Purdue University

Purdue e-Pubs
International High Performance Buildings
School of Mechanical Engineering
Conference

July 2018

A Python-Based Toolbox for Model Predictive

Control Applied to Buildings
Javier Arroyo
KU Leuven, Belgium; EnergyVille, Belgium, [email protected]

Bram Van Der Heijde

KU Leuven, Belgium; EnergyVille, Belgium; VITO, Belgium, [email protected]

Alfred Spiessens
KU Leuven, Belgium; EnergyVille, Belgium; VITO, Belgium, [email protected]

Lieve Helsen
KU Leuven, Belgium; EnergyVille, Belgium, [email protected]

Follow this and additional works at: https://docs.lib.purdue.edu/ihpbc

Arroyo, Javier; Van Der Heijde, Bram; Spiessens, Alfred; and Helsen, Lieve, "A Python-Based Toolbox for Model Predictive Control
Applied to Buildings" (2018). International High Performance Buildings Conference. Paper 282.
https://docs.lib.purdue.edu/ihpbc/282

This document has been made available through Purdue e-Pubs, a service of the Purdue University Libraries. Please contact [email protected] for
additional information.
Complete proceedings may be acquired in print and on CD-ROM directly from the Ray W. Herrick Laboratories at https://engineering.purdue.edu/
Herrick/Events/orderlit.html
 3 369, Page 1

A Python-Based Toolbox for Model Predictive Control Applied to

Buildings
Javier Arroyo1,2,3* , Bram van der Heijde1,2,3 , Alfred Spiessens2,3 , Lieve Helsen1,2
1
University of Leuven (KU Leuven), Department of Mechanical Engineering,
Leuven, Belgium
2
EnergyVille, Thor Park,
Waterschei, Belgium
2
VITO NV, Boerentang 200,
Mol, Belgium
* Corresponding Author

ABSTRACT

The use of Model Predictive Control (MPC) in Building Management Systems (BMS) has proven to out-
perform the traditional Rule-Based Controllers (RBC). These optimal controllers are able to minimize the
energy use within building, by taking into account the weather forecast and occupancy profiles, while guar-
anteeing thermal comfort in the building. To this end, they anticipate the dynamic behaviour based on
a mathematical model of the system. However, these MPC strategies are still not widely used in practice
because a substantial engineering effort is needed to identify a tailored model for each building and Heat
Ventilation and Air Conditioning (HVAC) system.
Different procedures already exist to obtain these controller models: white-, grey-, and black-box modelling
methods are used for this end. It is hard to determine which approach is the best to be used based on the
literature, and the best choice may even depend on the particular case considered (availability of building
plans, Building Information Models (BIM), HVAC technical sheets, measurement data). Nevertheless, the
vast majority of researchers prefer the grey-box option.
In this paper a Python-based toolbox, named Fast Simulations (FastSim), that automates the process of
setting up and assessing MPC algorithms for their application in buildings, is presented. It provides a
modular, extensible and scalable framework thanks to its block-based architecture. In this layout, each of
the blocks represents a feature of the controller, such as state-estimation, weather forecast or optimization.
Moreover, the interactions between blocks occur through standardized signals facilitating the inclusion of
new add-ons to the framework.
The approach is tested and verified by simulations using a grey-box model as the controller model and a
detailed Modelica model as the emulator. A time-varying Kalman filter is applied to estimate the unmeasured
states of the controller model.
FastSim is developed and used in a research environment, however this automated process will also facilitate
the implementation of MPC for different building systems, both in virtual and real life.
Keywords: Python toolbox, FastSim, model predictive control, MPC, building management system, HVAC

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 2

1. INTRODUCTION

The buildings sector is responsible for over 30% of total final energy use of all sectors of the economy
worldwide. Despite significant policy efforts to improve energy efficiency in buildings, building energy use
has risen by nearly 20% since 2000 (Building Energy Performance Metrics, 2015). Model Predictive Control
(MPC) is a technique that has proven to increase the energy efficiency of buildings while keeping their indoor
comfort (Coninck & Helsen, 2016). These controllers use a mathematical model of the building together with
weather and occupancy forecast to optimize the control signal. Concerns have shifted from whether MPC
will maintain thermal comfort at a reduced cost to how easily and reliably it can be implemented.
Therefore, there is an actual need of tools that facilitate the implementation of efficient controllers in buildings
and that allow a fair comparison between them. This paper describes the Fast-Simulations-Toolbox or
FastSim, a framework to set-up and assess control algorithms. Section 2 elaborates on the description of
FastSim. Section 3 verifies the toolbox with a simulation example where a detailed model written in Modelica
is used as an emulator. Finally, Section 4 draws the main conclusions of this paper.

2. DESCRIPTION OF THE TOOLBOX

FastSim is a Python based toolbox that helps in the implementation and assessment of control algorithms
in buildings. It provides a modular environment for the application and testing of the controllers. Even
though several types and variations of control algorithms exist, all of them are interfaced in the same way:
at the beginning of each sampling time period Ts , measurements are taken from the plant, then the controller
processes this information and returns a signal consisting of the controllable inputs that should be applied
during that time-step. FastSim standardizes these signals with a pandas data-frame format, which is the
most commonly used package for data-handling in Python, allowing the users to easily plug other control or
plant blocks without hampering the overall functionality of the simulation.
Within FastSim each block is a class that inherits all attributes from the parent class. At the same time, each
class is instantiated inside its parent class. The parent class that holds all the others is called "Simulation".
This class instantiates the controller and plant classes, that are going to inherit the attributes from Simulation
(simulation time, sampling time, states names, . . . ). In the same way, the controller and plant blocks may
instantiate other sub-classes. This architecture results in two main advantages:
• Each block can use all those attributes down-streamed from its parent class for its functionality, without
the need to declare or calculate them again. This ensures that the attributes used in each sub-block
are the same as those from the parent.
• Each block is contained inside the parent class, which makes hierarchy much clearer and natural.
Moreover, this architecture helps in the organization of the attributes generated during the simulation
because they will be stored within the blocks that generated such attributes. Finding and using these
attributes is an important task in the post-processing needed for controller assessment.
This paper presents the case with MPC as controller and an interface with Dymola as an emulator model
(plant). Figure 1 shows a schematic presentation of this case where it is possible to appreciate that the
controller also contains other sub-blocks: a predictor, an observer and an optimizer. These sub-blocks
inherit the attributes from the controller ensuring a coherent implementation. In the diagram, mk are
the measurements taken from the plant model at the beginning of the actual interval k; x̂k are the states
estimated by the observer; d(k,k+N ) are the predicted disturbances during the time horizon Th , i.e. for N
future intervals; and uk is the vector with the control signals to be applied during the actual interval k. All
these signals act between the blocks in a pandas data-frame format. Figure 2 shows the work-flow of the
blocks in a time-line scale. At the beginning of each time step the predictor forecasts the future disturbances
and the observer provides an estimation of the actual states to solve an optimization problem according
to a chosen objective. The solution of the problem is sent as control signals to the plant that returns a
measurement allowing the observer to estimate the new state. The process is then repeated shifting the
optimization horizon one time step. Each block of the diagram (see Figure 1) is explained at each of the

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 3

Simulation
MPC
Optimizer
∑k+N
min j=k J(xj , uj )
uk ,...,uk+N

uk
Controller model Plant
s.t.
x≤x≤x
u≤u≤u

d(k,k+N ) x̂k

Predictor Observer mk

Figure 1: Block diagram of the MPC

following sub-sections.

2.1 Simulation
The simulation block contains all high-level information like beginning and end time of the simulation,
sampling time, variables’ names. . . It also instantiates a controller and a plant block and orchestrates the
interaction between them.

2.2 Plant
The plant block is an interface with a virtual modelling and simulation environment or with an actual
physical building management system. In both cases the plant reads the control signals at the beginning of
each time-step and sends the measurements at the end. The tool integrates two plant blocks developed so
far.
The first developed plant block uses Dymola (“Dymola (Dynamic Modeling Laboratory) Users Manual”,
n.d.) to emulate the physical behaviour of the building system. For the simulation of the first time-step, a
.mos file is written. This is a scripting file with the specifications to translate the model and simulate the first
interval period. For the following steps, the generated dymosim.exe is directly called. This executable file
contains all information of the model. Therefore, it enables the simulation from an initial state defined in a
text file called dsin.txt and returns the final state into dsfinal.txt. Right before each step, the name of the file
dsfinal.txt is changed to dsin.txt to use the last state vector from the previous time-step as the initial state of
the new time-step. The controllable inputs are written in a dsu.txt file that is read by dymosim.exe.
The second developed plant block uses the Functional Mock-up Interface (FMI) standard (Project, 2018) to
emulate the building behaviour from a model in a Functional Mock-up Unit (FMU). The advantage of an
FMU is that it can be generated from any modelling environment. In this case, the plant block uses pyfmi,
which is a Python package with methods to load the FMU, set the initial states, write the inputs and run a
simulation (Andersson et al., 2016).

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 4

x0 x1
Plant Plant

m1 m2

u0 u1
Observer Observer

x̂1 x̂2

Initial States
Specifications x̂0 , Spec.
OCP (0) OCP (1) OCP (2)
Controller Model
Ts , Th

d(t0 , t0 + Th ) d(t1 , t1 + Th ) d(t2 , t2 + Th )

Predictor Predictor Predictor

t0 Ts t1 Ts t2

Figure 2: Work flow of the MPC

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 5

2.3 Controller
The main goal of the controller block is to read the measurements from the plant, process information and
return the controllable inputs to the plant. Two controllers have been implemented so far: a simple hysteresis
controller and an MPC. For the MPC block, a controller model is needed. This is usually a simplified model
of the plant, suitable for optimization. This model may also be used by the observer block to estimate the
initial states.
The main challenge when developing an MPC controller is related to obtaining a satisfactory controller model
(Cigler et al., 2013). For this reason, the developed MPC block contains a package to easily build grey-box
models. This type of models use a combination of prior physical knowledge and monitoring data for the
estimation of their parameters. They are structured as thermal resistance and capacitance network models
with lumped parameters and have demonstrated to be a comprehensive and accurate solution (Coninck &
Helsen, 2016; Coninck et al., 2015). The structure of the network is a choice of the designer who can easily
build the network connecting nodes with simple commands. The values of the parameters are estimated
using the GreyBox Toolbox (Coninck et al., 2015). At any node, the sum of the heat flows flowing into that
node is equal to the sum of heat flows flowing out of that node, leading to Eq. 1:

∑ Tj − Ti ∑ Q̇k
Ṫi = + , (1)
j
Ci Rij Ci
k

where Ti is the temperature of node i, Tj is the temperature of node j that is adjacent to i; Ṫi is the derivative
of the temperature of node i; Ci is the thermal capacitance of node i; Rij is the thermal resistance between
nodes i and j; finally, Q̇k represents each of the heat flows entering the node from internal gains, solar
irradiation, radiators. . . After specifying nodes, resistances, external heat flows and boundary temperatures,
the package automatically builds the associated state-space system as shown in Eq. 2.

} ( ) ( )( )
Ẋi = Ṫi ; ∀i = 1, . . . , n Ẋ A B X
=⇒ = (2)
Ym = Tm ; ∀m ⊂ i Y C D U

In Eq. 2, n is the number of nodes in the network that are represented with the index i and m is the subset of
measured states; U are the inputs to the system which concatenates both, the set of controllable inputs, and
the set of disturbances; A, B, C, D are the state-space matrices of the system containing all model parameters.
These matrices are automatically derived from the RC-model that is used by the controller. It is also possible
to introduce the state-space matrices directly into the controller, allowing not only grey-box models but also
black-box models and white-box models as controller models as long as they can be represented as state
space matrices. A methodology for obtaining linear state space building energy simulation models from
white-box models is presented in (Picard et al., n.d.).
Until now, the derivation of the controller equations assumes continuous-time equations. However, the
continuous state-space matrix representation of the system is discretized with a finite time-step in order
to explicitly calculate the system’s states at all intermediate time steps. The discretization method used
is the zero-order hold method, which assumes that the control inputs are constant during each time step.
For the discretization, the Python Control Systems Library is used (Control Systems Library for Python,
2018).

2.4 Predictor
The predictor is the MPC block that provides information about relevant external influences (mostly weather
conditions) during the optimization horizon. It predicts the disturbance inputs, i.e. outside temperature,
solar irradiation, . . . The tool integrates a block able to download weather data from (weather forecast
webpage, 2018) provided the latitude and longitude coordinates of the building are known. It also integrates
blocks to read weather data statically from .TMY and .CSV files. A block able to predict occupant behaviour
has not yet been integrated.

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 6

P rediction
x̂0 x̂k−1
xpk = Ax̂k−1 + Buk−1
P0 Pk−1
Pkp = APk−1 At + Qk

k →k−1
U pdate
x̂k Pk = (I − KC)Pkp Pkp C t mk
K=
x̂k = xuk CPkp C t + R
xuk = xpk + K(mk − Cxpk )

Figure 3: Block diagram of the time varying Kalman filter

2.5 Observer
The observer estimates the initial states of the system at the actual time-step x̂k from the previous state
estimate x̂k−1 and the measurements of the current time-step yk . A time-varying Kalman filter has been
developed and is already integrated into the toolbox. It is used for stochastic estimation from noisy sen-
sor measurements. The Kalman filter is essentially a set of mathematical equations that implement a
predictor-corrector type estimator that is optimal in the sense that it minimizes the estimated error covari-
ance (Laaraiedh, 2009). It works by assessing the process covariance matrix and accordingly weighting the
states estimated by the controller model and the measurements to get a state estimate with the information
coming from both sources. The Kalman matrix varies over time, giving more weight to the predicted states
or the measurements depending on the process and measurement noise. After some time, a time-invariant
matrix should be achieved.
The estimation process is presented in Figure 3. In the first time-step, it is required to initialize the states
x̂0 and the covariance matrix P0 . The process has two main steps. The first step is the prediction of the
new state xpk and covariance matrix Pkp . For that purpose, the same state-space matrices of the controller
model are used. uk−1 is the vector of inputs applied during the previous time-step and Q is the process noise
matrix, a diagonal matrix with an estimation of the variance of the model error for each state.
The second step is the update of the Kalman gain matrix, K, and state vector xuk with the new measurements
mk . R is the measurement noise matrix. Note that K is an estimate of the truthfulness of the model
compared with the measurements and varies over time. This gain is used to weight the correction factor
for the predicted state xpk . This correction factor is called "innovation" and is the difference between the
measurements and the expected values of the measured states obtained with the controller model. Once the
update is made, the covariance matrix is also adjusted with the Kalman gain and the process is repeated
again for the following time-step.

2.6 Optimizer
The optimizer is the algorithm that aims to find the optimal controllable inputs to the building by minimizing
the objective function and meeting the constraints associated with the comfort bounds and physical system
limitations. The constraints also include the equations of the controller model to include the dynamics of
the physical system into the optimal control problem. The objective function penalizes the heating cost and

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 7

the discomfort as can be seen in Eq. 3.

∑
N
min (uk + wδk )∆t, (3a)
x,u
k=1
xk+1 = f (xk , uk , dk ) ∀k (3b)
xk − δ k ≤ xk ≤ xk + δ k ∀k (3c)
uk ≤ uk ≤ uk ∀k (3d)
δk ≥ 0 ∀k (3e)
(3f)

Where N is the number of time steps contained in the prediction horizon, x are all states of the building
system that should be kept between the lower and upper comfort bounds: x and x; u stands for the
controllable inputs and represents the thermal power used to heat the building; u and u are the minimum
and maximum possible values of the controllable inputs; δ is the discomfort vector, which represents the
lower and upper deviations, respectively, of the actual temperature relative to each comfort bound. This
discomfort is weighted with a constant w that accounts for the different orders of magnitude between power
and temperature. Note that these temperature deviations are possible because the comfort is introduced as
soft constraints into the optimization problem. f is the model of the building that predicts the future states
x for a given set of controllable inputs and disturbances d.
FastSim automatically builds the optimization problem from the controller model in Pyomo which is a com-
piler and solver interface for optimization problems in Python (http: / / www .pyomo .org/ , 2018). Pyomo
can interface with a multitude of optimization solvers, and depending on the solver’s capabilities both linear
and non-linear problems with or without integer variables can be solved. Bonmin (COIN-OR, 2018) has been
chosen as the default solver for the optimization block of the toolbox because of its proven good performance
and its possibility to solve mixed-integer non-linear problems.

3. A SIMULATION EXAMPLE

A simulation of four winter months has been carried out to verify and test the performance of FastSim.
The MPC controller block is used with the time varying Kalman filter and the optimizer presented in the
previous section. The plant is an interface with Dymola where a detailed emulator model is running. Each
time-step the emulator runs a simulation of one sampling time period that is chosen to be one hour. During
this period, the heating inputs are those calculated by the MPC and the weather information is taken from
a TMY file. This detailed emulator model is considered to be an accurate representation of a nine-zones
residential building. However, to challenge the controller in an even more realistic scenario, the plant adds
white noise with a standard deviation of 0.3◦ C to each measurement before passing the measurement vector
to the controller.
At the controller side, each time-step the predictor reads the same TMY weather file used by the emulator
to take the (perfect) weather forecast during the prediction horizon; then the Kalman filter estimates the
initial states of the controller model from the noised measurements; finally, the optimizer estimates the
optimal inputs according to this information. These inputs are read by the plan and the process is repeated
again.
The lower and upper comfort bounds are time varying, being less restrictive during night-time and during
the weekend periods. The only technical constraint is a maximum heat power supply of 2 kW to each zone.
No cooling is considered in this example.

3.1 Emulator model

The emulator model is used for the virtual representation of a real building. In this example, the emulator
model is a detailed white-box model of a nine-zones residential dwelling that has been modelled using the

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 8

Modelica IDEAS library (Baetens et al., 2015). The IDEAS (Integrated District Energy Assessment by
Simulation) Modelica library is a modelling environment that enables thermal simulation of the building
envelop and its HVAC systems.
The emulator model is considered to be an accurate representation of a real building. In this particular
example, the model has 426 continuous time states and 26078 differential-algebraic equations defining the
heat transfer and other physical processes of the building at one minute sampling time accuracy. However,
only the nine air-zones temperatures are being measured and communicated to the controller. The heating
inputs are idealized, i.e. we assume that we can inject a desired amount of heat to each zone.
To initialize the emulator model it is run for one day with realistic inputs right before the beginning of the
four-months simulation. At the first time-step the observer assigns the value of 20◦ C to the unmeasured
states.

3.2 Controller model

A decentralised multi-zones grey-box model is identified as a controller model. The model is decentralised
because no interaction between zones is included. However, at the end of the identification process, the zones
are aggregated and the model is treated as a whole. Centralised models have been tried as well, but they
increased the complexity of the overall model without improving its prediction performance.
The month of January is chosen to generate data from the emulator to train and cross-validate the con-
troller model. Thus two sets of data are generated during this period. For the sake of data generation, an
hysteresis controller tries to keep the temperature of the building within 21◦ C and 23◦ C by tracking only
the temperature of the living room. The controller switches on 3 kW of thermal power when the sensor
detects that the temperature is under the lower comfort bound and switches it off when the higher comfort
bound is surpassed. The heat is distributed proportionally to the area of each zone. This type of control
mimics a traditional, rudimentary rule-based controller and leads to enough level of excitation to obtain rich
data.
Once the data is obtained it is split in two weeks for training and two weeks for cross-validation. The
GreyBox toolbox (Coninck et al., 2015) is used to estimate the parameters of the RC model. This toolbox
enables the estimation of grey-box model parameters in a semi-automated process and from model structures
built in Modelica. The model of each zone is identified independently, trying several zone-structures between
one and three states and different initial guesses for each. At the end, the attempt that provides the best
result in cross-validation is chosen. The controllable input to each zone model is the heating delivered to
the zone and the disturbances are the ambient temperature and solar irradiation.
The controllability of the aggregated model has been successfully tested. Then, it has been used as controller
model following the steps explained in Section 2.3. The green bars in Figure 4 show the Root Mean Square
Error (RMSE) for the obtained model in auto- and cross-validation in units of ◦ C. These are the RMSEs of
the model for a pure simulation during the two weeks of training data and the two weeks of cross-validation,
respectively.

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 9

1.0
Standard deviation of the measurement noise
Cross validation (simulation)
Auto validation (simulation)
0.8 One step prediction
One step updated

0.6
RMSE

0.4

0.2

0.0
1 2 3 4 5 6 7 8 9
Zone

Figure 4: RMSE of the controller model for each zone temperature for pure simulation of the
training and cross-validation period, for the one-step ahead prediction and for the one-step
ahead prediction updated with the Kalman filter. The dashed line represents the standard
deviation of the white noise associated with the measurement error. The states updated
with the Kalman filter are always more accurate than the one-step ahead predictions and the
measurements.

3.3 Results
The simulation results prove the good performance of the MPC within the FastSim framework. The first
sub-plot of Figure 5 shows the emulated temperatures for the nine zones, the comfort bounds (in green) and
the ambient temperature (in blue). The second sub-plot depicts the thermal power released to each zone
and the total solar irradiation per square meter (direct plus diffuse). From these graphs, it is possible to
appreciate how this controller keeps all zone temperatures at the lower part of the comfort range and takes
advantage of the less restrictive periods ensuring comfort at the lowest operational cost.

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 10

30
25
z_1
20 z_2
z_3
15 z_4
z_5
∘C

10 z_6
z_7
5 z_8
0 z_9
T_e
−5

2000
1750 Q_1
1500 Q_2
Q_3
1250 Q_4
Q_5
Watts

1000 Q_6
750 Q_7
Q_8
500 Q_9
irr
250
0
2017∘12∘01 2017∘12∘15 2017∘12∘29 2018∘01∘12 2018∘01∘26 2018∘02∘09 2018∘02∘23 2018∘03∘09 2018∘03∘23
time

Figure 5: Evolution of the zone temperature and the heat inputs during four months of
simulation. The MPC controller keeps the temperatures of the zones within the comfort
bounds while minimizing the energy use.

Even though the controller model complexity is noticeably lower than the complexity of the emulator model
(23 states versus 426), it forecasts the future thermal behaviour of the building with enough accuracy thereby
keeping the zone temperatures well within the comfort bounds. The total discomfort for all zones is of 540
Kh for the four months of simulation.
The Kalman filter safeguards the balance between measurement and process error by estimating the initial
states of the controller model. In Figure 4 the RMSE for the one-step ahead prediction of the controller
model and the RMSE when the Kalman filter updates such predictions with the measurements are also
shown. These are the two main steps of the Kalman filter process. From Figure 4 we appreciate that
the two-weeks simulation leads to larger prediction errors in cross validation. It is also worth noting that
the updated states are always more accurate than the predictions of the controller model and than the
measurements that are white-noised around their true value with a standard deviation of 0.3◦ C (see dashed
grey horizontal line in Figure 4). This standard deviation is considered a reasonable value for a conventional
thermostat.
As explained in sub-section 2.5, the elements of the time-varying Kalman filter vary over time to find a

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 11

trade-off between model predictions and measurements. The evolution of these elements’ values over the
three first days of simulation are depicted in Figure 6, which shows that the filter needs around two days
to achieve a time-invariant value of the elements of its matrix gain. Once they converge to their steady
values, they remain constant during the rest of the simulation. However, in a real case where the noise is
not perfectly white they would keep on changing, adapting themselves over time.

1.50

1.25

1.00

0.75
K

0.50

0.25

0.00

−0.25

12-01 04 12-01 16 12-02 04 12-02 16 12-03 04 12-03 16

time

Figure 6: Evolution of the elements of the Kalman gain matrix for the first three days of
simulation. These elements weigh the information coming from the measurements and the
one-step ahead predictions of the controller model to estimate the initial states. After two
days, the filter converges towards a steady gain that remains constant for the rest of the
simulation.

The results for the first zone (living room) during the first week of December are plotted in Figure 7 to
show in more detail the evolution of the temperature. The blue line represents the temperature as emulated
by the emulator model, i.e. what is considered in this simulation example the real temperature value. The
blue triangles are the measurements as given to the controller model. They do not coincide with the blue
line. This is due to the measurement error that has been artificially added. The red line is the living room
temperature as predicted with the controller model, i.e. its one-step ahead forecast. Finally, the yellow
line is the temperature state updated with the Kalman filter from the measurement and model prediction
information. Notice that this updated temperature is always between the predictions of the controller model
and the measurements.
One benefit of the Kalman filter is that it provides a Gaussian distribution for the state estimates, whose
mean is given by x̂(t) and mean square value of the estimation error is given by the trace of the covariance
matrix P (Moura & Zhang, 2016). Consequently, the standard deviation around each state estimate is
calculated as the square root of the trace of P and enables to draw a confidence interval around x̂(t) that is
depicted in the figure with the dashed yellow lines. According to the properties of a Gaussian distribution,
we can ensure that the true state should be within such interval with a probability of 68.2% (Ribeiro, 2004).
Such covariance enables to pursue robust control strategies, but this has not been implemented in this
paper.

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 12

28 True value
Measured
Updated
26 Confidence interval
Predicted

24
∘C

22

20

18

2017-12-01 2017-12-02 2017-12-03 2017-12-04 2017-12-05 2017-12-06 2017-12-07 2017-12-08

Figure 7: Evolution of the states of the living room as simulated with the emulator model
(true values), predicted with the controller model and updated with the Kalman filter for the
first week of simulation. The Kalman update always falls between the measurement and the
one step ahead prediction of the controller model.

4. CONCLUSIONS

This paper describes FastSim, a Python-based toolbox that facilitates the implementation and assessment
of control algorithms in buildings. More specifically, it automatizes the process of setting-up an MPC
controller in a building, enabling a broader implementation of MPC. Moreover, the toolbox has a modular
and extensible architecture and is projected to be used not just for testing at simulation level but also for
real applications.
FastSim has successfully been tested by simulation for a four months Winter period. In the simulation, a
detailed Modelica model plays the role of an emulator from where artificially noise-corrupted measurements
are taken. MPC is used with a decentralised multi-zones grey-box model as controller model. This model
has demonstrated satisfactory prediction performance even for a simulation of two weeks in cross-validation.
In general, the MPC manages to keep thermal comfort within all zones at a low energy cost. A time-varying
Kalman filter is used to estimate the initial states of the controller model. The matrix-gain elements of the
filter achieve convergence after two days. Moreover, it has been shown that the states updated with the
Kalman filter always lead to higher accuracy than the one-step ahead prediction of the controller model and
the measurements.
As a conclusion, we can state that the FastSim toolbox is a step forward towards the broad implementation
of MPC in practice.

ACKNOWLEDGEMENT

The authors gratefully acknowledge the financial support of the European Commission in the H2020 pro-
gramme under Grant Agreement no. 731231 Flexible Heat and Power that has made possible the research
leading to these results. The work of Bram van der Heijde is financed by VITO through a PhD Fellowship,
and through the project "Towards a Sustainable Energy Supply in Cities", which receives the support of the
European Union, the European Regional Development Fund ERDF, Flanders Innovation & Entrepreneurship
and the Province of Limburg.

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018
 3 369, Page 13

REFERENCES

Andersson, C., Åkesson, J., & Führer, C. (2016). Pyfmi: A python package for simulation of coupled dynamic
models with the functional mock-up interface (Vol. LUTFNA-5008-2016) (No. 2). Centre for Mathematical
Sciences, Lund University.

Baetens, R., Coninck, R. D., Jorissen, F., Picard, D., Helsen, L., & Saelens, D. (2015). Openideas - an open
framework for integrated district energy simulations. Proc. 14th Conference of International Building
Performance Simulation Association, Hyderabad, India, Dec. 7-9 . (https://github.com/open-ideas)
Building energy performance metrics (Tech. Rep.). (2015). International Energy Agency.

Cigler, J., Gyalistras, D., Siroky, J., Tiet, V.-N., & Ferkl, L. (2013). Beyond theory: the challenge of
implementing model predictive control in buildings. Proc. 11th REHVA World Congress Clima 2013 .
COIN-OR. (2018, April). (https://projects.coin-or.org/Bonmin)
Coninck, R. D., & Helsen, L. (2016). Practical implementation and evaluation of model predictive control
for an office building in Brussels. Energy and Buildings, 111 , 290–298.

Coninck, R. D., Magnusson, F., Akesson, J., & Helsen, L. (2015). Toolbox for development and validation
of grey-box building models for forecasting and control. Journal of Building Performance Simulation, 9 ,
288–303.
Control systems library for python. (2018, April). http://github.com/python-control/python-control.

Dymola (dynamic modeling laboratory) users manual [Computer software manual]. (n.d.).
Laaraiedh, M. (2009, June). Implementation of Kalman Filter with Python Language. The Python Papers
Journal.
Moura, P. S., & Zhang, G. D. (2016). Ce 295: Energy systems and control (Vol. LUTFNA-5008-2016).
University of California, Berkeley.
Picard, D., Jorissen, F., & Helsen, L. (n.d.). Methodology for obtaining linear state space building energy
simulation models. Proceedings of the 11th International Modelica Conference. September 21-23, 2015,
Versailles, France.
Project, M. A. (2018, April). The functional mock-up interface standard. (http://fmi-standard.org/)

Ribeiro, I. (2004, February). Gaussian probability density functions: Properties and error characterization.
, 5.
(2018, April). (http://www.pyomo.org/)
weather forecast webpage, D. (2018, April). (https://darksky.net/dev)

5th International High Performance Buildings Conference at Purdue, July 9-12, 2018