colorado. The temperature forecasts compared to 2015 temperatures are illustrated in
Figure 1.

Figure 1: Historical vs. Forecast Hourly Temperatures

2.2 Model Types

Based on the literature review, we selected five candidate machine learning methods
that were found to have reasonable computational times and predictive performance: (1)
Mave, which is based on random forest and developed at UC Berkeley’s Center for the Built
Environment [13], (2) Ridge Regression, (3) Artificial Neural Network (ANN), (4) K-Nearest
Neighbor (k-NN), and (5) Autoregressive with eXogenous inputs (ARX).
(1) Mave Mave uses a random forest machine learning technique with the inputs and
outputs that can be set according to the available training data. Decision trees are popular
because of their simplicity and interpretability. Random forests are an ensemble learning
methodology where the performance of a number of weak learners is ”boosted.”
With a RF, there is an ensemble of C trees T1 (X), T2 (X), ..., TC (X), where X = x1 , x2 , ..., xm
is a m-dimension vector of inputs. The resulting ensemble produces C outputs Yˆ1 =
T1 (X), Yˆ2 = T2 (X), ..., YˆC = TC (X). YˆC is the prediction value by decision tree number
C. The output of all these randomly generated trees is aggregated to obtain one final pre-
diction Ŷ , which is the average values of all trees in the forest. The decision trees in this
project only consider outdoor air temperature and the previous value of energy consumption
as input variables. Both these features were allowed to be tried in an individual tree, and
the maximum depth of a tree is restricted to 4.
(2) k-NN The k-NN model uses the scikit-learn KNeighborsRegressor in Python, which
is a regression based on k-nearest neighbors [28]. The model uses month of year, day of week,
hour of day, normalized outdoor air temperature, and normalized hourly electricity data for
training. The prediction uses the 15 closest neighbors (k = 15), based on Euclidean distance,
and uniform weights, meaning all 15 nearest neighbors are weighted equally instead of weights
by distance.

UC Berkeley, CE295
144

(3) Ridge Regression The ordinary least squares with L2 regularization (aka ridge re-
gression) fits a linear model while penalizing the size of parameters based on the L2 norm.The
input training variables (predictors) for ridge regression contain the scaled temperature, the
dummy month of year, day of week and hour of day. A cross validation was conducted based
on one building’s data to determine the weighting term for the regularization penalty, and
this weighting term was applied to the rest of the buildings. In this case, the weighting term
λ = 0.6.
The least squares minimization model is given by:
X
min ||wT xi − yi ||22 + λ||w||22 (3)
w
i

where xi are the predictors for each hour, w are the weights for each predictor. The model
is fit with the closed form solution: w∗ = (X T X + λI)−1 X T Y
(4) ANN An Artificial neural network stores knowledge from past observations and
makes it available for the prediction of future values. An ANN uses hidden layers where the
number of hidden layers depends on the nature and complexity of the problem. ANN’s do
not require any information about the system as they operate like black box models and learn
relationship between inputs and outputs. The schematic diagram below shows a feed-forward
neural network architecture, consisting of two hidden layers. While there are different neural
network strategies in literature e.g. feed-forward, Hopfield, Elman, self-organizing maps, and
radial basis networks, this study uses a feed-forward neural network and back-propagation
algorithm. The algorithm is summarized as follows:

1. Use the training data to propagate through the neural network to obtain the desired
output.

2. Initialize weights using small threshold values.

3. Calculate input to the j-th node in the hidden layer using the equation below:

n
X
netj = wi jxi − θj (4)
i=1

4. Calculate output from the j-th node in the hidden layer using the two equations below:
Xn
hj = fh ( wi jxi − θj ) (5)
i=1

1
fh (x) = (6)
1 + exp(−λh x)
5. Calculate the input to the k-th node in the hidden layer using equation below:
n
X
netk = wk jxj − θk (7)
j=1

UC Berkeley, CE295
145

6. Calculate the output of the k-th node of the output layer using equation below:
Xn
y k = fk ( wk jxk − θk ) (8)
j=1

1
fk (x) = (9)
1 + exp(−λk x)
Schematic diagram of a feed-forward artificial neural network used is shown below.

Input Hidden Output

layer layer layer

Input 1

Input 2

Input 3 Ouput

Input 4

Input 5

(5) ARX An Autoregressive with exogenous Inputs Model combines information from
a previous hourly observation with an exogenous input. In this case, the exogenous input
is the average hourly load for a given hour of day and day of week, to capture the load
variability unrelated to temperature. Rather than using the prior hours load data for the
autoregressive component of the model, we use the prior hours of temperature data. This
modification is in anticipation of the forecasting step; when forecasting into the future under
climatic conditions, we do not have prior load data and only have new temperatures as
inputs. This formulation also has a physical interpretation: the current energy loads depends
on how much heating or cooling the building conducted in the previous hours in response to
temperature.
The ARX model is given by:
L
X
P̂arx (t) = αl ∗ T (t − l) + P̂avg (t) (10)
l=1

where T (t − l) are the temperatures for l prior hours and the P̂avg (t) is the average hourly
load for the given hour of day and day of week from the training data. The model coefficients
αl are the weights for each prior hours of temperature, and the number of prior hours L is 3.
The model is fit with the closed form solution: α∗ = (X T X)−1 X T Y where X is the matrix
of lagged temperatures T (t − l) and Y is the observed hourly load minus the P̂avg (t).

UC Berkeley, CE295
146

2.3 Model Fit, Model Gating, and Forecasting Methodology

After the data collection and pre-processing steps described in Section 2.1, the following
steps and Figure 2 summarize the analysis methodology, which is described in more detail
below.

1. Split dataset into 12 months of training and 12 months of testing data.

2. Train each model using appropriate predictors, including: month, day of week, hour
of day, hourly outdoor air temperature (scaled by maximum value), and 2015 hourly
electric load (scaled by maximum for each building).

3. Use test data to predict hourly electricity loads based on given datetime and normalized
outdoor air temperature inputs.

4. Calculate and compare CV(RMSE) and MAPE for each model for each building.

5. The CV(RMSE) gate will select the model that performed best (produced the lowest
CV(RMSE)) during the prediction step. Remove buildings which do not achieve at
least 25% CV(RMSE) for future load forecasting.

6. Use model with lowest CV(RMSE) from step 5 to forecast hourly electricity load with
modified outdoor air temperature inputs from four climate change scenarios.

Figure 2: Analysis Methodology

In Step 1 we partition the pre-processed dataset of 46 buildings data into training and
testing periods for each individual building, with 2015 data used for training and 2016 data
used for testing. Overall the average hourly load and standard deviation of the test data is
well-balanced with the training data, as illustrated by Figure 3.

UC Berkeley, CE295
147

Figure 3: Average and SD Hourly Test and Training Energy Consumption by Building

In Step 2, for each building, we train and fit each of the 5 models described in Section 2.2
using the relevant date, time, and air temperature (scaled by maximum value) predictors.
We record the fitted parameters and in Step 3, we use those fitted parameters to predict
hourly energy loads for 2016. We inverse the scaling using the saved scaling parameters for
each variable to return the predictions to the original kWh units.
In Step 4, we calculate the CV(RMSE) and the MAPE on each model’s predictions
compared to the test data for each building. The CV(RMSE) is derived by normalizing
the RMSE with the mean of the data and has the advantage of providing a unit-less value
that allows for performance comparison across models and buildings. The CV(RMSE) and
MAPE are defined as follows:
q n
Σi=1 (yi −yˆi )2
N
CV (RM SE) = ∗ 100 (11)
ȳ
Σn
i=1 (yi −yˆi )
N
M AP E = ∗ 100 (12)
yi
We then perform a gating process in Step 5 that focuses on CV(RMSE): buildings for
which none of the models obtain a CV(RMSE) equal to or below 25% are removed from the
dataset for the climate change electricity load forecasting because they are deemed ”unpre-
dictable”. The 25% CV(RMSE) threshold originates from the American Society of Heating,
Refrigerating and Air-Conditioning Engineers (ASHRAE) Guideline for Measurement of
Energy, Demand, and Water Savings. Buildings may exceed the threshold for a number of
reasons, such as variable building occupancy or some type of change that occurred during the
data collection period (e.g. retrofit or change of tenant). Having no knowledge of building
occupancy or construction, we removed all buildings which did not meet the threshold; this
resulted in removal of 20 of the 46 buildings.
Finally, in Step 6, for each building which meets the threshold, the model with the lowest
CV(RMSE) is selected to forecast building load under climate change. To do this, we use
the fitted parameters from the best model for each building, and input the forecast hourly
temperatures from each of the four RCP scenarios to forecast hourly loads. Implicit in this
step is the assumption that the relationship between outdoor air temperature and building
load remains constant over time, and that the only change is the outdoor air temperature

UC Berkeley, CE295
148

under climate change. We do not forecast any building loads with k-NN, because that is a
clustering method that relies on observed load data for a time period close to the forecast
period and is only appropriate for short-term and stable (i.e. consistent) energy loads.

3 Discussion
Based on the results of the CV(RMSE) 25% threshold and gating process, as seen in
Figure 4, it is evident that not all buildings can rely on outdoor air temperature and datetime
inputs for accurate electricity load predictions, and that no single machine learning method
is the most robust under all situations. Among the models used in the analysis, there is
approximately an even split between which model performed the best among the 46 buildings:
MAVE (35% of buildings), Ridge Regression (19% of buildings), ANN (23% of buildings), and
ARX (23% of buildings). ARX and ANN have prior hours of temperature data as inputs, and
did equally better in the predictions. Even though Ridge Regression has the highest inputs
utilization rate followed by k-NN, they did not win among other models. Additionally, the
MAPE and CV(RMSE) across the buildings are high (greater than 100% in some instances),
and approximately half (20 of 46) buildings did not meet the 25% CV(RMSE) threshold for
any model. This suggests that outdoor air temperature and datetime inputs do not have
a strong enough correlation with electricity loads for all commercial buildings. This could
be due to non-weather-dependent loads, such as plug loads and equipment, representing a
larger portion of building electricity consumption.

Figure 4: CV(RMSE) for all models by building

For the remaining 26 buildings that met the threshold, the climate change forecasting
exercise results show that, on average, building electricity load is predicted to increase in
both annual consumption and peak demand under increasing RCP scenarios (i.e. increasing
outdoor air temperature). However, there are some buildings that see reduced overall elec-
tricity consumption and/or peak demand under the climate change scenarios. Looking into
the building-specific data reveals that these particular buildings experienced peak demand
during the winter months of the training data, and could be heating-dominated buildings
(assuming heating is provided by an electric system), or at least not cooling-dominated build-
ings. Therefore, increasing outdoor air temperatures results in a predicted reduction in loads
for those buildings and thus reduce overall electricity consumption. Building 15 in Figure

UC Berkeley, CE295
149

6 is one of the examples that had lower energy consumption in summer (i.e. warmer out-
door air temperatures) than in Winter, and low annual consumption overall. Under climate
warming scenarios, its total energy consumption decreased because the model is trained to
interpret warmer temperatures with lower electricity consumption. The other buildings in 6
are examples of predictions for mid and high annual consumption buildings. Without addi-
tional knowledge of these buildings’ consumption patterns and characteristics, we can only
speculate that some buildings will experience lower electricity demand in warming climates
due to reduced heating needs, which again highlights the importance of data availability for
prediction accuracy.

Figure 5: Percent change in total annual building electricity loads

There are a number of limitations of this study. The dataset only included commercial
buildings from one zip code in Colorado, and those buildings may undergo retrofits, change
in use, or even demolition prior to the forecast period of 2045 to 2065. However, the re-
sults more broadly imply that future climate may impact building electricity consumption
differently depending on the scenario and underlying building characteristics. Utility plan-
ners should examine current buildings’ peak demand usage when interpreting forecast results
because strategies that are currently in place or being pursued may not be sufficient. Addi-
tionally, current building design strategies and codes could incorporate climate forecasting
information to more appropriately design buildings and mechanical systems to be resilient
or adaptable to changing temperatures and building cooling loads.

UC Berkeley, CE295
150

Figure 6: Predicted, actual, and forecasting load for 3 sample buildings

4 Summary and Future Work

While the specific results of this study only apply to a particular climate zone and set
of buildings, we demonstrate a method of estimating a range of possible climate impacts
on electricity loads for a set of commercial buildings, given publicly available data. The
analysis could be repeated elsewhere in the U.S. and help planning authorities or utilities
anticipate demand changes and also target energy efficiency measures in areas where the
load would increase or change significantly under climate change. Our findings also show
that no one particular machine learning method can generally be considered to be better
than others, different machine learning models perform differently based on data, conditions

UC Berkeley, CE295
151

and a number of factors. It is best to either test a number of models and select the best
based on a performance metric or use an ensemble of models to forecast building energy
consumption. While we did the former, the latter could be an extension of this work.
Other extensions of this work include applying this method to datasets in other cities and
to include gas data in the analysis to get a more comprehensive understanding of building
energy consumption. This approach could also be used to identify temperature-sensitive
buildings that need to be retrofitted in regions that face climate change impacts.

References
[1] U.S. Energy Information Administration [EIA]. (2012). Commercial Building Energy
Consumption Survey: Energy Usage Summary, Table 1 (March 2016).

[2] Ettenson, L. & Heavey, C. (2015). California’s Golden Energy Efficiency Opportunity:
Ramping Up Success to Save Billions and Meet Climate Goals. Natural Resources De-
fense Council (NRDC) and Environmental Entrepreneurs (E2).

[3] Zhao, H. X., & Magoulès, F. (2012). A review on the prediction of building energy
consumption. Renewable and Sustainable Energy Reviews, 16(6), 3586-3592.

[4] Westergren, K.E., Högberg, H., & Norlén, U. (1999). Monitoring energy consumption
in single-family houses. Energy and Buildings, 29(3):247–57.

[5] Bauer, M., & Scartezzini, J.L. (1998). A simplified correlation method accounting
for heating and cooling loads in energy-efficient buildings. Energy and Buildings,
27(2):147–54.

[6] Pfafferott, J., Herkel, S., & Wapler, J. (2005). Thermal building behaviour in summer:
long-term data evaluation using simplified models. Energy and Buildings, 37(8):844–52.

[7] Dhar, A., Reddy, T.A., & Claridge, D.E. (1998). Modeling hourly energy use in com-
mercial buildings with Fourier series functional form. ASME Journal of Solar Energy
Engineering, 120:217–23.

[8] Lei, F., & Hu, P. (2009). A baseline model for office building energy consumption in
hot summer and cold winter region. In: Proceedings of international conference on
management and service science. p. 1–4

[9] Kimbara, A., Kurosu, S., Endo, R., Kamimura, K., Matsuba, T., & Yamada, A. (1995).
On-line prediction for load profile of an air-conditioning system. ASHRAE Transactions,
101(2):198–207.

[10] Yan, B., Li, X., Shi, W., Zhang, X., & Malkawi, A. (2017). Forecasting Build-
ing Energy Demand under Uncertainty Using Gaussian Process Regression: Fea-
ture Selection, Baseline Prediction, Parametric Analysis and a Web-based Tool.
https://doi.org/10.26868/25222708.2017.142

UC Berkeley, CE295
152

[11] Raftery, P. & Hoyt, T. (2016). Mave: software automated Measurement and
Verification. Center for the Built Environment, University of California Berkeley,
https://github.com/CenterForTheBuiltEnvironment/mave.

[12] Aydinalp, M., Ugursal, V.I., & Fung, A.S. (2002). Modeling of the appliance, lighting,
and space cooling energy consumptions in the residential sector using neural networks.
Applied Energy, 71(2):87–110.

[13] Li, Q., Meng, Q.L., Cai, J.J., Hiroshi, Y., & Akashi, M. (2009). Applying support vector
machine to predict hourly cooling load in the building. Applied Energy, 86(10):2249–56.

[14] Zhao, H. X., & Magoulès, F. (2010). Parallel support vector machines applied to the
prediction of multiple buildings energy consumption. Journal of Algorithms & Compu-
tational Technology, 4(2):231–49.

[15] Touzani, S., Granderson, J., & Fernandes, S. (2018). Gradient boosting machine for
modeling the energy consumption of commercial buildings. Energy and Buildings,
158:1533-1543.

[16] Burger, E. M., & Moura, S. J. (2015). Gated ensemble learning method for demand-side
electricity load forecasting. Energy and Buildings, 109, 23-34.

[17] Granderson, J., Price, P. N., Jump, D., Addy, N. & Sohn, M. D. (2015). Automated
measurement and verification: Performance of public domain whole-building electric
baseline models. Applied Energy 144, 106–113.

[18] Granderson, J., Touzani, S., Custodio, C., Sohn, M. D., Jump, D., & Fernandes, S.
(2016). Accuracy of automated measurement and verification (M&V) techniques for
energy savings in commercial buildings. Applied Energy, 173, 296-308.

[19] Auffhammer, M., & Aroonruengsawat, A. (2011). Simulating the impacts of climate
change, prices and population on California’s residential electricity consumption. Cli-
matic Change, 109(1), 191–210. https://doi.org/10.1007/s10584-011-0299-y

[20] Efficiency Valuation Organization (EVO). (n.d.). [Efficiency Valuation Organization

(EVO)]. Retrieved May 3, 2018, from https://evo-world.org/en/

[21] James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). Introduction to Statistical
Learning, with Applications in R. New York: Springer New York Heidelberg Dordrecht
London.

[22] Gutmann, E., Pruitt, T., Clark, M. P., Brekke, L., Arnold, J. R., Raff, D. A., &
M., R. R. (2014). An intercomparison of statistical downscaling methods used for water
resource assessments in the United States. Water Resources Research, 50(9), 7167–7186.
https://doi.org/10.1002/2014WR015559

[23] Downscaled CMIP3 and CMIP5 Climate and Hydrology Projections. (2018). Retrieved
May 2, 2018, from https://gdo-dcp.ucllnl.org

UC Berkeley, CE295
153

[24] Lukas, J., Barsugli, J., Doesken, N., Rangwala, I., & Wolter, K. (2014). Climate Change
in Colorado: A Synthesis to Support Water Resources Management and Adaptation.
Western Water Assessment, Cooperative Institute for Research in Environmental Sci-
ences (CIRES), University of Colorado Boulder.

[25] Guan, L. (2009). Preparation of future weather data to study the impact of climate
change on buildings. Building and environment, 44(4), 793-800.

[26] Linvill, D. E. (1990). Calculating Chilling Hours and Chill Units from Daily Maximum
and Minimum Temperature Observations. HortScience, 25(1), 14–16.

[27] Parton, W. J., & Logan, J. A. (1981). A model for diurnal variation in soil and
air temperature. Agricultural Meteorology, 23, 205–216. https://doi.org/10.1016/0002-
1571(81)90105-9

[28] Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ...
Duchesnay, É. (2012). Scikit-learn: Machine Learning in Python. arXiv:1201.0490 [Cs].
Retrieved from http://arxiv.org/abs/1201.0490

About the authors

Megan Dawe is an MS student in the Building Science program of the Architecture de-
partment.

Jin Pan is an MS student in the Energy, Civil Infrastructure and Climate (ECIC) program
of Civil and Environmental Engineering.

Samuel Fernandes is a PhD student in the ECIC program of Civil and Environmental
Engineering.

Julia Szinai is a PhD student in the Energy and Resources Group (ERG).

UC Berkeley, CE295
154

UC Berkeley, CE295
155

Air Conditioning Control System with Prediction of

Occupant Flow
Lorna Chen, Yao Han, Sikang Hu, Xinhao Wang,

Abstract
HVAC system is an essential component providing thermal comfort and adequate air
quality for individuals in built environment. However, the efficiency of HVAC system still
has large potential for improvement. In this project, the aim is to design a responsive air con-
ditioning(AC) system to improve the occupant comfort level and system economic efficiency
based on the prediction of occupant flow by applying innovative control law. Traditionally,
the response of AC system was defined by whether the room temperature exceeds desired
range. In this model, for operating hours, the temperature control would be initiated ahead
of time by predicting occupant flow from historical data and optimal indoor temperature
would be achieved before occupants’ arrival. This model is significant at the level to reduce
extreme temperature fluctuation and to maximize thermal comfort hours for occupants.

Introduction
Motivation and Background
The AC system of today has not changed much from that back in 20th century with
lack of control and responsive feedback. According to Quadrennial Technology Review from
Department of Energy, more than 76% of all U.S. electricity use and more than 40% of all
U.S. energy use as well as associated greenhouse gas (GHG) emissions were used to provide
comfortable, well-lit, residential and commercial buildings and to provide space conditioning
and lighting for industrial buildings. Nevertheless, the comfort level of customer is not closely
monitored and energy waste can be prevented through sophisticated control strategies. The
challenges for the control and optimization were to balance human comfort level and energy
consumptions. In order to address this issue, a responsive system was needed to correctly
predict the occupant flow in the built environment and allow optimal thermal experience for
the occupants.

The current control strategy for commercial AC system is simply on/off based on whether
the room temperature is beyond dead-band. However, there are two potential problems: 1)
The dead-band strategy could not prevent the temperature from going out of comfort zone or
would keep temperature zone too narrow compared with the comfort zone; 2) It’s unnecessary

UC Berkeley, CE295
156

to keep temperature in break time. Our project aims to improve these two problems through
customer flow estimation. With the customer flow estimation, our system can predict the
room temperature evolution and keep temperature in comfort zone precisely.

Focus of this Study

The aim of this project was to design a responsive AC system to maximize the occupant
comfort level and system economic efficiency based on the predicted occupant flow. From
the nature of temperature state equation, dynamic programming (Chapter 5) was adopted
in this project.

Literature Review
For the control technology part in AC realm, Zhao and Yu(2015) did a review on the
application of advanced control technologies in AC system and mentioned four main tech-
nologies: PID(proprtional-integral-derivative) control, artificial neural network, fuzzy control
and model predictive control. PID control extensively prevails in the modern AC control
system, because implementing the other three would require replacement or modification of
existing equipment. Besides, for artificial neural network, increasing computation and data
size decreases convergence speed; for fuzzy control, different designers have different clas-
sification principles which depend on experience; for model predictive control, large online
rolling calculation data causes data distortion and divergence.

For the occupant relevant AC control part, Godo, Haase, and Nishi(2017) proposed an
AC control based on pressure sensor in the chair detecting occupant location in a library.
Their idea on the control part is pretty simple: if there were occupants detected and the
comfort level was below certain limit, the heating temperature would be set to 20◦ C, other-
wise it would be set to 15◦ C. Wang, Feng and other eight researchers(2017) did predictive
control of indoor environment based on occupant number detected by video data and CO2
concentration. They incorporated occupant impact into their model by multiplying a factor
decided by sensitivity analysis of cooling load to occupant number. Basically, they used
real-time occupant data detected by sensors to adjust the AC on/off period to save energy
consumption. Their idea achieved faster AC response and made the indoor environment
more stable.

To summarize, the control technology used in the AC industry is fairly classic and there is
certainly room for improvement. As for the occupant related control, current achievement is
not satisfactory enough. With rapid development of computer technology, advanced building
automation era is sure to come, along with a giant leap in AC control strategy.

Key contributions
Our project explored the algorithm-based strategy to optimize AC system control with
occupant level in the built environment. The model includes the method of Markov Chain,
a predictive and probabilistic model, to determine occupant flow. The model is able to

UC Berkeley, CE295
157

adjust the AC operation power beforehand, further improving response speed and indoor
environment stability.

1 Modeling
1.1 Model Description

Figure 1: Project Overview

The setting was designated to build a simulation model for temperature control inside
a café. Figure.1 above demonstrated the system boundary where three terms changed the
observation states, 1) the energy exchange between interior and exterior state, 2) the working
power of the air conditioning units and 3) the heat dissipated from occupants inside the space.
The space was dimensioned to be 30 ft by 20 ft for length and width.

1.1.1 System Boundary

The system output is room temperature T (t). The system inputs include predicted am-
bient environmental temperature TA (t), predicted number of customers n(t) in the café and
thermal power removed by the AC unit Pe (t) and the AC mode s(t) ∈ {0, 1} corresponding
to off or on.

Notation Description Units

TA (t) predictable input t [◦ C]
n(t) predictable input t N/A
s(t) controllable input t N/A
T (t) performance output t [◦ C]

The stock is thermal energy stored in the room. The level or state variable, indicating
the amount of stored thermal energy, is T (t).

1.1.2 Energy Equation based on Conservation Law

The room temperature T (t) changes due to heat transfer with the outside environment
and with the AC unit. The First Law of Thermodynamics, in the form of Newton’s Cooling

UC Berkeley, CE295
158

Law, gives us:

d 1
C T (t) = [TA (t) − T (t)] − s(t)Pe + n(t)PI
dt R
where parameter C is the room’s thermal capacitance which was taken to be 0.5 [kW h/◦ C]
for this particular size of space. R is the thermal resistance between the room and outside,
25.3 [◦ C/kW ], and PI is internal heat generation per person, 0.1 kW .

1.2 Data Processing

There were two sets of data we implemented as inputs in the model. One was the out-
door temperature in the surrounding area. The other was the occupant flow or amount of
foot traffic entering the space. Both were recorded at hourly instances and then linearly
interpreted to the time interval as desired. The temperature data was collected at Climate
Data Online from agency of National Oceanic and Atmospheric Administration (NOAA).
The document provides a descriptive suite of statistics including the dates and hours when
it was collected, clouds coverage percentage for estimation of solar heat power if applicable
and lastly the mean air temperature observed by the stations. The temperature data was
referenced at the station in Oakland, CA. Due to the limitation of website service, only the
statistics from year of 1981-2010 was available.

Figure 2: The Average Trends of Occupant Flow from Monday to Sunday

The data for the occupant flow in the commercial space is difficult to obtain and often
the data is not accessible to the public. This data must be modeled to get a good estimate
of parameters. Thus, the occupant flow was evaluated in combination of the Popular Times
Graphs from Google, site visit and random sample generators based on Gaussian distribution.
Google Popular Times only specifies 7 foot traffic distribution of the café from Monday to
Sunday. From there, we recorded actual occupant data at a Starbucks Café at Albany, CA,
for a total of 10-hour interval. Then these numbers were scaled based on the trend to map
our average estimation as shown in Fig 2. For the simulation purpose, the occupant flow was

UC Berkeley, CE295
159

Level Range of Number of People Level Range of Number of People

1 0-2 6 24-29
2 3-7 7 30-34
3 8-13 8 35-39
4 14-18 9 40-45
5 19-23 10 > 46

Table 1: Griding of Occupant Flow

then randomly generated from Gaussian distribution with means determined by our average
model and standard deviation of 3.02 calculated from our 10 data points.

2 Simple Control
For the purpose of comparing and contrasting, a simple control model was incorporated
to test the effectiveness of our optimization strategy. The optimal temperature range was
defined to be from 24o C (75o F ) to 26o C (79o F ). The control rule is activated based on
the optimal time temperature. When the temperature is below the lower bound, heating
mode would be turned on. Alternatively, if the indoor temperature is higher than the upper
bound, cooling mode would be turned on. The cost was characterized in the similar fashion
as described in the Dynamic Programming section.

3 Dynamic Programming (DP)

3.1 Assumptions
In this project, the stochastic process for determining the trend of occupant flow was
assumed to follow the Markov Property. That means future state of occupant level depends
solely on present state. For the other random state variables such as outdoor temperature,
it was assumed that a reliable prediction of data would be provided and the errors between
actual data and predicted result were trivial.

3.2 Markov Chain Model

To begin with, the occupant flow was separated into k levels based on the principle that
occupant flow in same level will have similar influence on indoor temperature. In this project,
k is set to be 10, and the time interval is set to be 10 minutes (total time steps N = 144).

Since the flow of occupant follows Markov property, a series of matrices Pn , n = 0, 1, · · · , n−

1 can be formed to describe such stochastic process. Pn represent the transition matrix from
time step n to n + 1, in which pij is the probability that the system transfer to state j under
current state of i.
pij = P r(Xn+1 = j|Xn = i)

UC Berkeley, CE295
160

 
p00 p01 · · · p0k
p10 p11 · · · p1k 
Pn =  ..
 
.. .. . 
 . . . .. 
pk0 pk1 · · · pkk
As discussed before, different weekdays in a week have its own unique flow pattern and
there is no big difference between different weeks. With the data of occupant flow in past
year, those matrices can be calculated from dividing the frequency of specific transition by
the number of weeks. For instance, pij in Pn on Monday is calculated by
mij
P r(Xn+1 = j|Xn = i) = xij =
m
where mij is how many times the system transits from state i at timestep n, to state j at
timestep n + 1. m is the number of weeks. (There are 52 weeks available in this project.)

Figure 3: The Matrix of Markov Chain at n = 35

3.3 Stochastic Dynamic Programming

Given the Markov Chain model that dealt with the random process, a stochastic dynamic
programming can be formulated to solve the optimization problem.

3.3.1 Objective Function

The objective of this optimization problem is to minimize the expected cost, namely the
sum of electricity cost and penalty of temperature that went outside of comfort zone:
N
X −1
min J =E cu (n)∆tPe (k) + ccomf (T (n))
Pe (n),T (n),N (n)
n=0

UC Berkeley, CE295
161

N (n): Occupant flow level

T (n): Current room temperature, ◦ C.
Pe (k): Input power of Air conditioner, kW.
∆t: Time step, hour
cu (n): Electricity charge at every time interval, U SD/kW · h
ccomf (·): Penalty of discomfort, U SD

Subsequently, the penalty function ccomf (·) is a piecewise function which can be written
as 
2
(T − Tup )
 T > Tup
ccomf (T ) = 0 Tlow 6 T 6 Tup
 2
(T − Tlow ) T < Tlow


3.3.2 Constraints in the Model

1. To begin with, the AC units have maximum output for heating power and cooling
power:
max max
−Pcool 6 Preal (n) 6 Pheat
a negative sign is given to cooling power since in the dynamic equation, positive energy
numerically results a raising state while negative energy leads to a drop.

2. A hard bound of interior temperature are also imposed on this system, which means
even though there is leniency towards going out of comfort zone, ridiculous values (such
as 5 ◦ C and 40◦ C) are never allowed. That is

Tmin 6 T (n) 6 Tmax

3. There are also some equality constraints. First of all, the evolution of interior temper-
ature satisfies the dynamic equation mentioned before:

T (n + 1) = f (T (n), Preal (n), N (n))

4. The input power can be calculated given output power and corresponding efficiency:

Preal (n)


 Preal 6 0
Pe (n) = P ηcool
 real (n) Preal > 0

ηheat

3.3.3 Value Function and Principle of Optimality

The first step to implement a dynamic programming is to define the value function. Let
Vn (T (n), Nn ) be the minimum expected cost-to-go from time step n to the end, time step
N.

UC Berkeley, CE295
162

With such value function, the principle of optimality can be written as

Vn (T (n), N (n)) = min {cu (n)∆tPe (n) + ccomf (T (n)) + E [Vn+1 (T (n + 1), N (n + 1))]}
Pe (n)

= min {cu (n)∆tPe (n) + ccomf (T (n))

Pe (n)

+ E [Vn+1 (f (T (n), Preal (n), N (n)), N (n + 1))]}

= min {cu (n)∆tPe (n) + ccomf (T (n))
Pe (n)
X
+ pij Vn+1 (f (T (n), Preal (n), N (n)), N (n + 1) = j)}
j∈S

The boundary condition is

VN (T (N ), N (N )) = 0
And the optimal control strategy can be acquired by
Pe∗ (n) = γn (T (n), N (n)) = arg min {cu (n)∆tPe (n) + ccomf (T (n))
Pe (n)
X
+ pij Vn+1 (f (T (n), Preal (n), N (n)), N (n + 1) = j)}
j∈S

In consideration of accuracy and computational complexity, the interval of temperature

was set to 0.06◦ C, and the size of power grid was set to 60, which means the interval was no
more than 0.15 kilowatts. For the state variables that have no corresponding value in next
time step, interpolation is used to approximate. And increasing the density of grids didn’t
change Vk so much, implying the reliability of previous griding. Under such setting, the time
consumed for optimization was about 5 minutes.

3.4 Selection of AC Units

Besides the control strategy, model of AC units also has a significant influence on total
cost. First of all, AC with insufficient output power is not able to keep interior temperature
inside the comfort zone, leading to a high penalty. However, extremely powerful AC is
not economic as well for the high price, even though it can always maintain a comfortable
temperature. In order to cope with such trade-off between expense of AC and penalty of
discomfort, this project is designed to make a comparison between four different kinds of AC
units, and the table below shows the details.

Cooling Cooling Heating Heating

No. Price, USD
Power, kW Efficiency Power, kW Efficiency
1 2.7 5.51 2.9 4.58 2260
2 3.2 5.33 4.0 4.02 2700
3 4.2 5.46 4.7 3.86 3240
4 5.1 5.09 6.3 3.36 3600

Table 2: Parameters of Different AC Units

Pout
(Efficiency = )
Pin

UC Berkeley, CE295
163

4 Discussion & Results

A dataset with relatively large fluctuation in occupant flow level was selected as the test
set. And the results were presented as follows.

Figure 4: Output Power of AC Units

Figure 5: Interior Temperature under the Control of AC Units

UC Berkeley, CE295
164

Electricity Cost Economic Cost Penalty Total Cost

No.
USD/day USD/day USD/day USD/day
1 1.75 7.94 9.98 17.92
2 1.83 9.23 0.10 9.33
3 1.79 10.66 0.00 10.67
4 1.94 11.80 0.00 11.81

Table 3: Cost of Different AC Units

(Economic cost: sum of electricity cost and daily depreciation)

From Fig.4 and Fig. 5, the power of AC 3, 4 were not fully utilized during the process.
AC 1 had the lowest daily cost, but it was not capable to handle the performance during
the peak-time which led to extremely high penalty (even larger than the economic cost).
AC 2 found a balance between two scenarios. On the one hand, it kept temperature inside
comfort zone in most of the time. On the other hand, it was more cost effective and there
was no waste on surplus capability. Thus, AC 2 had the lowest cost in total, and was the
best choice for the setting of this project.

4.1 Comparing to Simple Control

Figure 6: Comparison Between Simple Model and Dynamic Programming

UC Berkeley, CE295
165

Electricity Cost Penalty Total Cost

Strategy
USD/day USD/Day USD/Day
Simple Control 3.80 9.21 13.01
Dynamic Programming 1.83 0.10 9.33

Table 4: Cost of Different Control Strategy

Under simple control, the delay of reaction to state changes caused unnecessary energy
consumption, making it hard to maintain comfortable temperature. Compared with simple
control, SDP control strategy significantly reduced electricity cost and discomfort. Armed
with statistic of random process, SDP would take future variance into consideration. There-
fore, SDP strategy made the AC warm up the room before opening time and adjusted the
energy consumption according to fluctuant electricity rate.

4.2 Expected Cost vs. Real Cost

Figure 7: Expected Cost vs. Real Cost

Since the objective function in stochastic dynamic programming is expected cost, the real
cost may have non-trivial deviation from it. Fig.7 shows the simulation result of 52 weeks
based on SDP strategy. The strategy was found to be reliable because real cost is smaller
than expected cost and the variance is reasonable.

4.3 Future Work

For the future consideration, the influence of solar radiation can be added in energy state
equation. It plays a significant role in prediction with current data from NOAA’s record
about cloud percentage. Secondly, more optimization techniques can be implemented. For
instance, the switching cost of shifting the AC power should be added to cost function. The
incentives were to allow more of smooth temperature transition and to keep AC system for
a longer service life. Overall, the model established was solid and it setups a framework for
more optimization of HVAC system in the future.

UC Berkeley, CE295
 166

Summary
The aim of the project was to design an autonomous control system for AC units to
account for different levels of occupant flow in daily life. Therefore, dynamic programming
algorithm coupled with Markov Chain model was employed in this project to determine the
optimal strategy across the time frame. In the end, two types of control were compared
and contrasted. One with simple strategy in which the system would be turned on once the
temperature was beyond comfort range. The other was the optimal strategy calculated from
SDP where we implemented cost function consisting of electric price, cost of AC units and
penalty for temperature beyond comfort range. As of the results, the optimal solution from
SDP has achieved what were stated in the cost function. Most of the temperature stayed in
the optimal range and it saved 51.8% energy consumption compared to the simple control
strategy.

References

Godo, S., Haase, J., & Nishi, H.(2015). Air Conditioning Control Using Self-Powered Sensor
Considering Comfort Level and Occupant Location. IECON 2015-41st Annual Conference
of the IEEE.
Wang, F., Feng, Q., Chen, Z., Zhao, Q., Cheng, Z., Zou, J.,...Reeve, H.(2017). Predictive
control of indoor environment using occupant number detected by video data and CO2 con-
centration. Energy & Buildings, 145, 155-162.
Zhao, Z. & Yu, N.(2017). The application of advanced control technologies in air condition-
ing system – a review. Advances in Building Energy Research, 11:1, 52-66.

Appendix
Code of Markov Chain
1 ## I d e n t i f y Markov Chain
2 # D i s c r e t i z e i r r a d i a n c e i n t o l e v e l s ( you can p i c k more than 10 i f you want )
3 l e v e l = 10
4 N g r i d = np . l i n s p a c e ( 0 . 0 , np . max( N 10min ) , l e v e l )
5 TT = np . a r a n g e ( N data . shape [ 0 ] )
6
7 t = np . r e m a i n d e r (TT/ 6 , 2 4 )
8
9 # P r e a l l o c a t e a r r a y s t o count t r a n s i t i o n s , and p r o b a b i l i t y t r a n s i t i o n
10 # matrix
11 c o u n t s = np . z e r o s ( [ l e v e l , l e v e l , 2 4 ∗ 6 ] )
12 P = np . z e r o s ( [ l e v e l , l e v e l , 2 4 ∗ 6 ] )
13 L e v e l s = np . a r a n g e ( l e v e l )
14
15 # Given v a l u e o f xi , r e t u r n t h e v a l u e c l o s e s t t o i t from x , y
16 d e f n e a r e s t i n t e r p ( xi , x , y ) :

UC Berkeley, CE295
 167

17 i d x = np . abs ( np . s u b t r a c t ( xi , x ) )
18 r e t u r n y [ i d x . argmin ( ) ]
19
20 f o r i d x i n r a n g e ( N data . shape [ 0 ] − 1 ) :
21
22 i i = n e a r e s t i n t e r p ( N data [ i d x ] , N grid , L e v e l s )
23 j j = n e a r e s t i n t e r p ( N data [ i d x +1] , N grid , L e v e l s )
24 kk = i n t ( round ( t [ i d x ] ∗ 6 ) )
25 # Increment t i m e s i r r a d i a n c e l e v e l g o e s from i i t o j j i n time s t e p kk
26 c o u n t s [ i i , j j , kk ] += 1
27
28 f o r i i in range ( l e v e l ) :
29 f o r kk i n r a n g e ( 2 4 ∗ 6 ) :
30 # Compute f r a c t i o n o f t i m e s i r r a d i a n c e l e v e l g o e s from i i t o j j i n
time s t e p kk
31 # out o f ALL t r a n s i t i o n s out o f l e v e l i i
32 i f np . sum ( c o u n t s [ i i , : , kk ] ) != 0 :
33 P [ i i , : , kk ] = c o u n t s [ i i , : , kk ] / np . sum ( c o u n t s [ i i , : , kk ] )

Code of Stochastic Dynamic Programming

1 # S t o c h a s t i c Dynamic Programming
2
3 # Simple c l o c k
4 s t a r t = time . time ( )
5 # C o e f f i c i e n t of discomfort penalty
6 alpha = 0.08
7 # Number o f A/C u n i t c a n d i d a t e s
8 n a c = P heat . s i z e
9 # Number o f occupant f l o w l e v e l
10 P e o p l e l e v e l = 10
11 # Time s t e p i n a day , i n t e r v a l i s 10 minutes
12 t i m e r a n g e = 24 ∗ 6
13 # Size of grid of temperature
14 g r i d s i z e t = 170
15 # S i z e o f g r i d o f Power
16 g r i d s i z e p = 60
17 # Temperature g r i d
18 t e m p g r i d = np . l i n s p a c e ( min ( t min ) , max( t max ) , g r i d s i z e t )
19
20 # Matrix t h a t s t o r e s v a l u e f u n c t i o n
21 v = np . i n f ∗ np . o n e s ( ( t i m e r a n g e + 1 , g r i d s i z e t , P e o p l e l e v e l , n a c ) )
22
23 # Matrix t h a t s t o r e s o p t i m i z o r
24 u s t a r = np . z e r o s ( ( t i m e r a n g e + 1 , g r i d s i z e t , P e o p l e l e v e l , n a c ) )
25
26 # I n i t i a l i z e value function
27 v [ time range , : , : , : ] = 0
28
29 s a v e e r r = np . s e t e r r ( a l l= ’ c a l l ’ )
30
31 # E v a l u a t e c u r r e n t temperature , g i v e s a p e n a l t y i f n e c e s s a r y
32 d e f c o m f o r t c o s t f u n c (Temp) :
33 i f Temp <= 26 and Temp >= 2 4 :

UC Berkeley, CE295
 168

34 return 0.0
35 e l i f Temp > 2 6 :
36 r e t u r n a l p h a ∗ (Temp −26) ∗∗2
37 e l i f Temp < 2 4 :
38 r e t u r n a l p h a ∗ (Temp −24) ∗∗2
39
40 # Convert output power t o i n p u t power
41 d e f pr2pe ( pr , n a c ) :
42 i f ( pr >=0) :
43 r e t u r n pr / e f h e a t [ n a c ]
44 else :
45 r e t u r n − pr / e f c o o l [ n a c ]
46
47 # Begin SDP f o r 4 d i f f e r e n t A/C u n i t s
48 f o r n in range ( n ac ) :
49 f o r i i n r a n g e ( t i m e r a n g e −1, −1, −1) :
50 f o r k in range (0 , g r i d s i z e t ) :
51 f o r i i in range (0 , P e o p l e l e v e l ) :
52 # Calculate f e a s i b l e f i e l d
53 l b = max( P c o o l [ n ] , − T 10min [ 0 ] [ i ] / R + (− C + 1 / R) ∗
t e m p g r i d [ k ] + C ∗ t min [ i ] − p I ∗ N g r i d [ i i ] )
54 ub = min ( P heat [ n ] , − T 10min [ 0 ] [ i ] / R + (− C + 1 / R) ∗
t e m p g r i d [ k ] + C ∗ t max [ i ] − p I ∗ N g r i d [ i i ] )
55
56 i f ( l b >= ub ) :
57 v [ i , k , i i , n ] = np . i n f
58 continue
59
60 p g r i d = np . l i n s p a c e ( lb , ub , g r i d s i z e p )
61
62 T next = ( ( T 10min [ 0 ] [ i ] − t e m p g r i d [ k ] ) / R + p g r i d + p I ∗
N grid [ i i ] ) / C + temp grid [ k ]
63
64
65 # C a l c u l a t e e x p e c t e d c o s t −to−go o f next s t e p
66 v n e x t = np . z e r o s ( [ P e o p l e l e v e l , g r i d s i z e p ] )
67 Ev next = np . z e r o s ( g r i d s i z e p )
68
69 f o r j in range (0 , P e o p l e l e v e l ) :
70 v n e x t [ j , : ] = np . i n t e r p ( T next , t emp gr id , v [ i + 1 , : , j ,
n])
71 f o r j j in range (0 , g r i d s i z e p ) :
72 i f ( np . i s n a n ( v n e x t [ j , j j ] ) o r v n e x t [ j , j j ] == np . i n f )
:
73 v n e x t [ j , j j ] = 1 e10
74
75 Ev next = P [ i i , j , i ] ∗ v n e x t [ j , : ] + Ev next
76
77 pe = np . z e r o s ( g r i d s i z e p )
78 f o r j in range ( g r i d s i z e p ) :
79 pe [ j ] = pr2pe ( p g r i d [ j ] , n )
80
81 i f ( i < 5 ∗ 6 o r i >= 20 ∗ 6 ) :
82 v [ i , k , i i , n ] = min ( c u [ i ] ∗ pe + Ev next )

UC Berkeley, CE295
 169

83 i n d e x = np . argmin ( c u [ i ] ∗ pe + Ev next )
84 else :
85 v [ i , k , i i , n ] = min ( c u [ i ] ∗ pe + c o m f o r t c o s t f u n c (
t e m p g r i d [ k ] ) + Ev next )
86 i n d e x = np . argmin ( c u [ i ] ∗ pe + c o m f o r t c o s t f u n c (
t e m p g r i d [ k ] ) + Ev next )
87
88 u s t a r [ i , k , i i , n ] = p grid [ index ]
89
90 end = time . time ( )
91 p r i n t ( end − s t a r t )

Acknowledgments
In this semester, all of us enjoyed studying CE295 a lot. We learned a variety of tools
for control and optimization. First of all, we would like to thank our instructor, Prof. Scott
Moura, not only for the interesting and inspiring materials, but also his rigorous attitude
during the course. Prof. Moura was so passionate and enthusiastic that we were immersed in
lectures and developed a deep understanding of all the materials. Secondly, we want to express
our thanks to our GSI Bertrand Travacca for all the extra workshops which supplemented the
course with some essential knowledge.

About the authors

All authors of this project are Master of Engineering student from UC Berkeley. Lorna
Chen, Yao Han, Sikang Hu are from Civil and Environmental department, majored in Sys-
tems Engineering. Xinhao Wang is from Mechanical Engineering department majored in
Energy.

UC Berkeley, CE295
170

UC Berkeley, CE295
171

Part IV

Batteries and Energy

Storage

UC Berkeley, CE295
172

Minimization of cumulative aging in batteries: a

grid-based approach
Yu-Hsin (Bryant) Huang, Yaser Marafee, Raja Selvakumar

Abstract
In this study we uniquely focus on the translation of portable size battery kinetic in-
formation to grid level analysis for a time-variant system. We characterize the effective
charging policies based on fluctuating demand and grid supply. We have constructed a dy-
namic programming framework to understand the effects of drawing power from the grid
when compared against the cumulative aging of the battery. We present a formulation that
includes two time-variant states and one control variable. We observe that with increasing
the tuning parameter α, we delay the charging/discharging of batteries and therefore ob-
serve the impact of cycle life on the battery. A capacity fade of less than 1% is observed
over a 24h time period, and fade of 7% is observed over a full week. Future work would be
keyed on simplifying some of the coupled and nonlinear constraints to additionally include
temperature dynamics and multiple battery nodes.

Introduction
Motivation and Background
Electrification of renewable energy integration and automobile transportation is essential
towards the reduction of greenhouse gas emission and therefore the impact of global warming
[1] The importance of energy storage devices can be seen in consumer electronics, electric
vehicles, and grid storage. With increasing diversification in renewable and intermittent
power supplies, larger integration of energy storage systems is imminent.
The main issue with batteries is aging during their lifetime due to the decrease in ca-
pacity, which leads to voltage decay and loss of power. There are numerous electrochemical
mechanisms to describe aging [2, 3, 4]. However, characterization is challenging due to di-
verse time scales and complex nonlinearities in these models. Dynamic parameter estimation
and battery state-of-charge (SOC) analysis remain complex topics.
As Figure 1 demonstrates, some of the studied impacts on battery aging is the cycle time,
depth-of-discharge (DOD), and storage life. It is observed that amongst these parameters,
it is the cycle lifetime that affects the lithium evolution the strongest. A more involved
discussion is presented in the Relevant Literature section; the purpose of this early statement
is to illustrate the direction of our research interests.

UC Berkeley, CE295
173

Figure 1. Evolution of active lithium upon aging of a Li-ion cell. Note that lithium evolution at
the anode increases more prominently with battery cycling. [1].

To that end, this project focuses on development of an optimization model which com-
bines the two challenging aspects of battery modelling: the grid storage and cell aging
mechanisms. With regards to the grid storage framework, we remodel the conventional grid
network as a source-sink problem. The input G(k) is defined as the grid power generated
at time k, which represents the variable power derived from combining nonrenewable and
renewable energy sources. Subsequently d(k) represents the time-variant demand response
that depends on the change in hourly demand from residential, commercial, and industrial
buildings. This transformation is shown in Figure 2. Note that this formulation is robust
for multiple batteries, although computational time scales exponentially with the addition
of each state. This progression is shown in the Appendix.
The second element of our project involves using a dynamic programming (DP) frame-
work to minimize cumulative aging of the batter(ies) presented in the system. We try to
investigate the capacity fade from parameters such as time, temperature, depth of discharge
(DOD), and discharge rate as presented in Equation 1:
−Ea
Qloss = B exp( )(Ah )z (1)
RT
where Qloss is the percentage of capacity loss, B is the pre-exponential factor, Ea is the
activation energy, R is the gas constant, T is the absolute temperature, and Ah is the amp-
hour throughput, which is expressed as Ah = (cycle number) · (DOD) · (full cell capacity),
and z is the power law factor. The work done by Wang et. al [5] presents a relationship for
the capacity fade from temperature and C-rate effects. The estimated values for a LiFePO4
battery are provided in the Technical Description section.
Our group consists of three chemical engineering masters students in the Product De-
velopment Program. We have all taken classes in electrochemical systems, ranging from

UC Berkeley, CE295
174

Figure 2. Grid transformation of a typical microgrid network into a max flow problem. The
inherent assumption is that we use prior knowledge on the evolution of G(k) and d(k) as supply
and demand. The remaining surplus or shortage will be taken from the batteries at each node i.
Figure adapted from [8].

mathematical fundamentals to first steps of solid-electrolyte interphase (SEI) modelling.

These experiences provide us an advantage to test the differences between black-box and
white-box modelling approaches. As a team, our goal is to learn advanced control and
parameter estimation techniques, so we hope to demonstrate this in our project results.
.

Relevant Literature
Battery aging can be classified into two main categories: calendar aging and cycle aging
[6]. The former is associated with the phenomena and the consequences of battery storage
and cycle aging corresponds to the influence of battery utilization time. Calendar aging
is the irreversible process of lost capacity during storage. The battery is degraded due to
self storage. Temperature effects greatly contribute to calendar aging as side reactions are
facilitated and cause capacity fade, as shown in Figure 3. The other principal variable under
investigation in calendar aging is the State of Charge (SOC). The state of charge is equivalent
to the ratio of the current battery energy to the maximum capacity. While operating at an
increased SOC may result in greater energy throughput, this also results in larger capacity
fade as indicated earlier in Equation 1. These mechanisms are described by electrochemical
models, performance based models, and equivalent circuit based (ECM) models.

UC Berkeley, CE295
175

Figure 3. Different battery modelling schemes. [7]

Focus of this Study

For the purposes of this study, we aim to consider the effects of temperature and cycle
aging on the battery SOC. As existing grid optimization models operate on the premise
of minimizing cost of electricity usage [10] or fuel usage [11] or some affine combination of
the two. We present here the first consideration of evaluating costs for a aging of larger
scale grid storage battery systems by using microkinetic analysis of effects of cycle time and
temperature on cumulative aging.

Key contributions
In this study we uniquely focus on the translation of portable size battery kinetic infor-
mation to grid level analysis for a time-variant system. We characterize the effective charging
policies based on fluctuating demand and grid supply.

Technical description
Assumptions
We first itemize the assumptions we have made for this model:

• Power can be drawn from the main power source, and we assume it has a maximum
capacity Gmax over all time N

• Power flow from the source is unidirectional (i.e. we only consider charging)

• The node represents a power distribution center, where the balance from power de-
manded and power transmitted is the power storage

• Battery temperature is constant with ambient conditions

UC Berkeley, CE295
176

• Power throughput is controlled at a constant C-rate1

• Grid batteries used are Li-ion

• Final state of charge is 0.3 (at the end of the simulation time step)

• Only one battery node is considered for the scope of the results of this project, while
the formulation is written for many nodes.

Formulation
We now define our grid and the appropriate parameters. Our operating framework was
first presented in Figure 2. Table 1 includes all the definitions for the variables and param-
eters used in this formulation.
The objective function is to minimize the cumulative aging across all nodes i ∈ [1, I], as
shown in Equation 2. Note that I represents the total nodes in the grid space. As afore-
mentioned, however, implementing a model for multiple nodes grows rapidly in complexity
without use of techniques like Approximate Dynamic Programming (ADP) and decoupling
the coupled affine constraint as given by Equation 7.
XX

min α ln Qi (k) + c(k)P0i (k) (2)
k∈N i∈I

EA Wh,i (k)
ln Qi (k) = ln(B) − + z ln ∀k ∈ 1..N, i ∈ 1..I (3)
RTi (k) Vi (k)
Wh,i (k + 1) = Wh,i (k) + |pb,i (k)|∆t ∀k ∈ 1..N, i ∈ 1..I (4)
Ei (k + 1) = Ei (k) − pb,i (k)∆t ∀k ∈ 1..N, ∀i ∈ 1..I (5)
P0i (k) = d(k) − pb,i (k) ∀k ∈ 1..N, ∀i ∈ 1..I (6)
X
P0i (k) ≤ G(k) ∀k ∈ 1..N, ∀i ∈ 1..I (7)
i∈I

Emin ≤ Ei (k) ≤ Emax ∀k ∈ 1..N, i ∈ 1..I (8)

0 ≤ Wh,i (k) ≤ Wh,max ∀k ∈ 1..N, i ∈ 1..I (9)

− Pbatt, max ≤ pb,i (k) ≤ Pbatt, max ∀k ∈ 1..N, i ∈ 1..I (10)

Equation 2 represents the cost minimization equation. α here is the key tuning parameter
that has thus far not been determined in literature surveys. In particular, if one tunes the
α
c(k)
ratio, this will present a minimization frontier that exists for a range of time evolving
values. While a Pareto Frontier wasn’t calculated for this project, this is an additional step
that we recommend.
1 P
This is based on the relationship between power and energy as: E = C-rate. Based on the work by
Wang et. al, we follow a C-rate of C/2

UC Berkeley, CE295
177

Equation 3 is the linearized logarithmic expression for the percent capacity loss. Note
mainly here that we have transformed the current throughput Ah,i to the power throughput
Wh,i . This is primarily done to describe the battery charging dynamics that are presented
in Equation ??. We describe the evolution of the battery dynamics in Equations 4-6. In
order to maintain some understanding of the energy state Ei , this transforms into a state
variable. Equation 7 now poses the main challenge for our model as with increasing numbers
of batteries, the optimization function will represent coupled dynamics. For future work, the
authors recommend starting with simple test cases (i.e. 2 nodes) with smaller grid sizes to
account for two-state dynamics repeated across two nodes. Equations 8-10 are simply the
state constraints on the variables.
 2

pb,i (k)
ρCVB Ṫi (k) = hAs Ti (k) − T∞ + RB Vi (k) ∀k ∈ 1..N, i ∈ 1..I (11)

Ti (k) + Ṫi (k)∆t = Ti (k + 1) ∀k ∈ 1..N, i ∈ 1..I (12)

Tmin ≤ Ti (k) ≤ Tmax ∀k ∈ 1..N, i ∈ 1..I (13)

Vi (k) = Voc,i (k) − Ii (k)RB ∀k ∈ 1..N, i ∈ 1..I (14)

Qcap, i (k) = Qmax 1 − Qi (k) ∀k ∈ 1..N, i ∈ 1..I (15)
0 ≤ Ii (k) ≤ Imax ∀k ∈ 1..N, i ∈ 1..I (16)

The above equations represent the temmperature dynamics of the system, which is where
the major complications arise in the battery state dynamics. We present the formulation
but recognize that due to computational limitations, we are unable to pursue the effects of
temperature evolution on battery aging. A further study could linearize and otherwise relax
some of the provided constraints.
We now discuss the dynamic programming (DP) framework for this project. Let V (k)
represent the cumulative capacity fade and power generation from time step k to total time
N. We define control variables pb,i (k) as uk ∀i and state variables Wh,i (k), Ei (k) as xk ∀i:
P


Vk (xk ) = min i∈I α · Qi (k) + c(k) · di (k) − ui (k)
uk ,xk
+ V (k + 1) ∀k ∈ 1..N (17)

We finally establish the boundary condition:

V (N + 1) = 0

Table 1 presents a cumulative list of all the parameters and values used in the formulation.

Discussion
We now present the main conclusions from this project. Figure 4 represents the battery
state evolution over a full day with only considering fluctuation in grid energy price, and not
considering the battery aging cost. This policy minimizes the use of the battery in general

UC Berkeley, CE295
178

Table 1. List of parameters and variables used in formulation

Symbol Description Units Value

α Fixed cost of capacity loss [$] 50
c(k) Grid energy cost at time k [$/kW] 0.02a
Qi (k) % capacity loss of node i at k [-]
B Pre-exponential factor [-] 30,330
EA Activation energy [J/mol] -31,500
R Gas law constant [J/mol*K] 8.314
T∞ Ambient temperature [K] 298
z Power law factor [-] 0.552
Vi (k) Voltage at node i at time k [J]
Wh,i (k) Power throughput [J]
Ei (k) Energy state [J]
pb,i (k) Power surplus/shortage [W]
P0i (k) Power delivered from grid to i [W]
G(k) Grid variable power supply [W] 50E3b
d(k) Variable demand [W] 25E3
∆t Simulation time step [h] 1
N Total simulation time [h] 24
c
Emin ,Emax Min and max energy limits [W-h] 792, 79200d
Wh,min ,Wh,max Min and max power throughput limits [W-h] 0, 24·Emax
Pbatt,max Maximum battery charge/discharge limit [W] 0.9·Emax /dt
a
Time variant cost calculated from hourly electricity given in CE295 HW3, Spring 2018
b
Value estimated based on existing grid capacities, this is the max value
c
This is derived from the C-rate.
d
Determined using the A-h throughput from Wang et. al [5].

and just meets the demand as required. This is different from the results shown in Figure
5, where as α increases, there is a noticable delay in charging/discharging from the battery.
This indicates that the DP recognizes the aging equation accounts for the time evolution of
the battery state and tries to draw more power from the grid (as long as its within the the
operating limits of the battery). This is a remarkable finding that shows the power of the
minimization function. Finally, Figure 6 proves this hypothesis by showing for a week-long
optimization, in the first day the demand does not draw power from the battery at all! So
cumulative aging is decreased because overall usage of the battery decreases itself. We’d
also like to note here that the calculated values of 7% capacity fade over a week long period
matches what we expect from literature values [5].
We would now like to enumerate the key takeaways from this project and further discus-
sion items:

• The equation for capacity loss is based on a 2.2Ah battery which lead to an unreason-
able capacity loss when the battery that is being minimized has a much larger capacity.

UC Berkeley, CE295
179

Figure 4. Grid based optimization using α = 0

Thus, we had to tune the equation to give reasonable capacity losses that would yield
similar quantitative results to the plots provided in literature.

• In order to incorporate more batteries in the system, we would need to account for
charge/discharge power for each battery in the constraints. We faced a great challenge
in trying to implement this in dynamic programming. Perhaps using a ADP framework
and relaxing some constraints could lead to convex optimization.

• We were able to corporate temperature as a third state, however, due to high com-
putational requirements, the grid size had to be reduced which resulted in inaccurate
interpolations because the grid size is small. Since the grid size ranges of the energy
state, power throughput, and temperature are all different, this yields different grid
sizes and therefore parameter sensitivity would be quite high.

• One would consider the practical values used in the simulation. Of course we have
estimated parameter values based on previous homeworks and some literature survey,
but this system should be robust to handle more realistic parameters. If not, the
authors advise exploring a modified capacity loss equation that may include some
post-exponential ”correction factor” that accounts for this.

Executive Summary
In this study we uniquely focus on the translation of portable size battery kinetic in-
formation to grid level analysis for a time-variant system. We characterize the effective
charging policies based on fluctuating demand and grid supply. We have constructed a dy-
namic programming framework to understand the effects of drawing power from the grid

UC Berkeley, CE295
180

Figure 5. Grid based optimization using different values of α

Figure 6. Week long optimization based on α = 50

UC Berkeley, CE295
181

when compared against the cumulative aging of the battery. We present a formulation that
includes two time-variant states and one control variable. Our results indicate control poli-
cies that fluctuate with the tuning parameter α, which represents the cost of losing power to
capacity fade. We observe that with increasing this tuning parameter, we delay the charg-
ing/discharging of batteries and therefore observe the impact of cycle life on the battery.
Future work would be keyed on simplifying some of the coupled and nonlinear constraints
to additionally include temperature dynamics and multiple battery nodes.

References
[1] Palacin, M. R.; de Guibert, A., ”Why do batteries fail?” Science 2016, 351 (6273), 7.

[2] Baek, K.W., Hong, E.S. & Cha, S.W. Int. J Automot. Technol. (2015) 16: 309.
https://doi.org/10.1007/s12239-015-0033-2

[3] Pinson, M. B., Bazant, M. Z (2013). ”Theory of SEI Formation in Rechargeable Bat-
teries: Capacity Fade, Accelerated Aging and Lifetime Prediction.” Journal of The
Electrochemical Society. Volume 160. Pages A243-A250.

[4] Vetter, J.; Novak, P.; Wagner, M. R.; Veit, C.; Moller, K. C.; Besenhard, J. O.; Win-
ter, M.; Wohlfahrt-Mehrens, M.; Vogler, C.; Hammouche, A., ”Ageing mechanisms in
lithium-ion batteries.” J. Power Sources 2005, 147 (1-2), 269-281.

[5] Wang, J.; Liu, P.; Hicks-Garner, J.; Sherman, E.; Soukiazian, S.; Verbrugge, M.;
Tataria, H.; Musser, J.; Finamore, P., ”Cycle-life model for graphite-LiFePO4 cells.” J.
Power Sources 2011, 196 (8), 3942-3948.

[6] Barre, A.; Deguilhem, B.; Grolleau, S.; Gerard, M.; Suard, F.; Riu, D., ”A review on
lithium-ion battery ageing mechanisms and estimations for automotive applications.”
J. Power Sources 2013, 241, 680-689.

[7] Jaguemont, J.; Boulon, L.; Dube, Y., ”A comprehensive review of lithium-ion batteries
used in hybrid and electric vehicles at cold temperatures.” Applied Energy 2016, 164,
99-114.

[8] Plett, G. L., ”Extended Kalman filtering for battery management systems of LiPB-based
HEV battery packs.” Journal of Power Sources 2004, 134, 262-276

[9] de Azevedo, Ricardo Cintuglu, Mehmet Ma, Tan Mohammed, Osama. (2016).
Multiagent-Based Optimal Microgrid Control Using Fully Distributed Diffusion Strat-
egy. IEEE Transactions on Smart Grid. 1-1. 10.1109/TSG.2016.2587741.

[10] G. C. Lazaroiu, V. Dumbrava, G. Balaban, M. Longo and D. Zaninelli, ”Stochastic

optimization of microgrids with renewable and storage energy systems,” 2016 IEEE
16th International Conference on Environment and Electrical Engineering (EEEIC),
Florence, 2016, pp. 1-5. 10.1109/EEEIC.2016.7555486

UC Berkeley, CE295
182

[11] F. A. Mohamed and H. N. Koivo, ”Online Management of MicroGrid with Bat-

tery Storage Using Multiobjective Optimization,” 2007 International Conference on
Power Engineering, Energy and Electrical Drives, Setubal, Portugal, 2007, pp. 231-236.
10.1109/POWERENG.2007.4380118

Appendix

Figure A 1. Computational scaling of increasing number of states and number of nodes processed.
Increasing the number of states increases the complexity exponentially.

Acknowledgments
We’d like to thank Professor Scott Moura for his extensive feedback and help in developing
the formulation for this project. Most helpful was breaking the problem down into solvable
portions and determining what is feasible and what is practical in emulating battery state
dynamics.
We’d also like to acknowledge Bertrand Travecca for providing astute observations on
simplifying some constraint of the problem and how to further develop the project analysis.

About the authors

Yu-Hsin (Bryant) Huang is a MS candidate in Chemical Engineering under the Product
Development Program (PDP) at UC Berkeley. His career interests include PHEV and
battery management.

UC Berkeley, CE295
183

Yaser Marafee is a MS candidate in Chemical Engineering under the Product Develop-

ment Program (PDP) at UC Berkeley. His career interests include process optimization
and sustainable fuels.

Raja Selvakumar is a MS candidate in Chemical Engineering under the Product Devel-

opment Program (PDP) at UC Berkeley. His career interests lie in the intersection of
machine learning and chemical process design.

UC Berkeley, CE295
184

UC Berkeley, CE295
185

States Estimation of Li-ion Battery

Junzhe Shi, Franklin Zhao, Ruitong Zhu, and Xin Peng

Abstract
Batteries are ubiquitous in all forms of electronics and transportation, and a key to the
store of clean and secure energy. For different kinds of batteries, Li-ion battery is the most
prominent one for their superior gravimetric and volumetric energy density. For the safe
operation of Li-ion battery, the state of charge (SOC) and state of health (SOH) estimation
are of great significance. Hence, the goal of the project is to design a robust observer which
can estimate the SOC and SOH of Li-ion batteries. In the project, the equivalent-circuit
model is used for the battery modeling with current and ambient temperature as inputs and
voltage as the measured output. The equivalent-circuit model includes three parts which
are an electrical model, a thermal model, and an aging model. To ensure the accuracy of
states estimation, the Extend Kalman Filter (EKF) is applied and examined in the project.
The battery system is constructed and simulated using MATLAB. The best observer built
in this project is a Voltage-Temperature (VT) observer which can accurately observe SOC
with great robustness, while SOH can be observed using open-loop observer. The robustness
of designed observer is tested using the wrong initial estimates and wrong model parameters.

Introduction
Motivation and Background
The identification of battery operation and aging in real life has been a long-desired
yet challenging goal, which includes multiple complex processes in complicated operating
conditions and environments. An accurate method to observe SOC and SOH of Li-ion battery
is in need. Meanwhile, batteries invariably work at varying thermal and aging conditions.
Thus, it is necessary for us to build a battery observation system to monitor operation and
aging of battery. The potential challenges also exist. We need to express a multi-control
problem via mathematical equations and combine electrical model, thermal model and aging
model. Besides, since all team members are major in Civil Systems, a lack of background
knowledge in electrical engineering can be a big challenge. However, our previous course CE
291F, Control and Optimization of Distributed Parameters Systems can be helpful for the
project. It gave us a background knowledge of partial differential equations, conservation
laws, linear stability, Kalman filter and so on. We all have experiences of building, controlling
and optimizing systems, including quench process, heat diffusion and Lighthill Whitham
Richards model.

UC Berkeley, CE295
186

Focus of this Study

In this project, we will focus on the SOC and SOH of Li-ion batteries. Based on
equivalent-circuit, the electrical, thermal and aging models will be developed for the observ-
ing system. Since battery monitoring and management can be the key to allowing innovation
in future designs because of their limit properties, our system may play an important role
in such an area, and significantly contribute to the energy saving and efficiency.

Literature review
The behavior of batteries is difficult to predict because of its non-linearity, and there
has been quite a few attempts to model and estimate the inner state of the system. To
better describe the behavior of batteries, a lumped-parameter electro-thermal model was
introduced to capture the correlation between the thermal and electric behavior [1], a semi-
empirical cycle-life model was established to investigate the attributes of capacity loss [2]. As
a combination of the two models above, a coupled electro-thermal-aging model is developed
in [3], which captured the systematic dynamics for lithium-iron-phosphate batteries. This
model provides a decent assumption for the battery behaviors, which we will go through in
detail at the section of mathematical model, as well as an open-loop observer for the Sate
of Charge (SOC) and State of Health (SOH). Inspired by the course material of CE 295, we
tried to go further to investigate more possible approaches to estimate the battery states.

Key contributions
Based on the thermal-electrical-aging model as discussed above, we build a robust ob-
server for the model, which provides users with an efficient method to monitor State of
Charge and State of Health for battery.

1 Technical Description
1.1 Mathematical Model
Our analysis is based on a coupled electro-thermal-aging model for lithium-iron-phosphate
batteries, which is introduced in [3]. The model consists of a two RC pair electrical model,
a two-state thermal model and a semi-empirical aging model.

1.1.1 Electrical Model

As shown in Figure 1, the electrical comprises an open-circuit voltage (OCV, VOC ), two
resistor-capacitor (RC) pairs (R1 , C1 , R2 , C2 ), and an ohmic resistor (R0 ). The state-space
model is given by:
dSOC I(t)
(t) = (1)
dt Cbat
dV1 V1 (t) I(t)
(t) = − + (2)
dt R1 C1 C1

UC Berkeley, CE295
187

dV2 V2 (t) I(t)

(t) = − + (3)
dt R2 C2 C2
Vt (t) = VOC (SOC) + V1 (t) + V2 (t) + R0 I(t) (4)
where Cbat is the nominal capacity of the battery, I(t) is the current (positive for charging),
and Vt (t) denotes the terminal voltage. Three state variables are SOC and voltages across
the two RC pairs V1 , V2 .

Figure 1: Electrical Model [3]

The electrical parameters are identified in [4]. In our model, we follow the equations listed
in the appendix to derive these parameters based on the state of charge (I < 0) or discharge
(I ≥ 0).

1.1.2 Thermal Model

Figure 2: Two-state Thermal Model

Since the core temperature can be higher than the surface temperature under high current
rates [4], a two-state thermal system was hereby introduced to capture both core and surface
temperature dynamics. As sketched in Figure 2, the radial heat transfer dynamics of a
cylindrical battery can be described as follow.
dTc (t) Ts (t) − Tc (t) Q(t)
= + (5)
dt Rc Cc Cc

UC Berkeley, CE295
188

dTs (t) Tf (t) − Ts (t) Ts (t) − Tc (t)

= + (6)
dt R u Cs Rc Cs
Rc , Ru , Cc and Cs represent the heat conduction resistance, convection resistance, core heat
capacity and surface heat capacity respectively, with their values shown in Table 1; two state
variables are core temperature Tc and surface temperature Ts ; the ambient temperature Tf
is treated as uncontrollable input.

Table 1: Thermal Parameters

Rc (KW −1 ) Ru (KW −1 ) Cc (JK −1 ) Cs (JK −1 )
1.94 3.08 62.7 4.5

Q(t) = |I(VOC − Vt )| is heat generation including joule heating and energy dissipated by
electrode over-potentials, based on equation (4), we can rewrite equation (5) as

dTc (t) Ts (t) − Tc (t) I(t)(V1 (t) + V2 (t) + R0 I(t)

= + (7)
dt Rc Cc Cc

1.1.3 Aging Model

The aging model is based upon a matrix of cycling tests from [2]. The experiment results
suggest that capacity fade depends strongly on C-rate and temperature in the cell at low
charge/discharge rates, while the sensitivity to depth-of-discharge is negligible. The semi-
empirical life model adopted the following equation to describe the correlation between the
capacity loss (∆Qb , in %) and the discharged Ah throughput(A, depends on C-rate),
 
−Ea (c)
∆Qb = M (c)exp A(c)z (8)
RTc

where M (c) is the pre-exponential factor as a funciton of C-rate, which is denoted by c.

The relation between the pre-exponential factor M (c) and C-rate are shown in Table 2. The
activation energy Ea and the power-law factor z are given by

Ea (c) = 31700 − 370.3c z = 0.55 (9)

Table 2: Pre-exponential Factor as a function of the C-rate

C-rate c 0.5 2 6 10
M 31630 21681 12934 15512

The model consider a capacity loss of 20% as the end-of-life (EOL) for an automotive battery.
The corresponding Ah throughput Atol and the number of cycles N are therefore calculated
as below.  1 z
20
Atol (c, Tc ) =    (10)
−Ea (c)
M (c)exp RTc

UC Berkeley, CE295
189

3600Atol (c, Tc )
N (c, Tc ) = (11)
Cbat
Each cycle correspondes to 2Cbat charge throughput, and since Atol is discharged Ah through-
put, the total throughput including both charged and discharged Ah should be 2Atol . Based
on this, the battery State-of-Health (SOH) is defined as:
Rt
|I(τ )dτ
SOH(t) = SOH(t0 ) − t0 (12)
2N (c, Tc )Cbat

where t0 denotes the initial time. SOH varies among [0, 1], SOH = 1 correspondes to a
brand new battery and SOH = 0 means 20% capacity loss, as known as EOL. The derivative
of SOH yields the battery aging model

dSOH |I(t)|
(t) = − (13)
dt 2N (c, Tc )Cbat

1.1.4 Model Coupling

Figure 3: Electro-Thermal-Aging Model Coupling

Combining the above three subsystems, the model dynamics are summarized as below.

dSOC I(t)
(t) = (14)
dt Cbat
dV1 V1 (t) I(t)
(t) = − + (15)
dt R1 C1 C1
dV2 V2 (t) I(t)
(t) = − + (16)
dt R2 C2 C2
dTc (t) Ts (t) − Tc (t) I(t)(V1 (t) + V2 (t) + R0 I(t)
= + (17)
dt Rc Cc Cc
dTs (t) Tf (t) − Ts (t) Ts (t) − Tc (t)
= + (18)
dt R u Cs Rc Cs

UC Berkeley, CE295
190

dSOH |I(t)|
(t) = − (19)
dt 2N (c, Tc )Cbat
Inputs in this model include current I( t), which is controllable, and ambient temperature
Tf (t), which is uncontrollable.

1.2 Observer & Simulation

All the parameters used in this project base on the parameter of an A123 Li-ion battery.
In the 1 hour of simulation, the controllable input is charging rate which is a constant value
0.9C. The uncontrollable input, ambient temperature, is assumed as a half sine wave. The
input values are presented in Figure 4.

Figure 4: Systems inputs

In the simulation outputs, SOC reached 90% after 1-hour charging. The terminal voltage
also increases as the SOC increases. SOH decreased as the time goes by. The core and
surface temperatures of Li-ion battery changed under the effects of the current and the
ambient temperature. The simulation outputs are showed in Figure 5.

UC Berkeley, CE295
191

Figure 5: Systems outputs

By adding white noises in the terminal voltage and surface temperature data with SNR
is equal to 60, the synthetic data is generated from this simulation model to produce mea-
surements of terminal voltage and surface temperature. The synthetic measurements are
presented in Figure 6.

Figure 6: Measured outputs

By using Extended Kalman Filter method, three different observers are created. The

UC Berkeley, CE295
192

components of these three observers are showed on Table 3.

Table 3: Three observers

Observers Control inputs Measured inputs
Voltage-measured observer Tf , Current VT
Temperature-measured observer Tf , Current TS
VT-measured observer Tf , Current V T , TS

In order to test the robustness of the three observers, these observers were tested with
wrong initial estimates and wrong parameter values respectively.

1.2.1 Voltage-measured observer

The test results of the Voltage-measured observer with wrong initial estimates and wrong
parameter values in showed in Figure 7 and Figure 8 respectively.

Figure 7: Estimation test of voltage-measured observer with wrong initial values

UC Berkeley, CE295
193

Figure 8: Estimation test of voltage-measured observer with wrong parameters

For the voltage-measured observer, the SOC and terminal voltage are well estimated.
The estimated surface and core temperatures of battery converge with their true values after
1000 seconds. The estimated SOH does not converge to it true value at all. The result shows
that SOH is unobservable. It is because the observability matrix of the voltage-measured
observer is not full rank. The null space to the observability matrix is span([0, 0, 0, 1, 0, 0]T ,
[0, 0, 0, 0, 1, 0]T , [0, 0, 0, 0, 1]T ). It means the states of Tc , Ts , and SOH are unobservable
for this observer. However, when the observer’s model is very accurate, the temperatures
will converge to their true values, even if temperatures are unobservable for the voltage-
measured observer. It is because the thermal system is asymptotically stable, the surface
and core temperatures will reach their equilibrium states after a long time.

1.2.2 Temperature-measured observer

The test results of the Temperature-measured observer with wrong initial estimates and
wrong parameter values in showed in Figure 9 and Figure 10 respectively.

UC Berkeley, CE295
194

Figure 9: Estimation test of temperature-measured observer with wrong initial values

Figure 10: Estimation test of temperature-measured observer with wrong parameters

For the temperature-measured observer, the surface and core temperatures of the battery
are well estimated. The estimated SOC and SOH does not converge to their true values.
The test result shows that SOC and SOH are unobservable for the temperature-measured
observer. The rank of the observability matrix of the observer is 4 which is not full rank.
The null space to the observability matrix is span([1, 0, 0, 0, 0, 0]T , [0, 0, 0, 0, 0, 1]T ). The null
space also indicates the unobservability of SOC and SOH for temperature-measured observer.

UC Berkeley, CE295
195

1.2.3 Voltage and temperature (VT) -measured observer

The test results of the VT-measured observer with wrong initial estimates and wrong
parameter values in showed in Figure 11 and Figure 12 respectively.

Figure 11: Estimation test of VT-measured observer with wrong initial values

Figure 12: Estimation test of VT-measured observer with wrong parameters

For the VT-measured observer, VT , TC , TS , and SOC of the battery are well estimated,
except the SOH. The rank of the observability matrix of the observer is 5. The null space to
the observability matrix is the span([0, 0, 0, 0, 0, 1]T ) which also indicates the unobservability
of the SOH.

UC Berkeley, CE295
196

1.2.4 Comparsion
To better evaluate the performance of these three kinds of observers, the observability
of 6 battery states of the observers are presented on Table 4. The root-mean-squared error
(RMSE) of three observers and two tests are showed on Table 5 and Table 6.

Table 4: The observability three observers

Observers SOC% V1 V2 TS TC SOH%
Voltage-measured observer Yes Yes Yes No No No
Temperature-measured observer No Yes Yes Yes Yes No
VT-measured observer Yes Yes Yes Yes Yes No

Table 5: RMSE of Extended Kalman Filter with wrong initial values

Observers VT TS TC SOC% SOH%
Voltage-measured observer 0.0163 0.0060 0.0607 0.2848 0.1009
Temperature-measured observer 0.0861 0.0055 0.0554 51.3642 0.1004
VT-measured observer 0.1063 0.0059 0.0594 0.2758 0.1009

Table 6: RMSE of Extended Kalman Filter with wrong parameters in models

Observers VT TS TC SOC% SOH%
Voltage-measured observer 0.0004 0.0090 0.0878 0.7590 0.0669
Temperature-measured observer 0.0034 0.0078 0.0798 51.4638 0.0669
VT-measured observer 0.0004 0.0083 0.0848 0.7551 0.0669

2 Discussion
As the results shown in the previous part, for temperature observer, SOC, SOH is unob-
servable. For voltage observer, TC , TS , SOH are unobservable. However, the temperature
can be affected by the resistance of the circuit. On the other hand, the resistance will be
affected by the temperature as well. Hence, the voltage observer can actually “observe” Tc
and TS by an indirect approach since after a long time, the temperature would converge.
Hence, we dicided to combine the templeature observer and voltage observer together
as a VT-observer. As the results are shown, VT-measured observer has an overwhelming
advantage. The observer can estimate most of the states because it the combines both V-
measured observer and T-measured observer. In this project, there is not any observer can
estimate the SOH. However, if an accurate battery model and initial values can be provided,
the SOH values can be well estimated by an open-loop observer.
In addition, from the simulation result we notice that SOC has a positive linear relation-
ship with charging time while SOH has a negative linear relationship with charging time. In
the simulation, it requires nearly 2 hours to replenish the SOC from 0% to 90%, with the
associated 0.005% SOH decay. The highest SOC is 87.4%, and the lowest SOH is 99.995%
compared to the original status. Such a trend indicates that charging time and battery
health is a trade off. This is interesting, since a balance between efficiency and safety, is

UC Berkeley, CE295
197

raised. If the highest efficency is needed, then the charging time should be minimized; If
the highest safety is needed, then the aging condition should be minimized. However, in
real-word cases, a balance point need should be find in different scenarios, which leads to
a new topic – optimal control for battery charging. Since batteries are storing most of the
energy that we use in our daliy life, the observer we developed, as such a useful tool for the
optimal control problems, will definitiely advance sustainablity in energy systems. Hence
it can be noticed that state estimation is a really powerful tool for maintaining the energy
system. In the future, a student team can focus on optimal control of the battery charging,
based on the observer we developed.

Executive Summary
In this project, we built the electric model, thermal model, and aging model, and im-
plemented the open-loop simulation as well as state estimation using EKF. The goal of this
project is to develop a battery observing system with high efficiency and robustness. Since
with the accurate initial values and model parameters, the open-loop observer can observe
the SOC and SOH pretty well, the trend of those curves are exactly as we expected. How-
ever, with wrong inital estimates and wrong parameter values, SOC cannot be observed by
temperature observer, and SOH cannot be observed by both observers. Hence, for SOC in
this scenario, it can be well-observed using our VT-observer; SOH can be observed using
open-loop observer with correct initial estimates and parameters. Future work will focus on
the optimal control of battery charging based on these observers we developed.

References
[1] C. R. Gould, C. M. Bingham, D. A. Stone, and P. Bentley, “New battery model and
state-of-health determination through subspace parameter estimation and state-observer
techniques,” IEEE Transactions on Vehicular Technology, vol. 58, no. 8, pp. 3905–3916,
2009.

[2] J. Wang, P. Liu, J. Hicks-Garner, E. Sherman, S. Soukiazian, M. Verbrugge, H. Tataria,

J. Musser, and P. Finamore, “Cycle-life model for graphite-lifepo4 cells,” Journal of
Power Sources, vol. 196, no. 8, pp. 3942–3948, 2011.

[3] H. E. Perez, X. Hu, S. Dey, and S. J. Moura, “Optimal charging of li-ion batteries with
coupled electro-thermal-aging dynamics,” IEEE Transactions on Vehicular Technology,
vol. 66, no. 9, pp. 7761–7770, 2017.

[4] H. E. Perez, J. B. Siegel, X. Lin, A. G. Stefanopoulou, Y. Ding, and M. P. Castanier,

“Parameterization and validation of an integrated electro-thermal cylindrical lfp battery
model,” in ASME 2012 5th Annual Dynamic Systems and Control Conference joint with
the JSME 2012 11th Motion and Vibration Conference. American Society of Mechanical
Engineers, 2012, pp. 41–50.

UC Berkeley, CE295
198

Appendix
The parameters of equation (1)-(4) are identified as follow [4], which varied across the
state of charge (I < 0) or discharge (I ≥ 0).
 
R1d I ≥ 0
R1 = (20)
R1c I < 0
 
2 Tref R1 ∗
R1∗ = (R10∗ + R11∗ (SOC) + R12∗ (SOC) )exp (21)
Tm − Tshif tR1 ∗

Table 7: PARAMETRIC R1 FUNCTION PARAMETERS

R10d R10c R11d R11c R12d
7.1135e-4 0.0016 -4.3865e-4 -0.0032 2.3788e-4
R12c Tref R1 d Tref R1 c Tshif tR1 d Tshif tR1 c
0.0045 347.4707 159.2819 -79.5816 -41.4578

 
R2d I ≥ 0
R2 = (22)
R2c I < 0
 
2 Tref R2 ∗
R2∗ = (R20∗ + R21∗ (SOC) + R22∗ (SOC) )exp (23)
Tm

Table 8: PARAMETRIC R2 FUNCTION PARAMETERS

R20d R20c R21d R21c
0.0288 0.0113 -0.073 -0.027
R22d R22c Tref R2 d Tref R2 c
0.0605 0.0339 16.6712 17.0224

 
C1d I ≥ 0
C1 = (24)
C1c I < 0
C1∗ = C10∗ + C11∗ (SOC) + C12∗ (SOC)2 + (C13∗ + C14∗ (SOC) + C15∗ (SOC)2 )Tm (25)

Table 9: PARAMETRIC C1 FUNCTION PARAMETERS

C10d C10c C11d C11c
335.4518 523.215 3.1712e+3 6.4171e+3
C12d C12c C13d C13c
-1.3214e+3 -7.5555e+3 53.2138 50.7107
C14d C14c C15d C15c
-65.4786 -131.2298 44.3761 162.4688

UC Berkeley, CE295
199

 
C2d I ≥ 0
C2 = (26)
C2c I < 0
C2∗ = C20∗ + C21∗ (SOC) + C22∗ (SOC)2 + (C23∗ + C24∗ (SOC) + C25∗ (SOC)2 )Tm (27)

Table 10: PARAMETRIC C1 FUNCTION PARAMETERS

C20d C20c C21d C21c
3.1887e+4 6.2449e+4 -1.1593e+5 -1.055e+5
C22d C22c C23d C23c
1.0493e+5 4.4432e+4 60.3114 198.9753
C24d C24c C25d C25c
1.0175e+4 7.5921e+3 -9.5924e+3 -6.9365e+3

Acknowledgments
The authors would like to thank Professor Scott Moura for guidance in this project.

About the authors

Junzhe Shi is a Systems Engineering student (Master of Science) at University of Califor-
nia, Berkeley.

Franklin Zhao is a Systems Engineering student (Master of Science) at University of Cal-

ifornia, Berkeley.

Ruitong Zhu is a Systems Engineering student (Master of Science) at University of Cali-

fornia, Berkeley.

Xin Peng is a Systems Engineering student (Master of Engineering) at University of Cali-

fornia, Berkeley.

UC Berkeley, CE295
200

UC Berkeley, CE295
201

Part V

Health and Environment

UC Berkeley, CE295
202

Mathematical Modeling of An Ecosystem Network

Kyra Chang, Tiffany Chang, Emily Farrar, Nate Tsang

Abstract
Complex systems are present in every facet of our environment. The vast biodiversity of
life and its interconnected networks are an area of research for ecologists and modelers alike.
In our project, we demonstrate the effects of various controllable inputs, such as adding
or removing species through introduction and hunting, on the population dynamics of an
ecosystem as well as three different types of species’ interaction: predator-prey, competition,
and commensal. The main research of the paper is focused on modeling predator-prey
relationships using food chain dynamics. In our first mathematical model, we predict how
populations evolve over time given interactions with each other and the environment. Our
second model utilizes dynamic programming to determine the optimal control strategy that
achieves a target population size given a set of boundary conditions.

Introduction
Motivation and Background
The stability of complex systems is central to the understanding of networks and com-
munities that have many interacting components. Examples of interactions include species
in ecology, human groups in sociology, markets in economics, computers in a cyber-physical
system, and neurons in our brains. The complexity of interactions between species within an
ecosystem and the question of dynamic stability is a topic of both theoretical and empirical
research in population ecology. In 1958, Charles Elton proposed the idea that simple com-
munities are less stable than diverse ones in that they are more vulnerable to invasions and
more likely to experience larger fluctuations as a result. There exists a complexity-stability
paradox where randomly generated ecological systems decreased in stability as they increased
in diversity in theory, but the opposite pattern has been observed in nature, leading to the
debate between diversity versus stability that has been studied by ecologists for the past 45
years. Recent studies are starting to explore dynamic interactions between the species to
more accurately depict the stability of realistic food web configurations and structures.
The relative state of a system is rated by three factors: richness, diversity, and complexity.
Richness refers to the total number of species in the system. Diversity is an index that
evaluates both species richness as well as the relative abundance of each species. Complexity
is a combined measure of the number of species and the strength of interaction and level of
connectivity between the species. To measure stability, the main factors to consider include

UC Berkeley, CE295
203

equilibrium, resistance, and resilience of the system. The system is considered to be stable
if it returns to equilibrium after encountering a perturbation. This can be a local as well as
global equilibrium point. The resistance of the system is the ability of it to resist change
when perturbed. A system with a higher level of resistance will be able to resist change
more easily than a system with a lower level of resistance. Lastly, the resilience of a system
measures the ability of the system to recover after being perturbed. A system with higher
resilience will return to equilibrium quicker than a lower resilience system.

Relevant Literature
The predator-prey model was first proposed by Alfred Lotka and Vito Volterra in the
1920s. This model has become the foundation of mathematical ecology and is widely used
to model predator-prey relationships. The Lolta-Volterra (LV) model is a pair first-order,
nonlinear differential equations that describe the dynamics of biological systems in which
two species interact, shown in (1, 2) [4].

ẋ = αx − βxy (1)
ẏ = −γy + δxy (2)
Where:
x is the number of prey
y is the number of predators
α is the growth rate constant for prey
β is the death rate constant for prey
δ is the growth rate constant for predators
γ is the death rate constant for predators

LV is a simple, deterministic model that relies on many assumptions, such as the following:
• Considers only two species.
• Prey have unlimited food so there is no interspecies competition.
• Predator’s food supply depends solely on the number of prey. As such, the predators
feed on prey and the food that the prey eat.
• The rate of change of population size is proportional to population size.
• Predators have unlimited appetite.
• The environment is held constant.
Since its proposal, researchers have altered the LV model to represent phenomena not pre-
viously considered, such as spatial heterogeneity, stochasticity, intra-species competition, and
the dynamics of two or more species. Interestingly, much of the recent literature continues
to rely on the basic LV model that is described in (1) and (2), without any augmentations.
In the paper by Gravel et. al (2015) [1], the authors study the trade-off between species
diversity, connectance, and interaction strength in determining the stability of a simulated

UC Berkeley, CE295
204

complex ecosystem. For each simulation, random interaction coefficients were applied before
solving for the equilibrium of the matrix. When only the positive species densities remained,
they subjected each matrix to a gradient of connectivity (termed ‘dispersal’ in the paper)
before numerically solving for the largest eigenvalue. They found that their meta-ecosystem
dynamics stabilized because of the effects of dispersal on the Jacobian matrix and its corre-
sponding eigenvalues. Cases with low dispersal are linearly affected, but in cases with high
dispersal, the effects are dependent on the ecological size of the metacommunity. Factors
such as species interaction and spatially structured landscapes (e.g., fragmentation and/or
habitat loss) can affect biodiversity in the ecosystem.
The stability of species coexistence is an important factor in maintaining the diversity
of an ecosystem. A paper by Chesson (2000)[7] evaluates how intra- and interspecies in-
teractions affect diversity maintenance, defined as the coexistence of species having similar
ecology in the same spatial region. Coexistence mechanisms can be defined in two main
ways: an equalizing mechanism, which minimizes average fitness differences between species,
or a stabilizing mechanism which increases competition between members of a single species
relative to competition between different species. Most studies focus exclusively on equal-
izing mechanisms, but Chesson argues that models of unstable coexistence would be more
robust if they also included mechanisms that model interactions between different species.
The paper by Chesson (2000) uses the LV model to discuss the basic principles of com-
petition. In terms of competition coefficients, the LV equations are written as (3):

1 dNi
× = ri (1 − αii Ni − αij Nj ) i = 1, 2 j 6= i (3)
Ni dt
Where αii and αij are absolute intraspecific and interspecific competition coefficients.

When parameterized in this method, intraspecific competition must be greater than in-
terspecific competition. While LV models growth rates are linear functions of density, models
with nonlinear growth rates can be written in the above form by making the competition
coefficients functions of density ( αij =fij Ni , Nj ) as long as the resident is at equilibrium and
the invader is at zero. A species is defined as an invader if it is at a low density compared
to the rest of the community.
A paper by Rafikov et al. (2005)[5] introduces two steps to solve the the optimal pest
control problem. The first step is to move the ecosystem pest at an equilibrium state below
the economic injury level by introducing the Pontryagin Maximum Principle. The Hamilton-
Jacobi-Bellman equation, based on the theory of dynamic programming, is used in this step
to obtain the pest control strategy through natural enemies’ introduction and stabilize the
pest density. Both of the steps are based on the predator-prey model.
For our implementation of optimal control problem, there is a complication with the
actual feasible value function. Sundström et al.[6] addresses this issue by introducing a
method for the optimal control of a one-dimensional dynamic model that avoids landing
in an infeasible solution space. In particular, after the preallocation of the Value Function
table, a new boundary based on the target number of species at time N is calculated by
proceeding backwards in time for k = N, N − 1, . . . , 0, which is xhigh (t) and xlow (t) in (1).

UC Berkeley, CE295
205

Figure 1: Value Function table with a new boundary

Focus of this Study

The goal of our research is to observe how ecological communities evolve over time subject
to chosen inputs and perturbations. We will be focusing on the effects of how the stability
of a system changes as a function of increasing complexity (e.g., network size, connectiv-
ity, and interaction strength). We will also delve into how ecosystems respond to external
management actions.

Technical Description
Our goal is to determine the optimal control strategy to achieve a desired populations
in a complex ecosystem. To achieve this, we will first develop a mathematical model to
represent the dynamical system. We will then optimally control this system to achieve a
target population size at a final time N with minimal intervention u.

Phase 1: Develop equations to describe the dynamical system.

Figure 2: Schematic diagram of a general three-species ecosystem

UC Berkeley, CE295
206

Population Relationships

We represent an ecosystem as a directed graph network with n nodes as shown in Figure

2. Each node represents a population of species xi , while each edge represents a relationship
between species xi and species xj . As such, this model is more general and complex than
a food chain, which has a radial nature. There are three types of inter-species relationships
that we will explore: predator-prey, competition, and commensal.
The parameters that describe the inter- and intraspecies relationships are organized into
a matrix R, as shown in (4) [2]. The diagonal entries rii represent the birth or death rates
for species xi . It is assumed that only the species at the bottom of the food chain has a
natural birth rate, while all other species have natural death rates.
 
r11 r12 . . . r1n
 r21 r22 . . . r2n 
R =  .. where rij ∈ R (4)
 
.. . . .. 
 . . . . 
rn1 rn2 . . . rnn

Interspecies relationships are characterized by the off-diagonal parameters rij . Positive

parameters (rij ) indicate that populations of xj are beneficial to xi , while negative parameters
(−rij ) indicate that populations of xj are adversarial to xi . For example, if we consider a
predator-prey relationship between two species x1 (e.g. elk) and x2 (e.g. wolves), r12 is
positive, while −r21 is negative. This is represented by the original Lotka-Volterra model,
represented in equations (5, 6). For equations (5-10), x1 and x2 must be positive real
numbers.

ẋ1 = x1 (r11 − r12 x2 ) (5)

ẋ2 = x2 (−r22 + r21 x1 ) (6)

Commensalism is a +/+ relationship where each species benefits from the other’s exis-
tence (e.g. birds and flowers). In this scenario, both r12 and r21 are positive. This phenomena
is represented with the equations (7, 8).

ẋ1 = x1 (r11 + r12 x2 ) (7)

ẋ2 = x2 (−r22 + r21 x1 ) (8)

Finally, competition is a -/- relationship where both species are negatively impacted by
the other’s existence. An example of competition would be two animals that eat the same
food or occupy the same physical space. In this scenario, both −r12 and −r21 are negative.
This phenomena is represented with the equations (9, 10).

ẋ1 = x1 (r11 − r12 x2 ) (9)

ẋ2 = x2 (−r22 − r21 x1 ) (10)

UC Berkeley, CE295
207

State-Space Equations

In a directed network, each species may exhibit one or more of the relationships de-
scribed above. As such, the population dynamics for each species xi in a system can be
represented using the general equation (11).
X
ẋi = (rij xi xj ) + rii xi (11)
i6=j

We describe the networked system using the generalized state-space equations for three
species in (12). The following two examples shows how (12) can be utilized to explain
different network types.

ẋ1 = x1 (r11 + r12 x2 + r13 x3 )

ẋ2 = x2 (r21 x1 + r22 + r23 x3 ) (12)
ẋ3 = x3 (r31 x3 + r32 x2 + r33 )
Next, we adapt the general state-space equations to reflect two types of networks: com-
mensal and food-chain.

Example 1: Three species commensal network

Figure 3: Commensal Network

Figure 3 is an example of a commensal relationship between a whale shark and remora, as

well as a predator-prey relationship, where the whale shark eats the krill. In this example,
the remora receives scraps of prey dropped by the shark and consumes parasites on the
shark. In return, it is protected from predators and receives a constant flow of water across
its gills. This relationship is represented by two positive relationship parameters, in green,
to represent that they both benefit from the relationship. We assume there is no interaction
between the remoras and krill. The dynamics of this network can be described by (13).

ẋ1 = x1 (r11 − r12 x2 )

ẋ2 = x2 (r21 x1 + r22 + r23 x3 ) (13)
ẋ3 = x3 (r32 + r33 )

UC Berkeley, CE295
208

Example 2: Three species food-chain network

Figure 4: Food Chain Network

The three species example in Figure 4 consists of coyote, a deer, and a blue grass popu-
lation. In a food chain, the top predator and bottom prey do not directly interact with each
other. The colored rij values in Figure 4 represent the intraspecies dynamics.

ẋ1 = x1 (r11 − r12 x2 )

ẋ2 = x2 (r21 x1 − r22 − r23 x3 ) (14)
ẋ3 = x3 (r32 x2 − r33 )

(14) is the set of state-space equations for the blue grass, deer, and coyote populations.
Unlike the commensal example where the remora and the shark have positive interspecies
relationships, parameter r23 is negative, indicating that the deer population decreases as the
coyote population increases. The remainder of this study will consider a food chain network
for either two or three species.

Phase 2: Evaluate Equilibrium and Stability.

Using the food chain scenario for three species, we can find where populations reach
a steady-state by identifying the equilibrium points. This is achieved by setting (14) to zero,
and solving for xeq eq eq
1 , x2 and x3 as seen in (15).

0 = xeq eq
1 (r11 − r12 x2 )
0 = xeq eq eq
2 (r21 x1 − r22 − r23 x3 ) (15)
0 = xeq eq
3 (r32 x2 − r33 )

UC Berkeley, CE295
209

There are two sets of equilibrium points, shown in (16) and (17), which are a function of
the parameters and are consistent with [3].

Set 1: xeq eq eq
1 = 0, x2 = 0, x3 = 0 (16)
r22 eq r11 eq
Set 2: xeq
1 = , x = , x =0 (17)
r21 2 r12 3

Next, we determine the stability of these equilibrium points by evaluating the Jacobian
(18) for the food chain model with respect to the equilibrium values. The system is unstable
if the eigenvalues have both negative and positive real parts, asymptotically stable if they
have all negative real parts, and marginally stable they have all non-positive real parts.
 
r11 − r12 x2 −r12 x1 0
J =  r21 x2 r21 x1 − r22 − r23 x3 −r23 x2  (18)
0 r32 x3 r32 x2 − r33

Substituting in the first set of equilibria points (16) into the Jacobian yields the matrix
(19) and eigenvalues (20). This is unstable because the first eigenvalue is positive, while the
second and third are negative.
 
r11 0 0
J =  0 −r22 0  (19)
0 0 −r33

λ(J) = (r11 , −r22 , −r33 ) (20)

Substituting in the second set of equilibria points (17) into the Jacobian yields the matrix
(21) and eigenvalues (22). The stability of the system is dependent on the third eigenvalue. In
particular, it is marginally stable when r32 r11 −r33 r12 ≤ 0 and unstable when r32 r11 −r33 r12 >
0.

r11 − r12r32r33 − r13r23r22

 
0 0
−r21 r33 −r23 r33 
J=  r32
0 r32 (21)
−r22 r31 −r22 r32
r23 r23
0
√ √ r32 r11 − r33 r12
λ(J) = i r11 r22 , -i r11 r22 , (22)
r12
The phase space diagrams in Figure 5 shows the population dynamics for two marginally
stable cases and one unstable case. In Cases A and B, the populations oscillate around the
equilibrium point due to its non-real eigenvalues. In Figure 5a, when r32 r11 − r33 r12 < 0, the
trajectory of the graph spirals downwards, changing from a 3D to a 2D system, depicting the
extinction of the coyote population. However, as shown in Figure 5b, when r32 r11 −r33 r12 = 0
(indicating a lower death rate for the coyote population), the system oscillates in a 3D system
without damping. Figure 5c is unstable (r32 r11 − r33 r12 > 0) and shows that the populations
spiral towards infinity as they deviate from the equilibrium. In all graphs, the different
colored orbits indicate different initial population sizes.

UC Berkeley, CE295
210

(a) Case A: (b) Case B:

r32 r11 − r33 r12 < 0 r32 r11 − r33 r12 = 0

(c) Case C:
r32 r11 − r33 r12 > 0

Figure 5: Phase space for a three-species food chain with marginally stable (Case A, B) or
unstable (Case C) equilibria.

As demonstrated in Figure 6, the farther the initial populations are from the equilibrium
point, the larger the oscillations and the greater the period. Figure 6a reflects populations
that initialize at the equilibrium, while 6b and 6c demonstrate non-equilibrium initial pop-
ulations x(0). This information also explains how the magnitude of perturbations from the
equilibrium impact the dynamics.

UC Berkeley, CE295
211

(a) x(0) = (12.5, 6.25, 0) (b) x(0) = (13, 6, 1)

(c) x(0) = (30, 10, 5)

Figure 6: Impact of initial population size on oscillatory dynamics for Case B. The more
the initialization deviates from the equilibrium, the more oscillatory the dynamics.

Phase 3: Minimize optimal controls to achieve a target population at the final

time step.

Optimization Problem - Two species system

Humans are not included as part of the ecosystem network, but can perturb the system
through a variety of controls (i.e., management techniques). For this problem, we consider
the possibility that predator x2 individuals can be removed through hunting or introduced
into the population at any time step. The removal of individuals is denoted −u, while the
introduction of individuals is denoted u. We aim to optimize the number of population
members added or removed to achieve a target population for two species at the final time
step N . The optimal control u∗ for ecosystem management is subjective and will vary based
on the specific scenario. For our purposes, we define the optimal control strategy as one
that minimizes the square value of the Euclidean norm. As such, the optimization problem

UC Berkeley, CE295
212

is formulated as follows:
N
X −1
min J = ||uk ||22 (23)
uk
k=0
With equality constraints:
xi0 = xiinit ∀i = 1, 2
x1k+1 = ∆t · x1k [r11 + r12 x2k ] + x1k ∀k = 0, 1, . . . , N − 1
x2k+1 = ∆t · x2k [r21 x1k + r22 ] + x2k + ∆t · uk ∀k = 0, 1, . . . , N − 1
And inequality constraints:
umin ≤ uk ≤ umax ∀k = 0, 1, . . . , N − 1
xmin ≤ xk ≤ xmax ∀k = 0, 1, . . . , N
x1,min
N ≤ x1N ≤ x1,max
N

x2,min
N ≤ x2N ≤ x2,max
N

Where:
k is the discrete time index
N is the time horizon
∆t is the size of each timestep
xk = [x1k , x2k ] is the state at time k
xik is the state for species i at time k, i ∈ 1, 2
uk is the control decision applied at time k
xiinit is the state for species i at initial time k = 0 , i ∈ 1, 2
xmin , xmax are the minimum and maximum population for every species for all time k
umin , umax are the minimum and maximum control value for species 2 for all time k
xi,min
N , xi,max
N are the minimum and maximum population for species i at time N, i ∈ 1, 2

Dynamic Programming

Since the dynamics of the system evolve with time, Dynamic Programming is used to
break the multistage decision problem into subproblems. The value function for this problem
is defined as follows:

Let Vk (xk ) denote the minimum control of species x2k from time step k to N, given that
the current state is the population sizes for each species xk = [x1k , x2k ].

The Principle of Optimality (PoO) can be written in recursive form as:

Vk (xk ) = min{||uk ||22 + Vk+1 (xk+1 )} (24)
uk

With the boundary conditions:

(
0 if x1,min
N ≤ x1N ≤ x1,max
N , x2,min
N ≤ x2N ≤ x2,max
N
VN (xN ) = (25)
inf otherwise

UC Berkeley, CE295
213

The optimal control action is:

u∗ (xk ) = arg min{||uk ||22 + Vk+1 (xk+1 )} (26)

uk

This formulation ensures that at every timestep in the time horizon k to N , the control
input uk and the corresponding states xk are optimized. The dynamic program calculates
3D value table, with the dimensions time k, the state of species x1 and the state of species
x2 . To limit the computational effort, the value table is discretized into a grid. Therefore,
the recursive PoO (24) involves computing xk+1 , and interpolating of Vk+1 (xk+1 ) onto the
grid.

1. Computing xk+1
For each timestep, all possible controls uk are evaluated for each possible state xk . To
simplify the formulation of ugrid , we rearrange each equality constraints as an upper
and lower limit on uk .

xmin ≤ x2k+1 = ∆t · uk + ∆t· (27)

xmin ≤ x2k+1 = ∆t · uk + ∆t· (28)
x2k [r21 x1k + r22 ] + x2k ≤ xmax (29)
xmin ≤ ∆t · uk + ∆t· (30)
x2k [r21 x1k + r22 ] + x2k ≤ xmax (31)
xmin − x2k − ∆t · x2k [r21 x1k + r22 ] ≤ ∆t · uk ≤ xmax − x2k − ∆t · x2k [r21 x1k + r22 ] (32)
xmin − x2k xmax − x2k
− x2k [r21 x1k + r22 ] ≤ uk ≤ − x2k [r21 x1k + r22 ] (33)
∆t ∆t
2. Interpolation of Vk+1 (xk+1 )
For each xk+1 that is not on a grid-point (i.e., is not an integer), a 2D interpolation is
required to calculate Vk+1 (xk+1 ).

Boundary Conditions Towards the Final Target Size

Inspired by [6], the boundaries for our model can be calculated using the following algo-
rithm:

1. Initialize with the lower bound of the partially constrained final state. x1k,low = x1N,min
and x2k,low = x2N,min

2. Proceed backward in time for k = N − 1, . . . , 0:

(a) Assume the states are unconstrained and solve the fixed point problem as shown:
i. Initialization:
x1,st=0 1 2,st=0 2
k,low = xk+1,low xk,low = xk+1,low (34)

UC Berkeley, CE295
214

ii. Iterate over st until a specified tolerance is achieved:

x1,st+1 1 1,st 2,st

k,low = xk+1,low − ∆t ∗ xk,low ∗ (R11 + R12 ∗ xk,low )

x2,st+1 2 2,st 1,st

k,low = xk+1,low − max{∆t ∗ (xk,low ∗ (R22 + R21 ∗ xk,low ) + uk )}
uk

(b) Check whether the solution violates the state constraints.

(c) If the constraints are violated, solve the general problem:

x1k,low = xmin (35)

max{∆t ∗ (x2,st 1,st

k,low ∗ (R22 + R21 ∗ xk,low ) + uk )}
uk

s.t. {∆t ∗ (x2,st 1,st 2,st 2,st

k,low ∗ (R22 + R21 ∗ xk,low ) + uk )} + xk,low = xk+1,low
umin ≤ uk ≤ umax
xmin ≤ xi,st
k,low ≤ xmax ∀i = 1, 2

(d) Store the solution xk,low with the respective uk,low and the cost-to-go Vk,low .

3. The upper bound of states x1 and x2 can be found using similar concepts as shown in
the previous steps.

To implement our boundary conditions in dynamic programming, it is necessary to check

if (36) fits, in which xik,next is all possible next states calculated from system dynamics.
Next, we only consider the feasible states from xik,next that will optimize the problem. If the
chosen state is on the boundary line, both the u∗ and Vk (xk ) will need to take the value from
uk,low/high and Vk,low/high .
xik,low ≤ xik,next ≤ xik,hi ∀i = 1, 2 (36)
Results

The parameters in our dynamic programming problem is as listed: N = 200, ∆t=0.1,

xmin =0, xmax =200, umin =-30, umax =30, xinit =[10,5].
The system dynamics without optimal control is shown in Figure 7. The Figure 8 is the
result for 38 ≤ x1N ≤ 42, 13 ≤ x2N ≤ 17. The simulation stops at x∗1 ∗2
N = 39 and xN = 14. The
Figure 9 is the result for 18 ≤ x1N ≤ 22, 18 ≤ x2N ≤ 22. The simulation stops at x∗1
N = 19
∗2
and xN = 23.

UC Berkeley, CE295
215

Figure 7: Original System Dynamics Without Control

Figure 8: Simulation Result for 38 ≤ x1N ≤ 42, 13 ≤ x2N ≤ 17

Figure 9: Simulation Result for 18 ≤ x1N ≤ 22, 18 ≤ x2N ≤ 22

UC Berkeley, CE295
216

Discussion
As shown in Figure 7, the two species modeled without the controls follow a stable tra-
jectory, with the prey species x1 peaking at around 42 members and the predator species x2
peaking at around 25 members at the final time step. However, in using optimal control,
the model aims to have a prey species between 38 to 42 members, and a predator popu-
lation between 13 to 17 members. Figure 8 displays the simulation result for the dynamic
programming model with optimal control; the control was applied to species x2 , in this case
the predator species. The optimal control scenario involves removing predator members (the
orange line in the population size graph) as the predator population peaks in the oscillatory
period. The dashed blue and red lines represent the solution space upper and lower bound-
aries for species x1 and x2 , respectively. The upper boundary line is the highest state xi, high
at each of the time steps that allows us to obtain the final state (and vice versa for the lower
boundary line) [6]. Figure 9 is a variation on the first optimal control model; in this scenario,
both predator and prey populations must be between 18-22 members at the final time step.
Similarly to the first optimal control problem, members of species x2 are removed after each
of the high points in the period.

Summary
One of the main takeaways from conducting a literature review and from our own experi-
mentation with stability and equilibrium analysis is that populations change in an oscillatory
pattern. A stable population, as modeled using the Lotka-Volterra model, will exhibit peri-
odic behavior with bounded trajectories [8]. In addition, the equilibrium trajectories and the
controllable inputs are highly sensitive to inter- and intra-species relationship parameters.
Slight variations in population growth rates and species interactions could cause destabi-
lization of the model. One limitation to our work is that the relationship parameters are
not well understood, which is reflected in the literature. We also found that modeling with
dynamic programming is exponentially more complicated to model as the number of species
modeled increases. The Lotka-Volterra model is not easily transferred to models with multi-
ple species. For example, a food chain with a odd number of species must include a logistic
term, while models of an even number should not contain a logistic term [8].

Acknowledgments
We would like to extend our gratitude to Professor Scott Moura and Bertrand Travacca
for their help throughout this process.

About the authors

Kyra Chang is a first-year Systems graduate student in Civil and Environmental Engi-
neering at the University of California, Berkeley.

UC Berkeley, CE295
217

Tiffany Chang is a first-year Energy, Civil Infrastructure, and Climate graduate student
in Civil and Environmental Engineering at the University of California, Berkeley.

Emily Farrar is a first-year Energy, Civil Infrastructure, and Climate graduate student in
Civil and Environmental Engineering at the University of California, Berkeley.

Nate Tsang is a first-year Energy, Civil Infrastructure, and Climate graduate student in
Civil and Environmental Engineering at the University of California, Berkeley.

References
[1] D. Gravel, F. Massol, and M. Leibold. “Stability and complexity in model meta-ecosystems”.
In: Nature Communications 7.12457 (2016). doi: 10.1038/ncomms12457.
[2] L. Devireddy. “Extending the Lotka-Volterra Equation”. In: (2016), pp. 2–10.
[3] E. Chauvet, J. Paullet, J. Previte, and Z. Walls. “A Lotka-Volterra three-species food
chain”. In: Mathematics Magazine 75.4 (2002), pp. 243–255. doi: 10.2307/3219158.
[4] A. Lotka. “Analytical Note on Certain Rhythmic Relations in Organic Systems”. In:
Proceedings of the National Academy of Sciences of the United States of America 6.7
(1920), pp. 410–415. doi: 10.1073/pnas.6.7.410.
[5] M. Rafikov and J. Balthazar. “Optimal Pest control problem in population dynamics”.
In: Computational Applied Mathematics 24.1 (2005), pp. 65–81. doi: 10.1590/S0101-
82052005000100004.
[6] O. Sundstrom, D. Ambuhl, and L. Guzzella. “On Implementation of Dynamic Program-
ming for Optimal Control Problems with Final State Constraints”. In: Oil Gas Science
and Technology - Rev. IFP 65.1 (2010), pp. 91–102. doi: 10.2516/ogst/2009020.
[7] P. Chesson. “Mechanisms of maintenance of species diversity”. In: Annual review of
Ecology and Systematics 31.1 (2000), pp. 343–366. doi: 10.1146/annurev.ecolsys.
31.1.343.
[8] S. Harris, J. Paullet, J. Previte, and J. Ranola. “A Lotka-Volterra Four Species Food
Chain”. In: Mathematics Subject Classification 92.01 (2005), pp. 1–10.

UC Berkeley, CE295
218

UC Berkeley, CE295
219

Blood Glucose Prediction

Mallika Bariya, Mohini Bariya

Abstract
In this project we develop a model for predicting blood glucose dynamics using auto-
regressive handling of the input variables - dietary glucose intake and physical activity -
and measured output (blood glucose levels as monitored by a continuous glucose meter).
We build a training and test data set by carefully recording time-series data for the input
variables and measuring blood glucose for an individual over the span of a week. The model
we develop is a starting point for predicting blood glucose levels in pre-diabetic patients,
potentially informing preventive therapies to stave off the onset of diabetes. Our predictive
model could also be incorporated into insulin injection control loops for diabetic patients.

Introduction
Motivation and Background
Diabetes is a chronic metabolic disease in which the body is unable to support normal
mechanisms for regulating blood glucose (BG) levels, either because it cannot endogenously
produce the insulin needed to store and absorb sugars, or because it develops insulin re-
sistance. As a result, sugars from food accumulate in the blood without a mechanism for
removal by absorption into tissues and cells, causing BG levels to become dangerously ele-
vated after eating. Diabetes can also result in sudden precipitous declines in blood glucose,
and is in general marked by high BG uctuations that can be life-threatening. As one of
the most abundant chronic diseases, with over 30 million Americans aicted and 1.5 million
more being diagnosed each year [1], there is an urgent need to understand how blood glucose
levels are impacted by daily activities and develop models for predicting blood glucose with
the vision of enabling preventive therapies. While much eort has historically been dedicated
to understanding the complex interplay of glucose and insulin in diabetic patients, it is also
valuable to model glucose dynamics in healthy individuals with functioning glucose regu-
lation systems. Non-diabetics do not rely on insulin injections or medication to stabilize
changing glucose levels as endogenous insulin production is metered internally by a complex
control system involving the pancreas, liver, and other organs. A similar system is at work
for pre-diabetics who, though able to generate insulin, show developing insulin resistance
that diminishes the eect of the internal control system. A predictive model for blood glucose
levels in healthy individuals could thus be extrapolated to pre-diabetics with minimal re-
structuring, and could be used to inform lifestyle changes and local preventive care to stave

UC Berkeley, CE295
220

o diabetes. Almost 85 million Americans are pre-diabetics [1], but with a healthcare system
that is focused on reactive as opposed to preventive treatment, there are few technological or
medical solutions to aid in monitoring developing health states. A predictive blood glucose
model that can capture the diminished response of the endogenous glucose control system
would thus be hugely impactful for curtailing the rising number of diabetes diagnoses.

Focus of this Study

The focus of this study is to build a predictive model for the blood glucose evolution of
a healthy, non-diabetic individual to explore the impact of influencing factors like physical
activity and food intake. This model will be based on longitudinal data from a single subject
to capture personalized model parameters.

Literature review
Many methods have been demonstrated for blood glucose prediction. Some studies,
such as by Lehmann et al, favor a first-principles approach, using compartment models
with simple differential equations to capture the interplay of food intake, physical activity,
insulin and blood glucose [2]. Others use time-series methodologies, relying on historic
measurements to predict future ones; for example, Eren-Oruklu et al use recursive linear
models to estimate parameter values that can change to accommodate most recent data and
any new disturbances [3]. Other studies use data-driven approaches and place less importance
on the dynamics of individual variables. Sandham et al use 1 artificial neural networks on
vast quantities of mined data to estimate weighting parameters for the different factors in
uencing blood glucose [4]. Naumova et al use meta-learning, creating algorithms that learn
which parameters are important before learning their actual values to improve portability
of the model across patients and even applications [5]. Using simple compartment models is
effective when the body produces no insulin and has no internal glucose control system. Then,
injected insulin acts as a glucose regulator and is a measurable input, allowing key compo-
nents of the governing differential equations to be known. In contrast, for healthy individuals
it is impossible to directly measure the actions of the body’s internal control mechanisms,
complicating the use of simple differential equations. Using machine learning tools like neural
networks poses challenges as well, requiring vast quantities of data to prevent overfitting while
also sacrificing model explainability and the ease with which we can slightly alter the model
structure to achieve better prediction-to-measurement coherence. As we are working with
limited data, we instead choose a simple model structure built on an intuitive idea of how
a glucose control system might work, while finding least squares estimates for parameter
values. Our linear model is simple to apply and understand, allowing an instructive and
potentially useful attempt at blood glucose prediction in healthy individuals.

Key contributions
While various complex physiological and data-driven models have been proposed for blood
glucose prediction, none achieve very successful results. We seek to create a simple, intuitive
model based on known influencing factors to determine each of their impact on blood glucose

UC Berkeley, CE295
221

evolution. While we cannot hope to achieve very accurate predictions with this approach,
our contribution is mainly towards understanding how food intake and activity individually
impact blood glucoes.

1 Data Collection
In our simplified model, we assume that blood glucose levels can be accurately predicted
based on food intake and activity. We collect food intake, activity, and blood glucose mea-
surements over 6 days in March. These measurements comprise the time series dataset on
which our models are trained and tested. In the following section, we describe how each
type of measurement is made, and how the different measurements are converted to time
series on the same discrete time grid (i.e. with identical sampling periods). In addition, for
each data source, we describe potential sources of error in the measurement process and data
pre-processing that could contribute to error in our final blood glucose predictions.

The sampling period of all the time series, denoted dT , was chosen to be 5 minutes, as
this was the resolution of the blood glucose data. For diabetic applications, a 5 minute
prediction time-step is reasonable for the purposes of glucose prediction as we would expect
the blood glucose to remain reasonably stable over this time period. However, if this is not
the case, a a shorter time step is desired, our models could be trained on higher resolution
data, or our 5 minute resolution predictions could be interpolated to a finer time scale. For
reference, with dT = 5, 6 days of data contain a total of 1728 data points per measurement
time series.

1.1 Activity Data

Figure 1: The Google Fit app estimates various activity metrics based on smart phone sensor
measurements. Estimates are streamed to a web interface.

We use Google Fit, a cellphone application, to record activity [6]. Google Fit estimates
steps walked and calories burned using GPS, gyroscope, and accelerometer measurements
from a smart phone. For our project, we choose to use Google Fit’s estimates of calories
burned, rather than steps walked, as representative of the subject’s activity. We believe
calories burned should be more directly linked to the removal of glucose from the blood as

UC Berkeley, CE295
222

sugars are used to fuel cell activity. Since the subject did not do any exercise during the test
period, and especially no stationary exercise such as on an exercise bike or treadmill, the
activity data predominantly tracks walking. Figure shows the Google fit interface. Google
fit records estimates every 1 hour, producing a time series ch [k]. We interpolate these onto
a dT = 5 min. time grid by assuming that calories are burned at a fixed rate every hour.
This produces the time series c[k] which is related to ch [k] as follows:
 
1 k
c[k] = ch
12 12

Potential errors introduced by the activity data would largely consist of inaccurate estimates
of calories burned. Google Fit uses its own algorithm to convert sensor measurements to
calories burned, taking into account the self-reported height and weight of the user. These
algorithms are unlikely to be perfect. However, we did ballpark checks that confirm the
caloric estimates are reasonable.

1.2 Food Data

Figure 2: An example of converting qualitative food data to a quantitative time series for
one banana.

In our simplified model, food is the input source of blood glucose. To create a time
series of food data, we start by recording the the type and quantity of every item of food
consumed, as well as the time period over which it was consumed. Some sample entries from
our records are included in Table 1.

Table 1: Sample food intake logs

Day Time Item

3/15/18 3:45pm-3:48pm half a doughnut with chocolate glaze
3/15/18 5:45pm-5:48pm 1 banana

The qualitative food data must be converted into a quantitative time series. The impact
of a certain food on blood glucose levels is related to two important properties of the food
item. These are the caloric content and the glycemic load [7]. The significance of the caloric
content is intuitive as it indicates how much sugar is contained in the food, and therefore

UC Berkeley, CE295
223

should be directly related to the quantity of sugar entering the blood stream. The glycemic
load reflects how quickly the sugars from a certain food can be released into the blood. This
depends on such aspects as the structure of the food, the amount of fiber it contains, etc [8].
Two foods may have the same caloric content, but if one has a larger glycemic load, it will
lead to a more rapid increase in blood sugar. Since they depend on such a variety of factors,
glycemic load values are determined through experimentation, with white bread setting the
standard with a glycemic load of 100.

We use both caloric content and glycemic load to generate our quantitative food intake
time series. To begin, we convert the quantitative food data into a time series of rect()
functions. Each rect() corresponds to one food item. The width of the rect() is equal to
the duration over which the food was consumed, while the height is chosen so that the area
under the rect() is equal to the caloric content of the food item. To account for the glycemic
load, we convolve each rect() with a Gaussian, where the width of the Gaussian is related
to the glycemic load. An example for one banana is shown in Figure 2.

1.3 Blood Glucose Data

Figure 3: The Dexcom Blood Glucose Sensor and online measurement reporting interface.
Note the fine needle attached to the sensor that penetrates the skin.

To measure blood glucose values over our test period, we use Dexcom, a continuous glu-
cose monitoring [9][10]. As shown in Fig. 3, the Dexcom device consists of a small data
conversion and transmission unit that sits above the skin, and a fine sensing needle that
penetrates the skin to access interstitial fluid (IF). This sensor makes periodic IF glucose
measurements every 5 minutes. IF glucose levels correlate one-to-one with blood, and time
delays between the two fluid compartments are compensated for in Dexcom’s signal process-
ing. Overall, the Dexcom sensor effectively measures blood glucose levels and the generated
data can be collected into a time series as shown in the right plot of Fig. 3. In this plot,
the black line shows the Dexcom sensor reading; the red and yellow lines depict cut-offs for
dangerously high or low glucose levels. The grey circles labeled with ’C’ set the times at
which the Dexcom sensor was re-calibrated via comparison with a fingerstick blood glucose
test. Calibration is necessary to compensate for drifting in the Dexcom signal that could
lead to underestimation or overestimation of blood glucose values. If there is a large discrep-
ancy between the Dexcom reading and the fingerstick control measurement at the calibration

UC Berkeley, CE295
224

point, the sensor is re-set to match the control. In our data, our sensor did not have much
drift and thus calibration points did not create significant discontinuities in the time series
(apart from day 1, when the sensor is first inserted and shows particularly noisy readings as
it equilibriates with its new in vivo surroundsings). This is important as it ensures that the
blood glucose data generated by Dexcom is accurate throughout the measurement interval.

2 Predictive Models
We train and test several different models for blood glucose prediction. We develop
the models incrementally to better understand how strongly each input variable impacts
blood glucose levels. In the following we begin by detailing the performance metrics used
to evaluate our models. We then describe each model and detail the results of training and
testing each model.

2.1 Performance Metrics

• Mean Squared Error. The mean squared error (MSE) is the classic metric for quanti-
fying the quality of prediction. Given a prediction time series ŷ[n] and the true time
series y[n], both of length N , the MSE is given by:
N
1 X
M SE(ŷ) = (ŷ[n] − y[n])2
N n=1
However, for the purposes of blood glucose prediction, the MSE can be difficult to
interpret. Given our highly simplified system in which only food and activity affect
glucose levels, if we are able to reliably predict when glucose rises and falls, the model
should be considered quite successful. This motivates the use of another metric: the
correlation coefficient.
• Correlation Coefficient. The correlation coefficient measures how correlated two ran-
dom variables are, in other words how closely they rise and fall together. For our time
series, ŷ[n] and y[n], the correlation coefficient is computed as follows:

PN
− µŷ )(y[n] − µy )
n=1 (ŷ[n]
ρ(ŷ) = s  
PN 2
PN 2
n=1 (ŷ[n] − µŷ ) n=1 (y[n] − µy )

N N
1 X 1 X
µy = y[n]; µŷ = ŷ[n]
N n=1 N n=1

The correlation coefficient is a useful performance metric for this project because it
indicates how well a model is doing at predicting the rises and drops in blood glucose.
Given our highly simplified system, a strong correlation between the true and forecasted
blood glucose levels would represent a success. As a rough, rule-of-thumb, ρ(ŷ) ≥ 0.7
would indicate that ŷ is a good prediction.

UC Berkeley, CE295
225

2.2 Autoregressive Model

The autoregressive (AR) model is a simple linear model which uses current and past
blood glucose measurements to predict future blood glucose levels. Given blood glucose
measurements g[n], the model takes the following form:
lg
X
ĝ[n + k] = αi g[n − i]
i=0

ĝ[n + k] is our k step ahead prediction, lg is the model order, or the number of past measure-
ment points we use to make our prediction, and αi , i = 0, ..., lg are the model coefficients.
We choose lg = 12, corresponding to the use of 1 hour of previous measurements. We fit the
model by solving the following ordinary least squares optimization problem.

α∗ = arg min ||Gn+k − Gn α||22

α
α∗ = (GTn Gn )−1 GTn Gn+k

where Gn and Gn+k are data matrices of glucose measurements constructed appropriately
from the g[n] time series. α is a vector of the model coefficients.

For k = 1, the model performs extremely well. The results are shown in Figure 4. This
tells us that in a healthy subject, the blood glucose evolves relatively smoothly. Therefore,
if the current and a few past values of the blood glucose are known, the blood glucose 5
minutes (k=1) into the future can be predicted reliably.

Figure 4: AR model results on training data (left) and test data (right)

As the prediction horizon (ie k) increases, the prediction quality falls steadily. As shown
in Figure 5, at some point, there is a leveling off in the MSE (and correlation), indicating the
time horizon at which the current blood glucose values are essentially useless for predicting
the k-step ahead future blood glucose values.

UC Berkeley, CE295
226

Figure 5: MSE (left) and Correlation Coefficient (right) versus the prediction horizon.

The AR model is accurate but not useful. Indeed, we could say that this model makes
sense from a pure data perspective but not from a physiological perspective. It does not
account for the external factors, such as food consumption and activity, that we know impact
blood glucose. Such a model would be even less useful for diabetic subjects, whose blood
glucose level is less regulated by the body and more prone to sudden, large changes due to
behaviour (such as eating or running).

2.3 Affine Food Intake Based Glucose Prediction

To focus on the effect of external factors (food consumption and activity) on future
glucose values, we train a model to predict one step (5 minutes) ahead glucose levels from
food intake data only. Given food intake values f [n], the model takes the following form:
lf
X
ĝ[n + 1] = γi f [n − i] + b
i=0

We allow for a bias term b as the level of the glucose measurements g[n] and food intake
time series f [n] differs. We fix lf = 24, corresponding to two hours of previous food intake
measurements. Again, we fit the model by solving the following OLS optimization problem.

γ ∗ = arg min ||Gn+1 − Fn γ||22

γ

γ = (FnT Fn )−1 FnT Gn+1

∗

Food intake is sporadic during the day. In fact, in our last two days of test data, the subject
only ate once. During the other time periods, f [n] = 0, and apart from a constant b, the
given model has no ability to predict the blood glucose level. Therefore, for the purposes
of this model, we restrict our data set to regions during which there is food consumption.
These time periods are shown in Figure 6. The blue periods correspond to training data,
while the yellow correspond to test data.

UC Berkeley, CE295
227

Figure 6: Data periods used for training (blue) and testing (yellow) the affine food intake
glucose prediction model.

The results of the food intake model are shown in Figure 7. We can see that there is a
large range in the test MSE/correlation. The correlation (which is a more intuitive metric
here) ranges from −0.01 to 0.63.

Figure 7: Test results for affine food intake prediction model

Glycemic load errors would severely deteriorate the performance of this model. This

UC Berkeley, CE295
228

might be the reason we see certain periods where the blood glucose predictions and mea-
surements seem to be anti -correlated. If the glycemic load is inaccurate, the blood glucose
rise due to food consumption could occur later than the model predicts. This is visually
expressed in Figure 8.

Figure 8: Potential effect of glycemic load errors on model predictions: Glucose rises are
predicted to occur before they actually do, as indicated by the dark arrows.

2.4 Affine Food Intake and Activity Based Glucose Prediction

We extend the affine food intake model to incorporate activity. We again train a one
step ahead glucose predictor, now using both food and activity time series data. Given food
intake time series f [n] and activity (estimated calories burned) time series a[n], the model
takes the following form:
lf la
X X
ĝ[n + 1] = γi f [n − i] + βi a[n − i] + b
i=0 i=0

We fit the model by solving the OLS optimization, fixing lf = 24, la = 24.
 ∗
 γ  2
   
γ  
  Gn+1 − Fn An
= arg min  
β γ β 
2
 
β

The results for the model are shown in Figure 9. Comparing with 8, we can see that the test
performance actually deteriorates compared to the affine food intake.

UC Berkeley, CE295
229

Figure 9: Training (left) and test (right) results for the affine food intake and activity based
glucose prediction.

To understand this performance deterioration, we consider the fitted coefficients, shown

in Figure 10. The β values, corresponding to the activity time series, do not make physical
sense. The magnitude of the coefficients increases with i, and the sign of the coefficients
oscillates. This is symptomatic of overfitting.

Figure 10: The OLS fitted coefficients for the affine food intake and activity based glucose
prediction model. β (left) corresponds to the activity time series while γ (right) corresponds
to the food time series.

To mitigate overfitting, we add regularization, specifically penalizing the γ coefficient

magnitudes only. This is a generalized Ridge regression optimization, which can be expressed
as:
 ∗
 γ 2 
     
γ    γ
  Gn+1 − Fn An
= arg min + γ β Σ
   
β γ β 2  β
 
β

UC Berkeley, CE295
230

To penalize the γ coefficients only, we set the

Σ matrix as follows:
 
λ 0 ... 0 0
0 λ . . . 0 0
Σ =  .. ..
 
. . . . . .. .. 
. .
0 0 ... 0 0

Σ contains all zeros except in the diagonal elements corresponding to the γ coefficients. Note
we penalize all γi equally with weight λ. The results for λ = 3000 are shown in Figures 11-
12. We can see that the test MSE reduces with regularization. The β coefficients are all
negative, which makes more physical sense since activity corresponds to removal of blood
glucose. However, the MSE and correlation are worse than for the purely food intake based
model, suggesting that either activity does not effect blood glucose in a linear and predictable
way, or our training data is insufficient to establish the relationship.

Figure 11: The regularized fitted β coefficients for the affine food intake and activity based
glucose prediction model. The regularization parameter was λ = 3000

Figure 12: Test results for the affine food intake and activity based glucose prediction with
regularization.

UC Berkeley, CE295
231

2.5 Neural Network Glucose Prediction from Food and Activity

The body’s regulation of blood glucose through its internal control mechanisms is likely
to be highly nonlinear. This limits the success of linear models like those we tested in the
previous sections. Our final model is a neural network with ReLu activations and one hidden
layer with 4 neurons. A diagram of the network is shown in Figure 13. The neural network
aims to predict the blood glucose from the food and activity time series.

Figure 13: Diagram of the neural network model used for blood glucose prediction from food
and activity data. Courtesy of deeplearning4j.org

The results of the model are shown in Figure 14. We can see that the neural network
manages to achieve a better correlation than the affine food intake model on the training
data. The performance on the test data is only slightly better than the affine food model.
It is interesting that both the neural network and linear model achieve a similar prediction
on the test data. This suggests that the poor prediction is not due to model issues, but
due to changes in the relationship between blood glucose levels and activity and food intake
between the training and test periods.

Figure 14: Training (left) and test (right) results for the Neural Network predictor.

Conclusion
In this project, we developed, trained, and tested several models for predicting blood
glucose from food intake and activity data. Our success in predicting blood glucose was

UC Berkeley, CE295
232

limited by a paucity of training and testing data and the simplicity of our models. We made
the assumption that blood glucose was predominantly affected by food intake and activity.
However, in a healthy, non-diabetic subject, the body has a complex control system at work
which regulates blood glucose levels. This system produces changes in blood glucose that
are difficult to predict from food and activity alone.
We found that in a healthy subject, an AR model can be used to predict future blood
glucose levels from current and past measurements, without the use of food intake or activity
measurements. This result is useful in a control system framework where continuous glucose
measurements are used to determine insulin delivery.
We also found that in our dataset, blood glucose prediction from food intake alone per-
formed better than incorporating both food intake and activity. This suggests that food
intake has a more direct impact on blood glucose, while activity has more complex effects
that are not easy to learn from our limited dataset.
Finally, we found that simple linear models performed nearly as well as nonlinear neural
networks on our limited dataset.

References
[1] “Diabetes statistics,” http://www.diabetes.org/diabetes-basics/statistics/.

[2] E. D. Lehmann and T. Deutsch, “A physiological model of glucose-insulin interaction in

type 1 diabetes mellitus,” Journal of biomedical engineering, vol. 14, no. 3, pp. 235–242,
1992.

[3] M. e. a. Eren-Oruklu, “Estimation of future glucose concentrations with subject-specific

recursive linear models,” Diabetes technology therapeutics, vol. 11, no. 4, pp. 243–253,
2009.

[4] W. A. e. a. Sandham, “Neural network and neuro-fuzzy systems for improving diabetes
therapy,” Engineering in Medicine and Biology Society. Proceedings of the 20th Annual
International Conference of the IEEE, vol. 3, 1998.

[5] S. V. P. Naumova, Valeriya and S. Sivananthan, “A meta-learning approach to the

regularized learning—case study: blood glucose prediction,” Neural Networks, vol. 33,
pp. 181–193, 2012.

[6] “Google fit,” https://www.google.com/fit/.

[7] “Glycemic index,” https://en.wikipedia.org/wiki/Glycemic index.

[8] “United states department of agriculture agricultural research service usda food com-
position databases,” https://ndb.nal.usda.gov/ndb/search/list.

[9] “Dexcom continuous glucose monitor,” https://www.dexcom.com/.

[10] https://mspoweruser.com/fitbits-ionic-smartwatch-getting-support-continuous-blood-
glucose-monitoring/.

UC Berkeley, CE295
233

About the authors

Mallika Bariya is a PhD student in Material Science at Berkeley. Her research is focused
on developing noninvasive sensors for personalized health monitoring.

Mohini Bariya is a PhD student in Electrical Engineering at Berkeley. Her research is on

using sensor data to improve situational awareness and enable integration of renwables
in the distribution grid.

UC Berkeley, CE295
234

UC Berkeley, CE295
235

Optimization of Rainwater Harvesting Systems for Single-Family

Residential Buildings in California
Ahmad Bin Thaneya, Fiona Greer, Minghui Zhang

I. Abstract: It is important to consider water used in buildings in California due to concerns of water
scarcity, as well as a growing understanding of the water-energy nexus and its relationship with the
climate. Rainwater harvesting systems (RHS) are one example of onsite water reuse systems proposed for
meeting non-potable water demands in buildings. There exist several questions regarding the efficacy of
RHS in meeting energy, greenhouse gas, and cost objectives. This research project will compare the use
of RHS in residential buildings with two California cities, Oakland and San Diego, and simulate RHS
performance in meeting established non-potable water needs through constraints of rainwater supply,
transmission energy, greenhouse gas emissions, and life-cycle costs. A multiobjective optimization
problem was solved in each city to determine the optimal operation of RHS in order to minimize cost and
energy use, and maximize the portion of demand that is supplied by rainwater. Through dynamic
programming and economic cost-benefit analysis, we found that RHS in California are not justifiable in
their current cost and emissions considerations unless a RHS subsidy is provided or costs of water savings
and emissions are adjusted.

II. Introduction
a. Motivation and Background:
Water is a vital resource used in buildings. It is important to consider alternative sources of water
in order to meet greenhouse gas reduction goals and to sustain society’s water demand. Total urban water
usage in California averages around 9.1 million acre-feet a year, with 4.2 million acre-feet used in urban
outdoor applications (Pacific Institute, 2014). Residential use accounts for 64% of the total urban water
use, while commercial use accounts for 23% (Pacific Institute, 2014). Onsite water reuse systems, such as
rainwater harvesting systems (RHS), have gained popularity in recent years as a strategy that could
address concerns about water scarcity and GHG emissions associated with the treatment, transport, and
heating of water. RHS, which capture, store, and use collected water from precipitation events, are a
water conservation technology that have the potential to address a building’s water needs, such as its
toilet and urinal flushing requirements.
The impetus to study RHS is bolstered by the fact that two of the members in this group are
currently working on a project for the California Air Resources Board (CARB) that studies the feasibility
of net zero carbon buildings. One of the net zero carbon building strategies being analyzed in the CARB
project is onsite RHSs. There are limitations in the CARB project that prevent detailed analysis of
whether RHS are a cost-effective and energy saving strategy. This project aims to address whether RHS
are a practical strategy.

b. Relevant Literature:
Recent interest in RHS adoption stems from the benefits that these systems can provide building
owners such as: (1) limiting costs to only system installation and operation; (2) allowing for water
sourcing without heavy reliance on complex distribution systems; (3) augmenting or replacing utility
sourced water for non-potable water needs, including toilet flushing; and (4) reducing customer utility
bills (Texas Water Development Board, 2005). Researchers and policy makers are also interested in the
potential societal benefits of RHS. These benefits include: (1) mitigating rising water demand through
decentralized sourcing; (2) conserving of water and energy; (3) reducing stormwater runoff and
associated pollutant loads; (4) replacing depleted groundwater supplies; and (5) aiding electricity utilities
by reducing summer demand peaks (EPA, 2015). The practicality of RHS adoption is contingent on

UC Berkeley, CE295
236

rainfall characteristics and RHS storage capacity (Texas Water Development Board, 2005). Table 1 lists
the main components of a typical RHS (Texas Water Development Board, 2005).

Table 1. Typical RHS Components.

Catchment Surface
Gutters
Downspouts
Debris and Dust Removing Mechanisms
Storage Tank (also known as a Cistern)
Water Delivery System
Treatment / Disinfection Equipment

Many home appliances require 20-30 psi of pressure for proper operation, making gravity flow of
water impractical. Therefore, water pumps are typically used to meet pressure requirements. The majority
of energy requirements for RHS is due to water pumping (Chiu et al., 2009). Relevant factors include lift,
water flow, pump efficiency, transmission efficiency, and friction loss of pipes.
Researchers have found that RHS cannot economically compete with local utilities in supplying
potable water; RHS might be competitive for non-potable uses (Texas Water Development Board, 2005).
A perceived lack of economic benefit is often cited as an impediment for widespread RHS adoption, as
upfront capital costs of RHS installation and annual maintenance fees make it difficult to compete against
low-cost and highly subsidized municipal water sources. However, many studies neglect full life-cycle
costing within their benefit-cost analyses. There is a growing need to incorporate social benefits of water
conservation and stormwater capture in addition to financial costing (EPA, 2015). A full life-cycle costing
analysis should include: RHS capital cost, maintenance costs, water conservation benefit, stormwater
management benefit, possible energy conservation and environmental benefits, and the displacement of
the need for water infrastructure development to meet growing water demand (EPA, 2015) (Farreny et al.,
2011) (Sample and Liu, 2014).
Economic feasibility analyses of RHS incorporate several performance criteria in order to gauge
adoption thresholds. Chiu et al. (2009) employed a benefit-cost (B/C) ratio to study the efficacy of RHS
adoption in Taiwan. The study found that when water and energy-saving benefits are combined, the B/C
ratio exceeds the threshold for adoption. If water and energy-saving benefits are analyzed separately, the
B/C ratios are lower than one, signifying an economically unattractive system. Hajani and Rahman (2014)
came to a similar conclusion using a reliability and life-cycle cost analysis for implementing RHS in rural
areas of Australia. Farreny et al. (2011) used a life-cycle costing framework to conduct a net present value
(NPV) study on implementing RHS infrastructure within dense Mediterranean neighborhoods. They
found that capital costs, operational costs, and maintenance expenses outweighed the financial benefit of
not having to use utility-sourced water. This study did not place any value on the benefits associated with
water conservation, stormwater capture, and displacement of the need for growing water supply
infrastructure. Their study found that RHS adoption is not cost-efficient, which is a common trend within
literature under current water pricing (Rahman et al., 2010).
Jenkins and Pearson (1979) studied RHS feasibility in California. They concluded that rainwater
systems in California are feasible. RHS could only partially supply domestic demand however, as the
high cost of the large tanks required to sustain demand during the dry season would make it too costly a
system. However, the authors did not define the difference between the costs, energy, and carbon
emissions of RHS compared to utilities. This is a gap in the literature that this study hopes to fill.
The majority of the literature assesses the performance of RHS using a behavioral model. The
behavioral model is a discrete-time mass balance equation of the water storage in the RHS tanks under
inputs of rainfall, demand, and one of two operating conditions. The operating conditions, yield after
spillage (YAS) and yield before spillage (YBS), are two approximations of the timing between water
demand and supply. In reality, these may occur continuously and at the same time over the course of a

UC Berkeley, CE295
237

day. In YAS, the demand is assumed to occur before any rain replenishes the tank on that day; the
opposite is true of YBS (Jenkins and Pearson, 1979).
The most common method for assessing the optimal system design is through sensitivity analysis.
Most studies attempt to determine optimum sizing of RHS using a variety of system performance indices,
such as water-saving efficiency (WSE) and overflow ratio (OR) (Palla et al, 2011), (Mun and Han, 2012).
WSE is the ratio between volume of rainwater supplied and water demand for a selected time interval. OR
is the ratio between the volume of rainwater exceeding the system capacity and the inflow to the RHS for
a selected time interval. The water-saving efficiency will appear in our multi-objective optimization
function. A notable pattern in RHS literature is a marked absence of optimization of RHS operation with
respect to cost, energy use and carbon emissions. Moreover, when performing benefit cost analyses,
researchers and policymakers have also failed to include the non-monetary benefits of RHS, such as
conservation and stormwater management benefits. This gap in the literature presents an opportunity for
us to apply our new optimization knowledge to a novel research question.

c. Focus of this Study:

The main objective of this study to analyze the ability of RHS to meet economic targets, energy
efficiency goals, and greenhouse gas emission reduction requirements, while being a reliable source for
meeting non-potable water demand. Ultimately, the goal is to assess the conflicting trade-off between
water conservation and emission profiles of RHS.

III. Technical Description

Data Sourcing and Data Prepping:

Rainfall data:
Daily rainfall data for each city were obtained from the National Oceanic and Atmospheric
Administration’s Climate Data Online portal. If multiple rain gauge data were available in the same city, the
gauge closest to the city center was chosen to obtain the best representation of the precipitation experienced by
most buildings in the city. Before we are able to optimize RHS operation, we must determine an
appropriate storage tank size. We followed a popular sizing technique, which ensures a storage capacity
sufficient to store water collected during rainy times to last through dry spells (Texas Water Development
Board, 2005). Given the wide range of dry spell lengths, we took the maximum dry spell from each
calendar year between 1970 and 2017, and took the average of these maximum annual dry spells to size
the tank for each city and building type. The average dry spell lengths and standard deviations are as
follows: 93 +/- 36 days for Oakland and 98 +/- 41 for San Diego. This information helped us design a
tank size that has the capacity to supply demand during an average water year. The size of the tank is
equal to the product of daily demand and the length of the dry period.

Building Parameters Description:

Building type characteristics, including catchment area and non-potable daily water demand, are
provided in Table 2. The data source for the catchment areas comes from a 2012 Department of Energy
report from which all building characteristics (e.g. square footage, catchment area) are derived (Arup,
2012). The non-potable water demand, which is comprised of toilet and urinal water demand, is based on
current research being conducted for the California Air Resources Board (CARB).

Table 2. Building Type Characteristics.

Building Type Catchment Area [m2] Non-potable Water Demand (m 3/day)
Single Family 97 0.0616

The daily potential harvestable volume, or daily RHS yield, is calculated using Equation 1. Sathe et al.
assume a collection efficiency of eighty percent (Sathe et al., 2012):

UC Berkeley, CE295
238

RHS Yield (m3) = Daily Precipitation (m)×Catchment Area (m2) (1)

Seasonality and variability of rainfall yield:

Figure 1 shows mean daily yield, demand and standard deviation for a representative city,
Oakland, from 1970 to 2017. These graphs show a dry season in the summer months, and large standard
deviations in daily yield indicate a need to capture the full range of rainfall dynamics. In Oakland, the
home’s demand is lower than rainwater yield for most of the year; in San Diego, the opposite is true.

UC Berkeley, CE295
239

Figure 1. Mean daily yield and demand for (a) Oakland single family homes and (b) San Diego single family
homes, and (c) standard deviation of daily (maximum) yield for Oakland, CA from 1970 to 2017.

Energy Data Sourcing:

Energy consumption data for RHS pumps were obtained through surveying literature on both
theoretical and empirical studies. This data was then compared to commercially available RHS water
pumps. The heterogeneity of building parameters in this study was taken into account when selecting
energy values (Cheng, 2002) (Viera et al., 2014) (Ward et al., 2011).
The electricity utilities of interest are presented in Table 3. Power mixes from each utility are
analyzed in order to determine the life cycle carbon emission factors used in estimating carbon emissions
from water use.
The energy intensities associated with municipal water use vary by hydrologic region in
California (Stokes et al. 2017). The variances in energy intensities stem from two critical factors: (1)
differing treatment processes and (2) differing transportation distances. Energy intensity related to
wastewater treatment largely depends on the number and type of unit process that each treatment plant
employs. In terms of energy intensity related to transportation distances, the fact that water is not
distributed evenly across the state means that in order to meet regional population demands, water is
diverted from the northern part of the state to the southern regions. The energy required to pump water
over the Grapevine increases the energy intensity of the water supply in the Southern California regions.
Energy intensities for the hydrologic regions of interest are presented in Table 3. They are adapted from
the 2020 regional intensities from (Stokes et al., 2017).

Table 3. Energy and Carbon Intensity Data.

City Hydrologic Energy Electricity Utility Life-Cycle Carbon Intensity (g
Region Intensity CO2e/kWh)
(kWh/m3)
Oakland San Francisco 0.69 Pacific Gas & Electric (PG&E) 145
Bay
San Diego South Coast 2.1 San Diego Gas & Electric 313
(SDG&E)

RHS Cost Sourcing:

The cost components for using water from a RHS and from a centralized, municipal source were
obtained from several sources. Federal and statewide technical reports provided estimates for RHS capital

UC Berkeley, CE295
240

and maintenance costs (EPA, 2015) (Texas Water Development Board, 2005) (Cabell Brand Center,
2009). RHS operational costs depend upon daily energy consumption for pump operation and vary by
each study city’s electricity provider. Operational costs are the only costs considered when using
municipally-provided water. Water utility costs also vary for each study city’s water utility provider.
The valuation of non-monetary benefits of RHS is crucial for a holistic, comparative analysis.
Estimating water conservation benefits is difficult because the price of water is volatile and subject to
local circumstances. For the sake of this analysis, water conservation pricing figures and projections of
growth from the EPA will be adopted (EPA, 2015). Similarly, valuation of stormwater runoff capture will
follow a method developed by Sample and Liu (2014), where the value of runoff capture will be based on
local stormwater utility charges. There currently does not exist any methodology or precedence for
costing the benefits of the displacement of water infrastructure development. Therefore, it will not be
included in this analysis.

Normalized costs and benefits of RHS are calculated in Equation 2:

CR = (Cc + Cm + λER) - (Bw + Bsc) (2)

Where:
CR = total cost of RHS ($/m3)
Cc = capital costs of RHS (normalized to annual basis) ($/m3)
Cm = annual maintenance cost of RHS ($/m3)
𝜆 = cost of electricity consumption ($/kWh)
ER=energy consumed for RHS pumping (kWh/m3)
Bw = benefit associated with water conservation ($/m3)
Bsc = benefit associated with stormwater runoff capture ($/m3)

Optimization Program:

Scope of Analysis:
The optimization program was solved on a daily time step. The daily time step was chosen
because the highest resolution of rainfall and demand data are at the daily level. Since demand and
rainfall occur simultaneously, the relatively coarse time step of 24 hours presents a challenge. We chose
to model the RHS using Yield After Storage (YAS); all demand for the day is assumed to occur before
any rain replenishes the tank on that day. This gives the most conservative estimate of RHS efficacy.
Based on RHS literature, a study period of at least 30 years is required to capture the full range of
rainfall dynamics (Mun and Han, 2012). To ensure that our simulation is representative of all water years,
we used almost 50 years of rainfall data spanning from 1970 to 2017 to study optimal RHS operation.
Due to lack of information about historical demand and utility costs, we simulated a hypothetical scenario
in which these factors remain at today’s levels throughout the study period.

Math Model:
The RHS was modeled following a behavioral model described by (Palla et al, 2011), which is
summarized in Figure 1. The primary constraint is a mass balance around the volume of water stored in
the tank (V), with an influx due to rainfall collection (Q) and outflux of demand satisfied by rainfall (Y)
and overflow (O) when volume stored exceeds storage capacity (S). The total demand for water (D) that
is not satisfied by rainfall (Y) is made up by municipal water supply (M). The state variable, fluxes, and
parameters in this model are listed in Table 4.

UC Berkeley, CE295
241

Figure 2. Rainwater harvesting mathematical model, to be used at a daily timestep. Adapted from: Palla, A; Gnecco,
I; and Lanza, L.G. 2011. Non-dimensional design parameters and performance assessment of rainwater harvesting
systems. Journal of Hydrology. 401, 65-76. Accessed on 2/6/2018 at 11:00 P.M.

Table 4. Variables, inputs, and parameters of the RHS model.

Symbol Description and equation
S RHS tank capacity
Y Rainwater supply (i.e. the amount of demand that is satisfied by rainwater storage). According to the YAS
operating rule, at any time t,
Y(t) = min(D(t), V(t-1))
M Municipal water use (i.e. the amount of demand that is satisfied by the utility)
V The stored volume in RHS tank. According to the YAS operating rule, at any time t,
V(t) = min(V(t-1)+Qt-Yt, S-Yt)
D Total daily water demand, which is assumed constant. At any time t, D = Y(t) + M(t)
O Overflow. At any time step, O(t) = max(0, (V(t-1) + Q(t) - S)
R Raw rainfall depth.
𝞍 Runoff coefficient, or the ratio between rainfall collected to the total rainfall. Typically set at 0.8 (Palla et al,
2011)
A Rainfall collection area.
Q Rain inflow to the storage tank. At any time t, Q(t) = 𝞍*R(t)*A
Smin Minimum technical capacity of the tank. Typically set at 0.5 cubic meters (Palla et al, 2011)

Objective function, optimization variables and constraints:

The optimization variables are the demand provided by rainwater (Y(t)) and the demand provided
by the utility (M(t)). These optimization variables were solved daily over for each city and building type;
simulations initialize with a tank storage of V0 = 0.5*S. Constraints are listed in Table 5. The variables
were optimized for cost, energy use, and water use efficiency at each time step over an average rainfall
year according to the following linear multi-objective function:

UC Berkeley, CE295
 242

Table 5. Constraints and their physical meaning for our optimization program.
Constraint Physical Meaning
Dt = Yt + Mt (5) Total demand (D) each day must always be satisfied either through RHS (Y) or
the utility (M)
Vt = Qt - Vt-1 - Yt - Ot (6) Mass balance of the rainwater stored in the tank
Qt = 𝞍 * Rt * A (7) Inflow of rainwater to the tank is proportional to rainfall and a loss factor.
Yt = min [Dt, Vt-1] (8) The yield of the rainwater storage will either partially or fully meet water
demand.
Ot = max [0, (Vt-1 + Qt) - S] (9) Allows for water overflow discharge when max storage capacity is reached.
Vt ≥ Smin (10) The storage in the tank must be above minimum technical capacity.

Dynamic Programming:
The optimization problem was solved using dynamic programming with the yield (Y) from the
rainwater harvesting tank as the control and storage (V) in the tank as the state variable.

The value function, in words, is as follows:

Let Vk(Yk, vk ) denote the minimum cost from time step k to terminal time step N, where the control, the
yield from the tank, is Yk and the resource level, the storage in the tank, in step k is vk.

The principle of optimality equation is:

Because there is no terminal cost at the end of the year, the boundary condition of the value function is:

VN(YN, vN) = 0 (5)

Due to computation time concerns, optimal yield from the RHS was only calculated over an
average water year. This average water year was calculated for each city and represents the daily average
rainfall that the city received from 1970 to 2017. Using an average water year allowed us to capture the
seasonal patterns of rainfall over the past several decades without needing to optimize each year
separately. For each optimization, the tank was initialized with a storage equal to half the tank capacity.

UC Berkeley, CE295
243

Comparative Scenario Analysis:

Two scenarios are analyzed in order to the assess the feasibility of implementing a RHS. The first
scenario, Scenario 1, utilizes an optimized combination of rainwater and municipally-sourced water to
meet non-potable demand. Scenario 2, which is the baseline condition, assumes that all non-potable
demand is met by each city’s water municipality. Energy usage, GHG emissions, and costs are compared
for each scenario. Figure 3 provides a graphical representation of the two scenarios. Table 6 summarizes
the main parameters used to conduct the analysis.

Figure 3. Flow chart describing the comparative analysis of the two scenarios conducted.

Table 6. Summary of main parameters used in the optimization program.

Oakland San Diego
Water Demand

Single-Family Non-Potable Water Demand [m3/day] 0.0616 0.0616

Municipal / Utility Data
Municipal Water Intensity [kWh/m3] 0.69 2.1
Utility Carbon Intensities [gCO2/kWh] 145 313
Municipal Water Costs [USD/m3] 1.22 1.71
Utility Energy Costs [USD/kWh] 0.309 0.251
RHS Data
RHS Operational Costs [USD/m3] 1.14 0.813
RHS Operational Energy Consumption [kWh/m3] 5.6 5.6

IV. Discussion

Figures 4 and 5 display the results of the dynamic programs conducted for RHS use in a single-
family residential home in Oakland and San Diego, respectively. The results reflect the team’s a priori

UC Berkeley, CE295
244

expectations that the utilization of harvested rainwater follows the expected pattern of yearly rainfall.
Reliance on RHS tends to be higher in the winter and fall than it is during spring and summer.
For the Oakland case (Figure 4), the non-potable water demand is met through RHS in a much
higher capacity than it is in the San Diego case (Figure 5). The observed trend is related to historical
rainfall patterns and the differing energy/GHG intensities of the two regions. Firstly, San Diego is
extremely more arid than Oakland. The average yearly rainfall is much higher in Oakland as clearly
indicated in Figures 4 and 5. Given the expected value of monthly rainfall data, the program attempts to
ration water sourcing from RHS in both Oakland and San Diego. Additionally, in equation (3), one of the
multi-objective criterion in the objective function is to reduce the energy intensity of the RHS system, and
in turn the embodied GHG intensity of the sourced water. Since the RHS has the same energy intensity in
both cities, the minimization objective is mostly contingent on the energy and GHG intensities of the
water supplies and power mixes in both cities. The energy intensity of San Diego’s municipal water
supply is almost three times that of Oakland’s. Additionally, the San Diego power mix GHG intensity is
roughly twice that of Oakland’s.

Figure 4. Dynamic program simulation results of RHS use in Oakland for an average rainfall year.

UC Berkeley, CE295
245

Figure 5. Dynamic program simulation results of RHS use in San Diego for an average rainfall year.

As can be seen in Figures 6 and 7, the results of the dynamic programming simulations indicate
that energy usage and GHG emissions are greater for Scenario 1 than for Scenario 2. This means that the
combination of using RHS water and municipally-sourced water to supply toilet demand is more energy
intensive than using only municipally-sourced water. The use of a RHS would be an appropriate GHG
mitigation strategy if it produced fewer emissions than municipally-sourced water. According to this
report’s system boundaries, a RHS is not an appropriate GHG mitigation strategy. This report only
considered using RHS to supply indoor non-potable demand (i.e., toilet flushing water). RHS might be a
more effective strategy for supplying irrigation demand, as there would be limited pumping requirements,
and thus fewer GHG emissions.
Although a RHS might not serve as an effective GHG mitigation strategy, it is an effective water
conservation strategy. Costs play an integral role when analyzing the water conservation benefits of RHS.
The dynamic programming simulations provide cost differences between the two scenarios. For Oakland,
the cost of using a RHS to meet non-potable demand is roughly commensurate with the cost of using only
municipally-sourced water. In San Diego, utilizing a RHS to partially supply demand costs less than
completely relying on municipally-sourced water. The cost components of the RHS only included
operating expenses and monetized benefits associated with water conservation and stormwater
management. If RHS capital costs were included, Scenario 1 would always be more expensive than
Scenario 2 as the planning horizon is one year. An interesting policy component to consider is the use of
subsidies to cover the capital costs of RHS. Some cities cover the cost of installing basic rainwater
capturing systems. For example, San Diego provides rebates of up to $400 per property for residents who
use rain barrels (City of San Diego 2018). At the state level, a proposed ballot measure, Proposition 72,
would reward homeowners for building rainwater harvesting systems by preventing them from having to
pay increased property taxes (Rogers 2018). As the cost of water increases in the future, water-stressed
regions, such as parts of California, might further consider incentivizing the use of RHS.

UC Berkeley, CE295
246

Figure 6. Energy intensity, GHG emissions, and cost comparison between the two modelled scenarios in Oakland.

Figure 7. Energy intensity, GHG emissions, and cost comparison between the two modelled scenarios in San Diego.

One of the more interesting revelations of the obtained results is the apparent tradeoff between
addressing GHG emissions reductions and water conservation goals. On the one hand, RHS perform well
in addressing water scarcity and resource depletion issues by making use of water that would have
otherwise gone to waste. Conversely, RHS are highly energy intensive in their operation. Having high

UC Berkeley, CE295
247

energy use requirements translates into a higher Global Warming Potential due to excess GHG emissions.
Optimization and control mechanisms are uniquely positioned to assess the tradeoff between the two
aforementioned environmental impact categories. Attempting to quantify this tradeoff through other
methods would be more difficult, and won’t offer the required resolution for an optimal analysis. One can
leverage dynamic programming tools to perform analyses on systems that evolve with time, such as RHS,
and are subject to conflicting constraints to obtain optimal system profiles.
Additional analyses can be performed on the water-sourcing profiles that resulted from the
dynamic programming simulation. By translating both the water conserved and the emitted GHGs into
monetary values, the trade-offs between the two criteria can be quantitatively analyzed. Total water
savings through RHS are expressed in [m3], while GHG emissions are expressed in gCO2e. Water savings
can be expressed in net savings by considering the monetary benefits of water conservation and
stormwater capture ($/m3). GHG emissions can also be expressed monetarily by considering California’s
Carbon Cap-and-Trade program (Global Warming Solutions Act of 2006 (AB32)). The Cap-and-Trade
program provides the social cost of carbon in $/gCO 2e. Table 7 provides the transformation parameters of
both values.

Table 7. Scalar parameters used to transform water savings and GHG emissions into costs.
Parameter Scalar Value
Water Conservation Benefit $1.81/m3
GHG Emission Costs $0.0000151/gCO2e (Climate Policy Initiative, 2018)

Only the Oakland case will be considered for this additional analysis. Figure 4 contains the total
amount of water savings in Oakland. The water sourced from RHS, which is displayed in green, is
considered “water-saved”, and can be directly translated into water conservation benefits. Figure 6
contains the total GHG emissions from using municipally-sourced and harvested rainwater to meet non-
potable demand in Oakland. The GHG emissions resulting from the RHS alone will be taken into
account for the analysis. The carbon costs are found by multiplying the net RHS emissions by the scalar
emission value found in Table 7. Table 8 shows the comparison between the water conservation benefits
of RHS and their respective GHG emissions trade-off for Oakland. Two cases are considered: one in
which capital and annual maintenance costs are included, and another in which capital and maintenance
costs are excluded.

Table 8. Water Conservation / GHG Emissions Trade-off analysis (a negative value signifies net costs, whereas a
positive value represents net benefits).
Including Capital + Excluding Capital + Maintenance
Maintenance Costs Costs

Water Conservation Benefits [$/y] -$1,300 $26

GHG Emission Costs [$/y] - $0.18 -$0.18

Table 8 reaffirms the idea that under current costing schemes, RHS water conservation benefits
do not justify their relative GHG emissions. The annual capital and maintenance costs dwarf the overall
benefits by 50 times. In order for the water conservation benefits from RHS to justify their GHG
emissions, a full annual subsidy of costs would be needed ($1300). Another way in which water
conservation benefits of RHS can be justified is either through increasing the monetary benefit of water
conservation to $92/m3, or through increasing the costs of GHG emissions to $0.086/gCO 2e. Both values
would lead to the trade-off costs and benefits to breakeven. The efficacy of changing costing schemes to
such high values is quite unreasonable. While it is true that water is severely discounted and the costs of

UC Berkeley, CE295
248

GHG emissions are highly undervalued, it seems unlikely that any political or administrative body would
approve of the needed valuations, especially considering the scale at which water is consumed and GHGs
are emitted.
In terms of the overarching analysis of RHS, future recommendations for research should look
into possibly setting energy efficiency targets for RHS, and how those targets could affect the optimal
water sourcing scheme. It would also be interesting to see how RHS perform against other water
conserving technologies such as efficient household water appliances (e.g., toilets, dishwashers, cloth
washers). Specifically, the analysis should consider opportunity costs by looking into whether the needed
subsidies for RHS should be used for other water appliances that may conserve more water. Lastly, in
terms of the dynamic programming aspect of the analysis, stochastic rainfall that is characterized
statistically rather than deterministically should be used in the analysis for increased robustness.

V. Summary

This report presents an analytical framework for assessing the feasibility of implementing
rainwater harvesting systems (RHS) in California to meet non-potable demand in single-family buildings.
Analyzing the feasibility of RHS is especially relevant and important for California because of the state’s
history of water-scarcity and the state’s energy intensiveness of its water supplies. The team chose to look
at Oakland and San Diego for the study, as these two cities represent best and worst case scenarios in
terms of rainfall and energy intensities of their respective water supplies.
The team developed two scenarios to study: (1) Using a combination of harvested rainwater and
municipally-sourced water to meet toilet flushing water, or non-potable, demand in a single-family home
and (2) Utilizing only municipally-sourced water to meet non-potable demand. The team created an
optimization program with the objective of minimizing the costs and energy usages associated with
implementing a RHS and of maximizing the water saving efficiency of the RHS. Using dynamic
programming, the team solved the optimization problem for a representative average year of rainfall. For
both the Oakland and San Diego case studies, energy and greenhouse gas emissions are greater for the
scenario that utilizes harvested rainwater than for the scenario that only utilizes municipally-sourced
water. When including net benefits of stormwater management and water conservation, RHS in both
cities cost less than municipally-sourced water. However, when capital costs are included, RHS are not
cost-effective.
From the results of the dynamic program, it is clear that RHS are currently not an affective GHG
mitigation strategy in California. The water conservation benefits of RHS are clear. Homeowners could
be more likely to implement RHS if local governments or policy-makers provided subsidies to cover the
capital costs of the systems. One way in which the RHS water conservation and GHG emission trade off
can be quantitatively assessed is by translating both values into costs. The translation can be performed
using current costing schemes for water conservation benefits issued by the EPA, and the current cost of
carbon through California’s Cap-and-Trade program. The results reaffirm the notion that the water
conservation benefits of RHS do not justify their GHG emissions. As water availability becomes more
uncertain in the future, RHS might become a more acceptable strategy for conserving water.

UC Berkeley, CE295
249

VI. References

ARUP. 2012. The Technical Feasibility of Zero Net Energy Buildings in California. Report ZNE/219664,
Pacific Gas and Electric.

Cabell Brand Center. 2009. Virginia Rainwater Harvesting Manual. Second Edition. Accessed on
3/18/2018.

Cheng, C.L. 2002. Study of the inter-relation between water use and energy conservation for a building.
Energy Build. 34, 261–266.

Chiu, Yie-Ru; Liaw, Chao-Hsien, and Chen, Liang-Ching. 2009. Optimizing rainwater harvesting
systems as an innovative approach to saving energy in hilly communities. Renewable Energy. 34, 492-
498.

City of San Diego. 2018. “Rainwater Harvesting Rebates”.

https://www.sandiego.gov/water/conservation/rebates/rainbarrel. Accessed on May 2, 2018.

Climate Policy Initiative. 2018. “California Carbon Dashboard”. http://calcarbondash.org/. Accessed on

May 3, 2018

Farreny, R; Gabarrell, X; and Rieradevall, J. 2011. Cost-efficiency of rainwater harvesting strategies in

dense mediterranean neighbourhoods. Resources, Conservation and Recycling. 55, 686-694.

Fewkes, A. 2000. Modelling the performance of rainwater collection systems: towards a generalized
approach. Urban Water. 1, 323-333.

Hajani, E. and Rahman, A. 2014. A. Reliability and cost analysis of a rainwater harvesting systems in
peri-urban regions of Greater Sydney, Australia. Water, 6, 945–960.

Jenkins, D.; and Pearson F. 1978. Feasibility of Rainwater Collection Systems in California. Technical
Report. NTIS 197906. California Water Resources Center.

Mun, J.S. and Han, M.Y. 2012. Design and operational Parameters of a rooftop rainwater harvesting
system: Definition, sensitivity and verification. Journal of Environmental Management. 93, 147-153.

National Oceanic and Atmospheric Administration, Climate Data Center. https://www.ncdc.noaa.gov/

Accessed on 16 Mar 2018.

Palla, A; Gnecco, I; and Lanza, L.G. 2011. Non-dimensional design parameters and performance
assessment of rainwater harvesting systems. Journal of Hydrology. 401, 65-76.

Pacific Institute. 2014. Water Reuse Potential in California. Issue Brief IB:14-05-E.

Rahman, A., Dbias, J., and Imteaz, M. 2010. Sustainability of rainwater harvesting systems in multistory

UC Berkeley, CE295
250

residential buildings. Am. J. Eng. App. Sci.: 3:889-98.

Rodgers, Paul. 2018. “New state ballot measure would reward people who build rainwater collection
systems”. Bay Area News Group.

Roebuck, R.M.; Oltean-Dumbrava, C.; Tait, S. 2011. Whole life cost performance of domestic rainwater
harvesting systems in the United Kingdom. Water Environ. J., 25, 355–365

Sathe, Amul, Michael Noreika, and Miriam Morris. 2012. On-Site Water Generation: An Analysis of
Options and Case Study. California Sustainability Alliance. http://sustainca.org/sites/default/files/On-
Site_Water_Generation-Final_Report.pdf.

Sample, D.J. and Liu, J. 2014. Optimizing rainwater harvesting systems for the dual purposes of water
supply and runoff capture. J. Clean. Prod., 75, 174–194.

Stokes-Draut, J. et. al. 2017. Evaluating the Electricity Intensity of Evolving Water Supply Mixes: The
Case of California’s Water Network. Environmental Research Letters.

Swamee, P. and Sharma, A. 2008. Design of water supply pipe networks. New York: John Wiley.

Tam, V., Tam, L., and Zeng, S. 2010. Cost effectiveness and tradeoff on the use of rainwater tank: an
empirical study in Australian residential decision making. Resour. Conserv. Recycle.; 54:178-86.

Texas Water Development Board. 2005. The Texas Manual on Rainwater Harvesting. Third Edition.
Accessed on 3/18/2018.

U.S. Environmental Protection Agency. 2013. Rainwater Harvesting: Conservation, Credit, Codes, and
Cost Literature Review and Case Studies. Report EPA-841-R-13-002. Accessed on 3/18/2018.

Vieira, A.S. Beal, C.D. Ghisi, E. Stewart, R.A. 2014. Energy intensity of rainwater harvesting systems: A
review. Renewable and Sustainable Energy Reviews, Volume 34, Pages 225-242, ISSN 1364-0321.

Ward, S., Butler, D., and Memon. F. 2012. Benchmarking energy consumption and CO2 emissions from
rainwater-harvesting systems: An improved method by proxy. Water and Environment Journal 26(2)

UC Berkeley, CE295
251

UC Berkeley, CE295
252

Wildfire Modeling: A Case Study on the 2013 Rim Fire

Wes Adrianson, Daniel Herron, Emily Peterson, and Borna Poursheikhani

Abstract
Motivated by recent wildfire events in California, this project seeks to model the spatiotemporal
evolution of past and future fires. Using semi-empirical fire equations and ecological inputs such as
vegetation dryness, wind speed, and terrain data we aim to predict the rate of wildfire spread using a
dynamical model to adapt to real-time changing conditions. We hope that this tool could one day be used
to advise evacuation plans, inform bulldozer lines/pre-burn regions, and guide wildfire mitigation efforts.
Our model will build upon previous research efforts in this field by linking the dynamical analysis of
Mandel with the spatial construction presented by Ntaimo, using Richard Rothermal’s mathematical
wildfire spread model as our empirical foundation. We apply our model to the Rim Fire in California,
modeling the spread of the leading edge of the fire and using a least squares regression to improve our
daily spread predictions. We hope that this implementation of machine learning will fill in the knowledge
gaps that currently exist in first principles understanding of wildfire spread.

UC Berkeley, CE295
253

Introduction
Motivation and Background
As average global temperatures rise, we have seen exacerbated extreme weather events. Higher
temperatures during the spring and summer along with snow melting earlier in the spring have led to hot,
dry conditions. These arid circumstances lead to increased probability, duration, and intensity of wildfires.
Data from the National Interagency Fire Center shows that wildfire-burned areas in the United States have
been increasing since the 1980s. As climate change continues, wildfire frequency and intensity is only
expected to increase. In order to protect the lives and properties of people in high risk areas, there is a
need to prevent these fires or lessen their damage whenever possible.

Recently, California’s Napa and Sonoma Counties experienced a devastating wildfire killing over 40
people and destroying 8400 structures. The suspected origin of the fire was high winds causing power
lines to throw sparks that ignited nearby vegetation. Ideally, we can learn from this misfortune and
prevent similar occurrences in the future. There are many factors affecting the ease with which wildfires
spread. However, by taking into account wind speed, temperature, dryness of vegetation, and proximity to
power lines, our goal is to model the severity of a wildfire starting at various locations. The hope is that
firefighters may us this tool to prevent, and, if necessary, minimize the damage from wildfires.

Battling large scale wildfires includes pre-burning vegetation in the fire’s path in a controlled fashion, or
bulldozing and removing the trees to stop or stall the progression of the fire. Our goal is to create a tool
that firefighters and wildfire experts will be able to use to inform their evacuation plans and bulldozer
lines given a predicted rate of spread. Figure 1 shows a bulldoze fire prevention plan-- a potential real
world application of our modeling tool.

Figure 1​ - Croy Fire Bulldozer Map. Retrieved from http://www.fire.ca.gov/cdf/incidents

UC Berkeley, CE295
254

Literature Review
Mandel et al built a real-time coupled atmospheric-wildland fire modeling system that uses machine
learning to incorporate real data and strikes a balance between model complexity and fast execution by
relying primarily on just two partial-differential-equations in 2D. The equations are based on balance
equations for energy and fuel. Data in the model includes atmospheric conditions, fire, fuel, terrain, and
more. This method uses a first principles model combined with an ensemble Kalman filter technique with
regularization (Mandel et al, 2008). However, the first principles used to formulate this model are not
common in fire spread modeling.

In 1972, Richard C. Rothermel published a widely cited research paper for the USDA Forest Service
titled, “ A Mathematical Model for Predicting Fire Spread in Wildland Fuels”. This paper details the
development of a mathematical model for predicting the rate of fire spread and intensity. The only
necessary inputs of the model are the physical and chemical makeup of the fuel and the expected
environmental conditions. The model is used to assess fire spread and intensity in the National Fire
Danger Rating System. The derived mathematical model for quasi-steady state rate of spread (ft/min) is:

where I​R represents

​ the reaction intensity (btu ft​-2​ min​-1​), ​ξ is the propagation flux ratio, ϕ​W​ and ϕ​S​ are
wind and slope parameters, ⍴​b​ is the fuel particle density, ε is the fuel’s effective heating number, and Q​ig
is the heat of preignition (btu lb​-1​) (Rothermel, 1972).

In another influential paper, “DEVS-FIRE” (Ntaimo, et al) constructed a cellular space model with spatial
fuels data, terrain data, and temporal weather data to predict wildfires across time and space. The model
uses the DEVS, a modeling and simulation framework based on generic dynamical systems concepts, an
emerging tool for modeling complex adaptive systems like wildfires. Our approach is similar to the
DEVS-FIRE model in that it combines Rothermal’s first principles model with a systems technique.

Two of the most widely accepted fire behavior predictive models are FARSITE and BehavePlus.
FARSITE relies on Huygen’s principle of wave propagation, where fire growth is simulated as a 2D
elliptical wave using spatial GIS data. The firefront is projected over a finite time step, in which local
raster information on fuels, topography, and weather predicts a speed and direction of fire spread at each
2.5m x 2.5m cell using the Rothermel model. Aggregating these points around the fire perimeter creates
the model for 2D fire growth. FARSITE is currently used by both CALFIRE and the U.S. Forest Service
for training and operations during large wildfire events (Gollner et. al). The DEVS-FIRE model was
designed to be integrated with a stochastic optimization model for the deployment of firefighting
resources quickly and at low cost (Ntaimo, et al). Researchers noted that different GIS data or
applications of different resolutions resulted in significantly different fire spread shapes. Results weren’t
compared to the spread of real fires, clearly an area of future development in the space.

Two of the major inputs across all simulation methodologies is “class of fuel” and wind. There are 13
classes of fuel as spelled out by the 13 Anderson Fire Behavior Fuel Models. These fuel classes
consolidate twig size and moisture content into a “fuel class” which is directly linked to its flammability.

UC Berkeley, CE295
255

The wind used by Petrasova et al. was broken up into two main components—midflame velocity and
direction, both of which can be spatially variable and must be accounted for in an accurate fire spread
simulation. It is noted by all sources that variability in wind leads to high uncertainty in all fire simulation
models. Similarly, the data required to initialize and parametrize these models such as fuels, topography,
weather, etc., are also subject to large uncertainties and limited resolution (Gollner et al).

Figure 2:​ 13 Classes of Fuel

(Source: Petrasova, Anna et al.)

The original 13 Fire Behavior Fuel Models work well for predicting fire spread during peak season when
conditions are dry. However, these models do not work well for certain situations such as prescribed fires,
wildland fire use, simulating fuel treatment effects, and crown fires (Scott et. al). Scott et. al developed a
new fuel model that incorporated live herbaceous components as dynamic. This means that the load shifts
between live and dead depending on the moisture content. We elected to use the Standard Fire Behavior
Fuel Model for our model. The Fuel Model Parameters are included in Appendix II.

There is currently very little technology being implemented on the ground to help firefighters make
dynamical decisions in real time. There is no shortage, however, in preventative firebreak simulations.
Many fire simulations and spread prediction models use Tangible Landscape and the GRASS GIS
wildfire toolset to model impacts of different orientation firebreaks on the spread of a fire (Petrasova et
al.). Of course these models require knowing the source of the fire, which is unrealistic in a wildfire
scenario. Thus we hope to create a more dynamical system that can augment firefighting decision making
in real time. It is no secret that this is a paramount issue and yet, despite the development of a myriad of
fire models, “their use has been relatively limited operationally” (Gollner et al). This is in large part due
to the lack of fundamental understanding of fire physics as they relate to different regimes (Gollner et al).
Recently, some work has been done by Gollner et al. regarding the bridging of simulation gaps by using
data-driven modeling which essentially supplements low resolution simulations with real-time
observations of wildland fire dynamics.

Focus of This Study

The primary goal of this study is to integrate foundational fire spread theory with the methods we have
learned throughout this course to develop a unique approach to modeling wildfires. We will focus on

UC Berkeley, CE295
256

training a Rothermel-based ordinary least squares model on the 2013 Yosemite Rim Fire in order to
dynamically model and forecast one-dimensional fire spread in the future.

Technical Description
The basis of our model is derived from the fire spread fundamentals provided in Rothermel (1972). In
Rothermel’s analysis, the conservation of energy principle is applied to a unit volume of fuel to show that
the quasi-steady state rate of spread can be defined by the ratio between the heat flux received from the
fire and the heat required for ignition. The resulting rate equation is shown below,

where I​R represents

​ the reaction intensity (btu ft​-2​ min​-1​), ​ξ is the propagation flux ratio, ϕ​W​ and ϕ​S​ are
wind and slope parameters, ⍴​b​ is the fuel particle density, ε is the fuel’s effective heating number, and Q​ig
is the heat of pre-ignition (btu lb​-1​). It should be clear, then, that the rate of spread is very closely related
to the fuel characteristics (surface-area-to-volume ratio, moisture content, etc.) and environmental
conditions (wind and ground slope). These relationships were defined through empirical analysis, and the
results can be found in Appendix I.

Although the Rothermel Model provides a semi-empirical formulation of fire spread, it is very limited in
its scope of application -- a lack of first-principle understanding of fire physics has proven to be a huge
barrier in the field of wildfire modeling/prediction. The industry standard fire modeling techniques
include overload and under supply specific variables within the Rothermel to generate a “realistic”
prediction. This approach is highly subjective and leaves no room for reproducibility in the absence of a
fire expert. We hope that by applying machine learning, more specifically a least squares regression, we
will be able to bypass the need for first principles and provide an actionable, dynamic tool for wildfire
responders to use.

Discretization and Modeling

Several key parameters for the fire spread equations, such as fuel loading, fuel depth, and slope, vary
across space in the landscape. Satellite Lidar data was processed to provide fairly effective estimates of
these parameters, limited only by the level of pixel detail. We utilized ArcGIS to render these spatial
landscape data rasters before ultimately processing the model in Matlab.

Due to the pixelated structure of our data, we discretized the Rothermel fire spread model in space. To
demonstrate proof of concept, our initial model was 1-dimensional. In other words, it calculated the
spread of a fire along a line with parallel wind flow.

The first-generation 1-dimensional model utilized a discrete structure where each “pixel” in our
1-dimensional landscape is defined as 40 feet across. In this example, the landscape is completely
uniform, with typical parameter values for grassland fires, a constant 45-degree slope, and constant wind.
Each discrete unit of time is 1 minute. A “status” variable keeps track of whether a pixel is unburned,
burning, or burned.

UC Berkeley, CE295
257

s ε 0, 1, 2 (0 = unburned, 1 = burning, 2 = burned)

A “fuel load burned” variable keeps track of the percentage of each pixel burned during each time step.

F ε [0, 1]

The for-loops calculate a “rate of spread” at each time step for any pixels that are defined as currently
burning (s=1).
R ε [0, ∞]

Figure 4: ​Illustration of 1-Dimensional pixel

model

Our initial model includes 50 pixels (200 feet

total) and a timeframe of 120 minutes. The plot below shows the calculated and discrete rates of spread
across the 1-dimensional landscape.

Figure 5: ​Calculated and discrete

rates of spread

UC Berkeley, CE295
258

Figure 6: ​Model
animation, paused at three
points in time, t = 1, 40,
and 80 minutes.

The animation shows the spread of the fire (red) across the 1-dimensional landscape.

The second generation of our 1-D model applies the same methodology but uses actual GIS Data to
populate the mathematical model and conduct daily adjustments to the model parameters via historical
burn data. Spatial, elevation and fuel data was obtained for every 37.7 square meter pixel of the Rim Fire.
Each pixel was assigned a Fuel Model Code that corresponds with the Fuel Model Parameters in
Appendix II. This provided the fuel data inputs for the Rothermel equation for each pixel. Additionally,
we were able to calculate the slope based on USGS Digital elevation model data. The World Weather
Online API was used to collect the wind velocity and direction for every hour of the 23 day fire.

Figure 7​: LIDAR Vegetation Type Fuel Loading

Analysis. Areas of yellow and orange represent high
fuel loading.

We took the vector components of the wind and slope data such that all vectors included in the model are
parallel to the 1-D pixel trajectory, creating a chain of discretized fire-spread rates unique to each pixel.
The fire model is then run with these variable rates to observe the spread of the leading edge of the fire in
1-D.

UC Berkeley, CE295
259

Figure 7​: Rim fire burn record from 8.17.13 to

9.3.13 (shapefile provided courtesy of Berkeley
Stephens Lab on Wildland Fire Science).

The third generation of our 1-D model adopts machine learning in the form of a least squares regression,
pulling data from a real fire progression and using the error between actual fire spread and our model to
train our model and improve future predictions. Here, we first had to discretize the Rothermel equation
and multiply by the burn time across a given pixel to calculate a distance burned.

Next, we distributed the first term throughout the rest of the equation and include our optimization
variables, 𝛉. From left to right, each term in this equation can be viewed as the fuel, wind, and slope
influences on the total fire spread in a given pixel.

Our least squares regression is composed of a quadratic optimization that seeks to minimize the error
between the actual fire spread and our model predictions, using 𝛉​1​, 𝛉​2​, and 𝛉​3​ from the above equation.
We were then able to formulate our optimization program as follows, with X representing fuel, wind, and
slope input data for each pixel and Y representing the actual fire spread for each day.

UC Berkeley, CE295
260

For the purpose of this study, we simulated our model’s interaction with daily data by utilizing historical
fire progression maps. In a real world application, this tool would take in real time fire spread data (as
frequently as possible) to train the model quickly and predict the spread over the next few days.

Given the inherent unpredictability of wildfire spread, we hope that our machine learning augmentation
will fill in the gaps between Roethermal’s idealized model and reality. We hope to create a foundation for
dynamical fire modeling and prevention to help mitigate the increasing devastation caused by these
anthropogenically intensified natural disasters.

Discussion

The base Rothermel model successfully tracks the fire progress across 730 discrete units of space (37.7m
pixels) and 23040 discrete units of time (minutes). A variable titled ‘status’ keeps track of which pixels
have burned, are burning, and are not yet burned at each time step. The code calculates a unique value of
R for each pixel that is determined to be currently on-fire at each time step, referencing unique hourly
wind data and pixel-specific slope and vegetation fuel loading data. A general location parameter was
adjusted to tune the Rothermel model so that it would burn roughly the full distance of the fire in the
appropriate number of days. The results of the Rothermel forecast model are plotted with the orange line
in Figure 8. The actual daily burn record of the fire is plotted with the stepwise yellow line.

The second model incorporates learning via the ordinary least squares optimization described in the
previous section. The Rothermel model calculates a unique burn rate for each minute. At each new day,
the code conducts a CVX optimization to adjust the three theta-parameters, which primarily reflect
relative importance of fuel loading, slope, and wind in the model. The previous day’s Rothermel model
forecast is compared to the actual historical record of the fire. The CVX includes bounds of -15 and 15 to
prevent extreme theta definitions.

The Rothermal model with learning demonstrates a successful adjustment of theta parameters after the
rapid fire-burn in Day 4. Real burn data for some days, including Day 5, is missing. The forecast speeds
up again after Day 6, where the optimization again responds to the deficit between the model and the real
burn. After Day 8, the parameters are made smaller again by the least squares optimization, responding to
the model slightly overshooting the actual burn. All 3 theta parameters, representing relative importance
of fuel, wind, and slope in the model, were significant throughout. The theta with the largest magnitude
was wind and the second largest magnitude was slope. This indicates that variation in wind and slope
were generally more significant in predicting the real fire burn rate. The results of the Rothermel model
with learning are plotted with the blue line in Figure 8. The RMSE was 864.5, an improvement compared
to the base model RMSE of 966.7.

UC Berkeley, CE295
261

Figure 8​: Plots comparing the historical burn record with the base Roth. Model and the Roth. Model with
learning.

Rothermal Model Rothermal Model with Learning

RMSE 966.7 864.7

Summary

In the last year, many areas throughout California have been devastated by wildfires. As the impacts of
climate change become more apparent each day, it is clear that these events will continue to pose a serious
threat to the state. The main purpose of this project was to better understand the spatiotemporal evolution
of these fires by creating a dynamical model.

Using the foundational fire spread theory presented in Rothermel (1972) along with vegetation, wind, and
slope data, we were able to model the spread of the 2013 Yosemite Rim Fire. However, the Rothermel
model has several limitations when it comes to extreme fire events such as this. To account for these
shortcomings and to present a new approach to solving the problem of fire spread modeling we applied a

UC Berkeley, CE295
262

linear least squares optimization on parameters representing three of the most significant inputs: fuel,
wind, and slope. The goal here was to outperform the rate of spread estimates presented in the Rothermel
model, and our success can be confirmed by the RMSE values of 966.7 and 864.7, of the base and least
squares model, respectively. This is the first known demonstration of machine learning in wildfire
modeling.

While our model does provide some exciting new results, we believe there still exists significant room for
further research in the fire modeling space. Most importantly, in order to more accurately apply machine
learning algorithms to this problem, it is imperative that more detailed fire spread data be recorded and
made publicly available during the evolution of wildfires. When a model can only “learn” at the end of
each day, it is difficult to make accurate predictions about the hourly behavior of a fire. Additionally, the
model should incorporate data from many different fires in many different regions to best forecast a
variety of possible conditions. Of course, we would also like to expand the model to fire spread in
2-dimensions in future iterations.

Fire spread models are used to inform containment and evacuation efforts anytime a wildfire breaks out.
Improving these models will help insure citizens’ immediate safety and direct the efforts of firefighters
trying to mitigate the spread. In this regard, we hope that our report outlines a useful machine learning
approach that can be used to improve future models of wildfire spread.

UC Berkeley, CE295
263

References

Altinas, I., et al. "Towards Data-Driven Operational Wildfire Spread Modeling."

Coen, Janice L., and Wilfrid Schroeder. "Use of spatially refined satellite remote sensing fire detection
data to initialize and evaluate coupled weather-wildfire growth model simulations." ​Geophysical
Research Letters​ 40.20 (2013): 5536-5541.

Gollner, M., et al. "Towards Data-Driven Operational Wildfire Spread Modelling." ​Report of the
NSF-Funded Wildfire Workshop​. 2015.

Mandel, Jan, et al. "A wildland fire model with data assimilation." ​Mathematics and Computers in
Simulation​ 79.3 (2008): 584-606.

Ntaimo, Lewis, Xiaolin Hu, and Yi Sun. "DEVS-FIRE: Towards an integrated simulation environment
for surface wildfire spread and containment." ​Simulation​ 84.4 (2008): 137-155.

Petrasova, Anna, et al. "Wildfire Spread Simulation." ​Tangible Modeling with Open Source GIS​.
Springer, Cham, 2015. 105-113.

Rothermel, Richard C. "A mathematical model for predicting fire spread in wildland fuels." (1972).

Scott, Joe H., and Robert E. Burgan. "Standard fire behavior fuel models: a comprehensive set for use
with Rothermel's surface fire spread model." (2005).

UC Berkeley, CE295
264

Appendix I: Rothermel Parameters

UC Berkeley, CE295
265

UC Berkeley, CE295
266

Appendix II: Standard Fire Behavior Fuel Model Parameters

Source: ​Scott, Joe H., and Robert E. Burgan. "Standard fire behavior fuel models: a comprehensive set
for use with Rothermel's surface fire spread model." (2005).

UC Berkeley, CE295
267

Appendix III: Fire Model Input Data

Fuel Loading Slope

Actual Burn Record Wind (Northern projection)

UC Berkeley, CE295
268

UC Berkeley, CE295