G Model

EPSR-4738; No. of Pages 13 ARTICLE IN PRESS

Electric Power Systems Research xxx (2016) xxx–xxx

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Modeling study on flexible load’s demand response potentials for

providing ancillary services at the substation level
Ryan Tulabing a , Rongxin Yin a , Nicholas DeForest a , Yaping Li b , Ke Wang b , Taiyou Yong b ,
Michael Stadler a,∗
a
Energy Storage and Distributed Resources Division, Lawrence Berkeley National Laboratory, Berkeley, CA, USA
b
China Electric Power Research Institute, China

a r t i c l e i n f o a b s t r a c t

Article history: Demand response (DR) is an important component for the establishment of smart electricity grids. It can
Received 24 February 2016 decrease the system peaks through load shedding or shifting and optimize the utilization of the existing
Received in revised form 8 June 2016 grid assets, which delays the need for costly upgrades. DR can also enable the integration of intermittent
Accepted 9 June 2016
and distributed energy resources (DER) into the existing electricity grid. Fast DR from aggregated flex-
Available online xxx
ible loads can provide ancillary services (AS) to absorb grid disruptions and may replace the expensive
fast-ramping reserve generation units. This study presents a methodology for load aggregation based on
Keywords:
the prioritization of loads according to their flexibility. Different flexible load types are categorized as
Demand response
Distributed energy resources
thermostatically controlled loads (TCL), urgent non-TCL, non-urgent non-TCL, and battery-based loads.
Thermostatically controlled loads Models based on their physical behaviour are developed and simulations performed to apply the pro-
Battery-based Loads posed aggregation and control algorithm. Results show that the loads during peak hours can be shed off
Load prioritization without rebound demand spikes after the DR event commonly seen in other types of DR programs. The
Load aggregation algorithm also automatically adjusts the power demand according to the output of the distributed renew-
able generation, mitigating disruptions due to variations of the DER output. Additionally, the algorithm
is able to adjust the load demand dynamically according to the fluctuations of electricity price.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction price responsive demand [1–4]. However, the use of DR for ancil-
lary service is different from that of traditional DR applications as
Grid frequency control and power balancing are traditionally DR for AS require fast response and high accuracy. Recent studies
done by ancillary units commonly composed of fast ramping gen- have demonstrated the use of demand-side resources to provide
eration units such as gas turbines and diesel generators. However, ancillary service in the electricity market [4–6]. AS can be classified
as new rules to reduce greenhouse gas emissions are implemented, into three categories: regulation, flexibility, and contingency [7].
system operators are turning to non-fossil fuel powered resources. Ma et al. [5] and Kirby [8] described generalized DR product defi-
Increased penetration of renewable generation from solar and nitions for load participation in AS, energy, and capacity markets.
wind, which are intermittent and non-dispatchable, is further driv- Those DR product are defined by the response time, the length of
ing the need for fast ramping resources. From this point of view, the response, the time to fully respond and the event times being
the use of flexibility of demand-side resources and availability of called. Regulation service refers to the capacity to respond to ran-
real-time signals communication in the electricity grid enables the dom deviations from the scheduled net load. Response time for
interactions between the supply and the demand. this type of ancillary service vary between 30 s and 5 min and usu-
Over the past decades, considering the grid issues of power ally lasts for 15 min. Flexibility ancillary service addresses the large
imbalance and peak demand, demand response (DR) has proven unforeseen deviation of wind or solar output responding as fast
to be a viable option by load shedding and load shifting in response as 5–20 min for a duration of 1 h. Meanwhile, contingency service
to the need of grid. A number of studies have demonstrated the tra- is allocated for immediate response to sudden loss in generation.
ditional DR for emergency load relief, peak load management, and Contingency services are required to respond as fast as 1–10 min
holding for a duration of 30 min or less [7]. In the US electricity mar-
ket, such as CAISO (California ISO) and PJM region, frequency service
∗ Corresponding author at: 1 Cyclotron Road, Berkeley, CA 94760, USA. requires 4 s response to track automatic generation (AGC) control
E-mail address: [email protected] (M. Stadler). signals.

http://dx.doi.org/10.1016/j.epsr.2016.06.018
0378-7796/© 2016 Elsevier B.V. All rights reserved.

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
2 R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx

From the perspective of DR enablement, DR can be considered

broadly to fall into two categories: direct control and indirect con-
trol. Direct control enables the grid operators to directly turn on
or off the customer’s loads or change the operating setpoints (e.g.
thermostat control) after a short notice. Alternatively, under indi-
rect control, grid operators send requests to reduce load demand to
customers, who have the choice to participate or not. In the context
of AS described above, demand response has to be fast and auto-
matic (i.e. Auto-DR) [9]. The variability in customer response time
makes indirect control less reliable compared to the direct control
DR, and unsuited for this application. Different control method-
ologies for harnessing the DR potential of heating ventilation and
air conditioning (HVAC) system were evaluated by [10,11]. Results
show that significant demand relief can be taken from tempera-
ture reset control and pre-cooling control but with energy and cost
penalties. Highest demand relief can be attained by curtailment
control but can only be used in shorter duration as the indoor tem-
perature quickly exceeds the comfort level for most humans. For
demand response lasting for several hours, pre-cooling control is
among the best options [12–14]. However, with regards to DR for
AS, battery-based storage and electric vehicle (EV) have proven to
be viable options for grid application, ancillary services such as
frequency regulation in particular [15–18]. When considered as
Fig. 1. Aggregation of flexible loads.
flexible resources, a number of recent studies have demonstrated
the use of EVs for increasing penetration of renewable generation
resources. In [19], the authors design three a suitable modeling of the definition of flexibility for each load type, and the power allo-
electric vehicles with three types of controls (night charge, intel- cation algorithm that prioritizes loads based on their flexibility.
ligent charge and vehicle to grid) to analyze the impact of EVs on Section 4 uses the model detailed in Section 2 and control algo-
energy systems. The authors of [16] conduct several simulations rithms detailed in Section 3 to demonstrate the use of aggregated
to show the adoption of advanced centralized EV charging con- DR resource under several scenarios. Finally, the conclusion and
trol strategies and allow the integration of a larger number of EVs recommendations for future investigations are presented in Sec-
in the system. On the other hand, the adoption of a local level tion 5.
of control will allow a better operation performance of increas-
ing penetration of intermittent and variable renewable generation 2. Methodology
resources installed in isolated power systems. In [20], the authors
investigate the use of plug-in EVs (PEVs) to balance the fluctua- This study uses the bottom-up approach in modeling customer
tion of renewable energy sources and study the benefits of fleet loads and their corresponding DR potential. Individual models of
EVs in response dynamic price signals. In [21], the authors explore electrical loads are developed based on their physical properties
the potential of PEVs to balance variability and uncertainty from and behavior. The models are then aggregated in a simulation to
wind and solar generation resources. The authors demonstrate the mimic the load demand at the substation level where the proposed
use of a large number of PEVs to provide ancillary services in the controller should be located. Fig. 1 illustrates the setup of the con-
regulation (secondary frequency control) time frame by leveraging troller. Loads will send requests to turn on, then the controller will
emerging information and communication technologies, and con- prioritize all requests and grant permission according to available
clude that roughly 3 million PEVs with a charging rate of 3.3 kW power from the DER and the grid. It is assumed that the commu-
and no V2G capability would suffice to supply a large part of the nication line between the aggregator and the loads exists and that
regulation up and down demand in California. the central controller has the ability to prohibit a load from starting.
In the domain of demand response in buildings, recent studies Two conditions are required for a load to start: load request trigger,
show that using commercial and residential HVAC load control in and permission from the central controller. The central controller
grid operations can provide power regulation and ancillary services gathers the load status or simulate the load status, prioritize the
[22–28]. Most research have demonstrated the value of demand- loads according to their flexibility, then allocate power by send-
side flexible loads individually at each category of DR product. ing a “permission to start” signal to the loads. If a particular load
However, at the substation level, it is quite challenging to aggregate requests to start but no permission has yet been granted from the
different type of flexible loads at the same order due to different central controller, it shall wait until the permission to turn on is
response characteristics (e.g. response availability, depth and dura- given – inherently shifting the load demand.
tion). There is a need to provide a solution to aggregate, control
and optimize each type of flexible load’s DR resource in the grid 2.1. Load classification
operation.
In this study, we develop a suite of bottom-up physical mod- Electrical loads can be classified as: thermostatically controlled
els to quantify aggregate DR potential from residential sector. loads (TCL); non-thermostatically controlled loads (non-TCL); and
Specifically, we propose a general methodology for a DR con- battery-based loads. TCLs include HVACs, water heaters, refrig-
troller which aggregates common types of residential electrical erators, and freezers. Non-TCL loads can further be classified as
loads, EV and storage, and predicts their potential to provide urgent or non-urgent. Urgent loads are types of load that has to
demand response resource. The paper is organized as follows: respond to user’s request instantaneously after the switch is turned
Section 2 describes the methodology for modeling each type of on (e.g. lights, cooking appliances, entertainment appliances). Non-
flexible load in residential sector, as well as models for distributed urgent loads are those that can be started after some allowable time
energy resources (DER) like solar and wind. Section 3 introduces delay such as dishwashers, washing machines, and clothes dryers.

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx 3

where ṁ is the mass flow rate, and Cp is the specific heat capacity
of air or water (in the case of a water heater).
The common control algorithm for TCLs based on the temper-
ature set-point and the tolerance (deadband) is used in this study.
However, an additional decision gate has been established as pre-
requisite before a particular TCL turns on.

2.2.2. Battery-based load model

The usual charging process of battery-based loads starts with
a constant current (CC) mode followed by a constant voltage (CV)
mode [30]. At lower charge level, the charging current is fixed at
a certain value, based on the charger capacity while the voltage is
allowed to swing. When a certain charge level has been reached,
the charging voltage is fixed at a certain value while the current
is allowed to fluctuate. In this study, it is assumed that the power
demand (voltage × current) for charging stays constant while the
ratio between voltage and current changes.
Given the capacity of a battery in kilo-Watt-hours (kWh), the
length of time for charging (or discharging) is a function of the
charging power.
Fig. 2. Representation of equivalent thermal parameter (ETP) model.
CB (1 − SOC0 )
tf − t = (6)
Battery-based loads, including electric vehicles (EV) and stationary Pcharging charging
batteries, store energy and should be fully charged in a specified where tf is the predicted time when the charging finishes [h], t is
time as required by the user. Battery-based loads are assumed to be the current time [h], CB is the capacity of the battery (kWh), SOC0 is
interruptible and can be delayed as long as they meet their required the existing state of charge [%], Pcharging is the charging power [kW],
charge status on the time. By inference, TCLs and battery-based and charging is the charging efficiency.
loads have the capacity to store energy (i.e. thermal or chemical) The SOC can also be determined as a function of charging power
hence, they are inherently flexible. Non-urgent non-TCL loads are and the charging time.
also flexible as their starting time can be delayed.
Pcharging ∗ t ∗ charging
SOC = SOC0 + (7)
2.2. Load characterization CB
where t is the duration of charging (t = tf − t).
2.2.1. TCL model
The most common method to model TCLs is the equivalent ther- 2.2.3. Urgent non-TCL model
mal parameter (ETP) model. A schematic representation of this Urgent non-TCL (UNTCL) loads include lighting, and miscella-
modeling approach [29] is shown in Fig. 2. neous appliance (e.g. for entertainment, cooking, cleaning, etc.) that
In this representation, Ua is the coefficient of heat transfer directly interacts with the end-users. The models used to account
between the inside air and the outside environment, Um is the the demand variations from these type of loads are based on a
coefficient of heat transfer between the mass inside the thermal regression curve that fits the typical residential load curve from
zone and the inside air, Ca is the heat capacity of air, Cm is the heat a suite of prototype buildings. The derived regression functions
capacity of the internal mass, Qa is the net heat introduced to the depict the occupancy schedule on the time of the day. Since these
inside air, Ta , To and Tm are temperatures of the inside air, outside loads need to start instantaneously, their flexibility is zero and they
environment, and the internal mass respectively. Qa is the sum of are prioritized during the power allocation process. The following
the external heat gain (Qex ), internal heat gain (Qin ), and the heat equation expresses the total power demand of the urgent non-TCLs
gain/loss due to the TCL (QTCL ). (PUNTCL ).
The ETP model is described mathematically by the following
state space function [29]: PUNTCL (t) = Plight (t) + Pmisc (t) (8)

ẋ = Ax + Bu, y = Cx + Du (1) where Plight and Pmisc are power demand from lighting and miscel-
    laneous appliance loads with hourly operational schedules.
T˙a Ta
ẋ = , x= , u=1 (2)
˙
Tm Tm 2.2.4. Non-urgent non-TCL model
Some loads are not required to start instantaneously if allowed
⎡ UUa
 U ⎤ ⎡ ⎤
m m by the end-user. For load such as a dishwasher or clothes drier, a
− + To Ua Qa
⎢ C C C ⎥ + user may not be sensitive to the start time of the load, but is likely
⎦ , B = ⎣ Ca Ca ⎦
a a a
A=⎣ (3) to be sensitive to the finish time. Non-urgent non-TCL loads can
Um −Um
0 be characterized according to the earliest time it can start (tstart ),
Cm Cm
  duration of its operation (t), power demand during the operation
1 0 0 (Prated ), and the latest time it needs to finish its task (tfinish ) [31].
C= , D= (4)
0 1 0 PNNTCL = f (Prated , t, tstart , tfinish ) (9)
To account the heat gains or losses due to ventilation and leak-
ages, the following term is added in the differential function for 2.3. Power supply limitation
Ta :
In the traditional set-up of controlling the balance between
˙ p (To − Ta )
mC (5) the electricity supply and demand, the generators have to adjust

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
4 R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx

simultaneously and follow the variability of the demand. However, The amount of solar irradiance (GT ) received by the solar panels
this is not applicable to DER which rely on uncontrollable inputs is calculated based on the commonly used isotropic model defined
such as wind speed and solar radiation. A shift from this paradigm is by the following Eq. [32]
therefore necessary – load demand should follow or adjust accord-    
ing the available power supply. To demonstrate the capability of 1 + cos(ˇ) 1 − cos(ˇ)
GT = Gb Rb + Gd + Gg (15)
flexible loads to absorb the grid disruptions caused by electrical 2 2
loads and DERs, this study proposes a maximum power limit that
where Gb is the beam component of the solar irradiance, Gd is the
the loads can consume at a particular time.
diffused component, G is the irradiation received by a horizontal
Plimit = Pbase + PDER (10) plane, Rb is the ratio between the radiation on a tilted plane and a
horizontal plane, ˇ is the tilt angle of the solar panel, and g is the
If the base power (Pbase ) produced by traditional generators is ground albedo or the reflectance coefficient caused by the ground.
allowed to be constant, any fluctuations in the output of DER (PDER ) The irradiation received by a horizontal plane (G) can be calcu-
would be reflected in the load demand limit (Plimit ). Consequently, lated using the sky clearness index (KT ), which is defined as the ratio
at instances when DER output is high, more load demand requests between the irradiation measured at the sea level, and the extrater-
shall be granted to consume electricity. Conversely during low DER restrial irradiation measured outside the earth’s atmosphere [32]
production, the load consumption would automatically adjust and
delay the start-up of some flexible loads. G = KT Go (16)
Electricity price can also be included as an incentive to reduce
demand. As the electricity market price increase, demand from flex- Go is the extraterrestrial irradiation defined by the following equa-
ible loads should fall in response. Under this approach the market tion:
price factor will only be applied to the base generation and not to Go = Gon cos(z ) (17)
DER which generally operate at lower costs. Thus, Eq. (10) can be
modified to: where Gon is the extraterrestrial irradiance on a plane normal to
the solar beam and  z is the zenith angle.
Plimit = Pbase Kp + PDER (11)

where Kp is the price factor defined as:

Gon = Gsc (1.000110 + 0.034221 cos(B) + 0.001280 sin(B)
A ∗ limit
Kp = (12) + 0.000719 cos(2B) + 0.000077 sin(2B)) (18)
market
where Gsc is the solar constant (1367 Watts/m2 ) while B is defined
where A is a predetermined constant, limit is the price threshold,
as:
and market is the market electricity price.
 360 
Another approach is to define a total cost limit based on the price
B = (n − 1) n = day of the year (1 ≤ n ≤ 365) (19)
threshold and the allocated power. 365

Climit = Pallocated ∗ limit (13) Gd is calculated using the Orgill and Hollands correlation [32]
and Gb is determined by Eq. (20)
Climit is the highest expected cost [$] i.e. when the price of the
electricity to be purchased is limit . Upon power allocation, the sum Gb = G − Gd (20)
of electricity cost (i.e. demand × price) should not exceed the limit
Climit . Thus, if current electricity price is beyond the price threshold, The calculation of the solar irradiance followed by this study can
only the loads of higher priority (i.e. lower flexibility) shall be served be summarized as follows:
while the loads of lower priority will be shed off or shifted. This
approach has been adopted in the simulation stage of this study. • Predict the extraterrestrial irradiance Go based on the date, time,
The management of loads during periods of low DER produc- and site parameters (i.e. latitude, longitude, altitude). Derive the
tion high demand, or high market prices is done automatically by necessary data about solar angles and solar time.
the central controller. Loads are prioritized so that some loads with • Derive the clearness index KT based on atmospheric conditions.
more flexibility will be delayed while loads of higher priority are This value is related to the cloud cover. According to [33], cloud
served first. The prioritization algorithm for different loads is dis- cover is influenced by humidity more than by temperature. Thus,
cussed in details in the next section. Meanwhile, the models used for this study, a regression model was created to derive KT as a
to forecast DER output in this study are discussed in the following function of humidity.
sub-section. • Derive clear sky irradiance on horizontal plane (i.e. G = KT Go ).
• Derive the diffused component Gd using Orgill and Hollands cor-
2.3.1. Solar power model relation.
• Calculate the beam component (i.e. Gb = G − Gd ).
The output of the distributed solar PV (i.e. behind-the-meter
• Calculate Rb (i.e. Rb = cos()/cos( z ).
PV), which are usually installed on the rooftop, are calculated as:
• Finally calculate the solar irradiance GT .
Pinstalled ∗ GT ∗ inverter
Psolar = (14)
Gstandard The complete details of calculating the necessary angles and
where Psolar is the predicted power output from the solar panels solar time conversion is discussed by [32].
[Watts], Pinstalled is the installed capacity of behind-the-meter PV
panels, GT is the global solar irradiance received by the solar pan- 2.3.2. Wind power model
els [Watts/m2 ], Gstandard is the standard Irradiance value used by The collective performance of several wind turbines depends
PV manufacturers to test their products (i.e. 1000 Watts/m2), and on the output of each individual turbine which is expressed by
inverter is the inverter efficiency. The efficiency of the PV array has its designed power curve. This direct estimate approach has been
already been accounted for in the rated capacity as it is based on proven to be more accurate and reliable than other methods such
the actual PV performance at the standard conditions. as curve fitting and regression [34].

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx 5

In this study, the model of an individual wind turbine shall

be used to predict the power output of a collection of behind-
the-meter wind turbines. Such model specifies the cut-in speed
at which the turbine starts to produce electricity, rated speed at
which the turbine produces the rated capacity, and cut-out speed
(i.e. maximum speed) at which the turbine has to stop its oper-
ation for safety purposes. The typical mathematical expression of
the wind power curve is [30]:
⎛ ⎞
0 0 ≤ v ≤ vi
⎜ ⎟
⎜ a + bv3 vi ≤ v ≤ vr ⎟
⎜
P(v) = Pcapacity ∗ ⎜ ⎟ (21)
⎟
⎝ 1 v r ≤ v ≤ vo ⎠
0 v > vr
Fig. 3. Flexibility of thermostatically controlled loads in cooling mode.
where v is the wind velocity, vi is the cut-in speed, vr is the rated
speed, and vo is the cut-out speed. The coefficients a and b are
defined as:
v3i 1
a= , b= (22)
v3i − v3r v3r − v3i
To account the variations wind effectiveness due to air density,
the average velocity should be adjusted [34]:
  1/3
v = vave (23)
o
where vave is the average wind speed,  is the current density of air
in the site, and o is the density of air based on the ISO standard
atmosphere (1.225 kg/m3 ).

Fig. 4. Flexibility of thermostatically controlled loads in heating mode.

3. Load flexibility, prioritization, and power allocation

3.1. Flexibility of each type of flexible load Similarly, the flexibility of a TCL in heating mode can be derived
as:
The flexible loads are prioritized based on their capacity to Ta − Tmin
accept delays in operation. In this context, the urgent non- f = (25)
(Ts + tolerance) − Tmin
thermostatically controlled loads that has direct interaction with
the end-user and needs to start immediately after the switch is where Tmin is the minimum temperature contracted with the
turned on would have zero flexibility. Therefore, such types of loads end-user during the agreement to participate in the DR program.
should have the highest priority and be served first. This is to enable Negative values of flexibility for TCLs in heating mode means that
the end-users to do their daily routines with minimal disruptions the internal temperature has gone lower than the contracted lower
from the DR program. temperature limit.
Other load types as discussed above, have their own degree of
flexibility which shall be defined and discussed in this section. 3.1.2. Flexibility of battery-based loads
Battery-based loads like electric vehicles and grid energy stor-
3.1.1. Flexibility of TCLs age are prioritized according to the urgency of charging completion.
TCLs are prioritized according to the temperature deviation from This can be expressed as:
the setpoint. Loads with higher values of deviation shall be served
(tuse − t) − (tf − t)
first. The common thermostat control of a TCL allows a certain f = (26)
tuse − t
deviation from the setpoint (i.e. deadband). The flexibility of TCLs
in cooling and heating mode are illustrated in Figs. 3 and 4. The where f is the flexibility, t is the current time, tf is the expected time
flexibility of a TCL in cooling mode can be calculated as: to finish charging if charging is started at current time considering
its current status, and tuse is the time when the unit shall be used
Tmax − Ta
f = (24) and should already be fully charged. This should be specified by the
Tmax − (Ts − tolerance)
end-user upon plugging in the charger, hence, the charging stations
where Ts is the setpoint temperature, Ta is the current temperature should be able to take this information from the customers.
inside the thermal zone, and Tmax is the maximum temperature In this study, battery-based loads are assumed to be interrupt-
for Ta which can be specified in the contract agreement with the ible so that whenever there is insufficient power, charging should
end-users. cease, consequently moving the predicted finish time closer to the
Based on the flexibility equation, there is a possibility that the latest finish time. Meanwhile, whenever excess power is available,
flexibility value would be negative. This happens when the tem- charging should start automatically. Fig. 5 illustrates the flexibility
perature inside the thermal zone has exceeded the contracted of the battery-based loads which is based on delayed charging.
maximum temperature limit Tmax , a scenario that should be This study does not include fully modulated EV charging or bi-
avoided. Hence, in the power allocation algorithm, the goal is keep directional vehicle-to-grid (V2G) charging, because the charging
the flexibility values from being negative. control technology necessary to achieve these functionalities is not

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
6 R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx

Table 1
Power allocation to flexible loads after prioritization.

Load ID Demand [Watts] Flexibility value Permission status

5 1061.1 0.287 ON
2 807.5 0.334 ON
3 907.3 0.368 ON
1 2340.2 0.472 ON
8 432.7 0.477 ON
9 540.2 0.489 ON
10 659.1 0.632 OFF
7 1354.3 0.682 OFF
6 120.8 0.723 ON
4 2340.8 0.782 OFF

Fig. 5. Charging flexibility of battery-based loads.

higher flexibility. Such loads could be TCLs that are still far from the
temperature limit, battery-based loads which still have sufficient
time to be fully charged before they are to be used, or non-urgent
loads that can be shifted to operate in off-peak times.
An example of load prioritization and power allocation is pre-
sented in Table 1. The loads are sorted according to their flexibility.
Power is allocated first to loads with lower flexibility values and
they are given the permission to turn on. Notice that load 6 is also
allowed to turn on even its flexibility value is higher than load 7 and
load 8. This happens because the power demand of load 6 is small
enough that it can still be served by the remaining available power
that the generators can supply. The algorithm scans the entire list
of loads until all of the available power has been allocated without
Fig. 6. Flexibility of non-urgent non-thermostatically controlled loads. exceeding the limit.
After the status of each load are determined, total electricity cost
commonly available at this time. To be reflective of current technol- limit is determined by the product of the price threshold and the
ogy offerings, only binary interruptible charging control is included sum of all load demands that were permitted to turn on. Such loads
in the model. The introduction of more sophisticated EV charging are rescanned and their costs (i.e. demand × price) are aggregated
controls would of course provide additional DR flexibility from a with the constraint of not exceeding the total electricity cost limit.
resource that contained significant EV deployment, though such a Loads that did not make it to the cut-off will be reassigned with
scenario is not modeled in this study. an off status, hence it is not permitted to run for the next time
increment.
3.1.3. Flexibility of non-urgent non-TCLs
The non-urgent non-TCL loads can be delayed, hence it offers
a flexibility space between the predicted finish time based on the 3.3. Simulation
duration of its operation and the required latest time to finish.
In this section, the viability of the proposed aggregation algo-
tL − tf
f = (27) rithm is demonstrated via simulation. Fig. 7 shows the proposed
tL − t
aggregation algorithm as implemented in Python. This study does
where tL is the latest finish time, tf is the predicted finish time, and t not simulate the individual models of loads which were already
is the current time. Here the predicted finish time can be calculated validated by other authors [29,32,34]. The focus of this study is to
as: validate the proposed aggregation methodology. Simulations are
conducted on a set of buildings with varying load characteristics.
tf = t + tduration (28)
A hypothetical population of HVACs, refrigerators, freezers,
Similar to the flexibility function of the battery-based loads the water heaters, EVs, energy storage, and non-TCL loads were gener-
flexibility of the non-urgent non-TCLs already take into account ated using a Gaussian distribution of parameters (e.g. area, volume,
the user’s requirement for the operation to finish. It is important to insulation, Ua , Um , Ca , Cm , etc.). Then, the time varying parameters
note that unlike battery-based loads NNTCLs cannot be interrupted were initialized. The parameters are then updated using the cor-
once started. The flexibility value may reach a negative value which responding weather conditions of that particular time step. With
means that the expected finish time is already beyond the required such parameters the status of the loads were analyzed and their
latest time to finish. Thus, the constraint in this algorithm is to avoid flexibilities are calculated. The flexibilities of all the loads are then
reaching negative values for the flexibility (Fig. 6). ranked together during the prioritization process. Loads with lower
flexibility are then served first during the power allocation process.
3.2. Power allocation The allocated power is maximized to serve as much load requests
as possible while following the constraint that it should not exceed
After calculating the flexibility of the loads, they are prioritized the power supply limit. During this process the aggregated demand
according to their flexibility value. The available power is then allo- response is calculated as the total load demand minus the total
cated to the loads starting from those with lower flexibility. The allocated power. Next, the behaviour of the loads is emulated by
power allocation is constrained to the limit of the available power inputting the current parameters to their physical models. The
which is subject to the DER output and the market price. Hence, resulting values of the variables being monitored (i.e. temperature,
any instantaneous fluctuations caused by the DER is automatically charge level, job status) are recorded while the real-time data are
absorbed by the aggregated loads by shedding off the demands with updated. The algorithm proceeds to the next time step and repeat

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx 7

Fig. 7. Proposed algorithm for aggregation of flexible loads.

Table 2 Table 3
Range of values for the parameters of the HVAC models. Range of values for the parameters of the refrigerator models.

Parameter Lower Upper Average Parameter Lower Upper Average

Area [m2 ] 139 325 186 Volume [m3 ] 0.324 0.72 0.5
Ua [W/◦ C] 0.85*Area 1.98*Area 211 Ua [W/◦ C] 2.27 * Volume
Um [W/◦ C] 0.85*Area 1.98*Area 211 Um [W/◦ C] 66.67* Volume
Ca [J/◦ C] 3025*Area 3025*Area 3025*Area Ca [J/◦ C] 1210 * Volume * Air density
Cm [J/◦ C] 900*Area 900*Area 900*Area Cm [J/◦ C] 1500 * Volume
Tsetpoint [◦ C] 23.9 24.4 23.9 Tsetpoint [◦ C] 1.7 4.0 2.8
Tolerance [◦ C] 0.28 1.1 0.56 Tolerance [◦ C] 0.28 1.1 0.56
Tmax [◦ C] 26 28 27 Tmax [◦ C] 8 15 12
Tmin [◦ C] 15 17 16 Cooling [W] 350 * Volume
Heating [W] 56.75*Area 56.75*Area 56.75*Area
Cooling [W] 56.78*Area 113.56*Area 14,064
Table 4
Range of values for the parameters of the freezer models.

the process until it reaches the last time step specified at the begin- Parameter Lower Upper Average
ning of the simulation. Volume [m ] 3
0.24 0.72 0.3
The weather database used for this simulation is the Typical Ua [W/◦ C] 2.27 * Volume
Meteorological Year 3 (TMY3) developed by National Renewable Um [W/◦ C] 66.67 * Volume
Energy Laboratory. Meanwhile, the typical load demand for a rep- Ca [J/◦ C] 1210 * Volume * Air density
Cm [J/◦ C] 1500 * Volume
resentative building in Los Angeles, California, USA is based on the Tsetpoint [◦ C] 23.0 18.0 21.0
datasets compiled by openei.org that is associated with the weather Tolerance [◦ C] 0.28 1.1 0.56
database TMY3. Tmax [◦ C] 0 4 2
Cooling [W] 350 * Volume
Cp [J/kg ◦ C] 1005
3.3.1. Assumptions for TCLs
Thermal parameter values used for the ETP models of different
TCLs were taken from different publications, or assumed [10,35,36]. for in a constant value of air mass flow (i.e. 0.000001 kg/s). Since the
The assumed parameters for HVAC units are shown in Table 2. ambient temperature as well as the temperature inside the units are
The heat gains/losses due to ventilation and leakages in the acceptably stable, the specific heat capacity CP is assumed constant
building are taken into account by assuming that the air change (i.e., 1005 J/kg ◦ C). Meanwhile, the coefficient of heat transfer and
per hour (CPH) is 2.0. To calculate the air mass flow, a regression the heat capacity of air and food materials are assumed to be the
model for air specific heat capacity (Cp ) and air density were created same for both refrigerators and freezers.
as a function of the outside air temperature. Though there are some entries in Tables 3–5 with only one
For refrigerators and freezers, it is assumed that the units are value for Lower, Upper, and Average values, they are commonly
located inside the building, hence, the ambient temperature is the multiplied by a certain factor e.g. volume, thus, the actual val-
air temperature simulated by the HVAC units. The air changes due ues still vary according to the Gaussian distribution. Tolerance
to the opening of the refrigerator and freezer door are accounted values set the limit above or below the temperature setpoint. In

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
8 R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx

Table 5 Table 6
Range of values for the parameters of the water heater models. Range of values for the parameters for electric vehicles.

Parameter Lower Upper Average Parameter Lower Upper Average

Volume [m3 ] 0.0075 0.03 0.0225 Capacity [kWh] 5.2 85 24

Uw [W/◦ C] 75 * Volume Charging rate [kW] 3.3 20 3.3
Cw [J/◦ C] 4186 * Volume * Water density Charging efficiency [%] 0.75 0.9 0.8
Tsetpoint [◦ C] 50.0 65.0 60.0 Charging start [h] 17 21 19.5
Tolerance [◦ C] 0.28 1.1 0.56 Latest finish [h] 5 7 6
Tmin [◦ C] 35 45 40
Heating [W] 200,000 * Volume
Flow [kg/s] 0.0003 * Water Usage Table 7
Cp [J/kg ◦ C] 4186 Range of values for the parameters of energy storage.

Parameter Lower Upper Average

Capacity [kWh] 7 90 10
Charging rate [kW] 2.0 3.3 2.0
Charging efficiency [%] 0.85 0.95 0.92

Fig. 8. Graphical representation of regression functions for hot water usage.

that case, the operating deadband of TCLs in this study is from

(Tsetpoint − Tolerance) to (Tsetpoint + Tolerance). For water heaters, the
specific heat capacity of water is assumed to be constant (i.e.,
4186 J/kg ◦ C). The effect of the internal mass is neglected in the ETP Fig. 9. Hourly operating schedules of urgent non-TCL loads.

model for water heaters.

The mass flow variations caused by usage of hot water will have The status of EV batteries are also simulated based on the typical
an impact on the performance of water heaters. In this study, a driving schedule of the users, i.e. they leave the house at around
regression model for water usage has been created based on the 05:00 to 07:00 and return home at around 17:00 to 21:00. It is
water heater demand from the database for typical residential assumed that at the start of charging (i.e. when the EV is plugged in),
loads. The regression model is a function of the time of the day, the users would specify the required latest time to finish charging.
i.e. 0–24. To attain a better accuracy (i.e. for per minute values), the On the other hand the battery-based energy storage parameters are
total time horizon is divided into six parts and a regression model shown in Table 7. The discharging rate is assumed to be equal to
is created for each part. the charging rate.
⎛ ⎞ Energy storage devices are not bounded by a particular sched-
0.019 0≤t≤4
ule for charging start and the latest time to finish charging. Instead,
⎜ 0.6(t − 4) + 0.25 4<t≤8 ⎟
⎜ ⎟ they operate according to the balance between the power supply
⎜ −0.1(t − 8) + 0.99 8 < t ≤ 12 ⎟
Fm = 0.0003 ∗ ⎜
⎜ −0.01(t − 12) + 0.44
⎟ (29) and demand. If demand exceeds supply, the batteries will discharge
⎜ 12 < t ≤ 16 ⎟
⎟ to help in supplying power. Conversely, when demand is lesser than
⎝ 0.06(t − 16) + 0.14 16 < t ≤ 20 ⎠ supply, the battery will switch to charging mode. It is important
−0.14(t − 20) + 0.8 20 < t ≤ 24 to note that in discharging mode, the state of charge should stay
above 20%. Below this limit, the discharging should stop to mini-
where Fm is the mass flow, and t is the time in hours from 0 to 24 mize battery degradation [37]. Flexibility values are fixed at 9999
(see Fig. 8). Random time delay (i.e. +30 min) is introduced on each just to emphasize that in load demand perspective, they are of least
unit to mimic the realistic variation of load schedule. priority. In addition, it is assumed that 50% of the buildings have
battery-based energy storage device.
3.3.2. Assumptions for battery-based loads
The first assumption made for battery-based loads is that their 3.3.3. Assumptions for urgent non-TCL loads
charging process can be interrupted. Other assumptions for the dif- Urgent non-TCL loads considered in this study include the light-
ferent parameters of the battery model such as the capacity, and ing and miscellaneous loads (i.e. appliances for cooking, cleaning,
charging rate are based on actual values of existing EVs in the mar- entertainment, etc.). Regression models were created with similar
ket. The types or brands of units which are commonly bought by approached used to model the hot water usage. The total time hori-
customers are also considered as they depict the realistic average zon is divided into four parts and regression models for each part
battery capacity and charging rate of EVs that connects to the grid. were developed. The typical hourly operating schedules are pre-
The following range of parameter values were derived for electric sented in Fig. 9 to predict the power demand of lighting loads and
vehicles, as presented in Table 6. appliance loads.

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx 9

Fig. 10. Flexible loads’ aggregated demand and renewable generation on a summer day.

where Plight is the power demand by lighting loads [kW], Pmisc 2. Getting the baseline load demand. In this simulation, EVs and
is the power demand by appliance loads [kW], and t is the time of energy storage devices were added to the total load curve.
the day [h]. The total urgent load demand is the sum of Plight and Two baseline load curves were established: one curve includes
Pmisc . These loads are also assumed to be converted into heat which EVs but excludes the energy storage devices; the second curve
contribute to the internal heat gain that should be included in the includes both the EVs and battery-based energy storage devices.
HVAC model. This is to check if energy storage devices are necessary or the EVs
With regarding to the simulation sequence, the following steps are already sufficient to provide more flexibility to the grid.
of action were performed to model difference grid scenarios. 3. Load shedding simulation: The grid capacity which was derived
in simulation 1 was used as the power limit for the baseline curve
in simulation 2. The maximum increase or decrease of power
1. Calculation for the existing grid power capacity. The power supply at each time step (i.e. 1 min) was determined by adding
capacity of the existing grid was calculated by simulating a set the maximum ramping rate of the existing generators and the
of hypothetical loads without electric vehicles and energy stor- DER forecast at such period. It was assumed that the traditional
age devices. This is to check the viability of the algorithm to generators have a collective ramping rate of 1% of its capacity
provide DR and delay the necessary infrastructure upgrades and per minute [38]. The upper limit of the entire simulation period
construction of additional power plants while EVs and energy (i.e. 24 h) is the maximum capacity of the traditional generators
storage devices are integrated into the existing grid. The system plus the maximum DER output at a particular time. The total
capacity is determined by the following equation. The typical load demand should not exceed the power limit, hence, some
diversity factor (i.e. 3) for residential sector was adopted in this loads are automatically shifted by the control algorithm during
simulation. the power limiting hours.
 4. Response to DER fluctuations. A sudden drop of DER output was
(Individual peak loads) simulated by reducing the DER electricity production by 80%. The
System capacity = (30)
Diversity factor algorithm automatically reduces the load demand to maintain

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
10 R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx

Fig. 11. Load shedding during peak hours.

the power balance. It was assumed that 50% of the buildings has (i.e. green) overnight. The high cooling demand is logical since the
installed a rooftop solar PV system, and the roof area used for outside air temperature (OAT) is high. No other load exhibits and
solar PV varies from 25% to 70% of building’s floor area. On the obvious diurnal pattern. As shown in Fig. 10, market energy prices
other hand, the installed wind capacity was assumed to be 10% are presented to quantify each type of flexible load’ response to
of the grid capacity. higher energy price. Market prices will be relevant to scenarios
5. Response to market fluctuations. A hypothetical variation of presented in subsequent sections.
market price was simulated and the threshold price was chosen
to be $0.15 per kWh. The demand automatically reduces when
the price exceeds the threshold price level. 4.1. Scenario 1: system peak mitigation
6. Impact of the energy storage devices. It was assumed that each
building has an energy storage device that can help in managing The simulation results presented in Fig. 11 illustrate the
the grid. State of charge of the batteries were not allowed to response of the algorithm to mitigate large demand spike which
discharge below 20% of its capacity. This is to consider the safety occurs between approximately 18:00 and 23:00 in the base case
operation of the batteries and minimize degradation. scenario. To accomplish this, the controller relies predominantly on
shifting EV charging, while also utilizing flexibility in refrigeration
4. Results loads and NNTCLS. The magnitude of each type of flexible load con-
tribution is determined by the load prioritization algorithm. The
For the simulation results presented in this section battery- power allocation algorithm prevents the load demand to exceed
based energy storage technologies are not considered. Additionally, the available power supply. It can also be observed that there is
the electric vehicles act solely as a load and do not contribute in no rebound of power demand after the peak hours for TCL loads,
supplying power to the grid. This is done to emphasize the ability which is commonly seen after a DR event in other DR programs (e.g.
of the aggregation methodology in maintaining the power balance Time-of-Use).
without relying on dedicated, dispatchable resources like energy With such demand response, the performance of each load is
storage devices. slightly affected by participating in DR, but does not exceed the
To provide additional context for how the controller will uti- limits of its specified operation conditions, particularly tempera-
lize and influence the aggregate demand Fig. 10 shows the original ture for TCLs and completion status for non-TCLs. Fig. 12 shows the
load demand, i.e., before the proposed algorithm was applied. It also operating temperatures of HVACs, refrigerators, water heaters, and
shows the aggregated load demand as typically seen from the sub- freezers during the DR event. As shown in Fig. 12 the TCL temper-
station or aggregation point. Each demand load-type is summed atures are within the limits of the normal operation. The charge of
and grouped by color. As expected, the dominant loads are cool- the electric vehicles and the job status of the NNTCLs are also done
ing (i.e. blue) during daytime hours and electric vehicle charging before the required time as shown in Fig. 13.

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx 11

Fig. 12. Status of flexible loads to DER output disruption and load shedding. Fig. 13. EV and NNTCLs’ status in response to system peaks and power limitations.

4.2. Scenario 2: DER disruptions generations. In this scenario, the controller is also reducing the sys-
tem peaks caused by EV charging during evening hours. It can be
Fig. 14 illustrates how the controller is able to adjust the total seen that HVACs, freezers, and refrigerators provide all the load
load demand instantaneously in response to DER fluctuations. In shed capacity.
this figure, PV output within the system decreased by 80% for 1 h
(i.e. from 11:30 to 12:30). During such an event, traditional gen- 4.3. Scenario 3: market price fluctuations
erators will ramp up to meet the newly increased net demand,
however it is not capable of providing sufficient power instanta- Finally, this scenario introduces fluctuations in the market price
neously. Under these conditions, the ability of the controller to for energy, as illustrated in Fig. 10. The results from the simulated
provide fast DR is highly valuable. Fig. 14 demonstrates how the controller for this scenario are given in Fig. 15. A threshold price
controller utilized each load-type to respond the loss of renewable has been set at 0.15 $/kWh. When the market price exceeds the

Fig. 14. Response of flexible loads to DER output disruption and load shedding.

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
12 R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx

Fig. 15. Adjustment of the load demand based on the electricity price.

threshold price, the controller responds by reducing consumption In comparison with previous work, we create a suite of bottom-
similarly as when under a DR event, while allowing consumption up physical models of each energy system component for modeling
to increase when market prices remain below the threshold. The the DR potential from residential sector. Specifically, those models
high-price periods (i.e. from 13:00 to 15:00 and from 17:00 to are grounded in the physical constraints for thermostatically con-
18:00) are highlighted in the figure. As with scenarios 1 and 2, the trolled loads (i.e. HVAC, refrigerators, electric water heaters), EV
overall demand is still subject to the limitation of power supply. and storage, and distributed PV and wind generations. In addition,
Meanwhile, the load shed capacities from EVs, NNTCL, fridges, and the proposed power allocation algorithm for providing demand
freezers are deployed to prevent the aggregated load demand to response resources is unique base on each energy system compo-
exceed the supply limit. The operating conditions of the loads are nent’s load flexibility. Lastly, it is expected to provide an accurate
slightly affected but within prescribed limits. prediction of DR potential for day-ahead planning simulations in a
large scale due to the sufficiently low computational intensity. With
5. Discussion, conclusion and future work regarding to the limitation of this framework, we need to know a lot
about the composition of the feeder system to build up the model.
This paper provides the details of the development of an In addition, we also need to collect a lot of the equipment stock
algorithm for aggregation of flexible loads for demand response and conditions to build model for each system component under
application at the substation level. Its viability has been demon- the feeder system, because those models vary a lot by a number of
strated by simulations of the different flexible load models parameters, such as location, weather, building stocks, EV driving
aggregated by the proposed algorithm. Results show that the algo- and storage usage behaviors. In addition to those uncertainties in
rithm can effectively mitigate system demand peaks with minimal model buildups, the computational intensity may be too high to
compromise in the load operations. It is also observed that sudden generate for higher granularity or real-time simulation in a large
spikes on the load demand (rebound effect) commonly observed scale.
in other DR programs such as Time-Of-Use are avoided in this Further studies of this methodology may include the improve-
methodology. The premise is that the algorithm should run in the ment of the models of different flexible loads and optimization
background during normal operations in the same manner as spin- algorithm to close the gap between the supply and demand with-
ning reserves operate. The algorithm can react instantaneously and out the help of energy storage devices. In actual implementation
can provide fast demand response or auto-DR to absorb grid dis- of this technology, models of the flexible loads may no longer be
ruptions caused by fluctuations in DER output or market price necessary. Instead of simulating the models, the loads may directly
of electricity. With all these potential advantages, the proposed send its current status to the central controller periodically. It is
methodology proves to be a good candidate to assume the role also recommended to implement the same algorithm in grid sim-
of ancillary services and assist in accelerating the adoption of the ulation software to verify its viability when electrical behaviors of
intermittent renewable energy resources. the grid is taken into account.

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018
G Model
EPSR-4738; No. of Pages 13 ARTICLE IN PRESS
R. Tulabing et al. / Electric Power Systems Research xxx (2016) xxx–xxx 13

Development and analysis of business models based on this type [17] S. Han, S. Han, K. Sezaki, Estimation of achievable power capacity from plug-in
of aggregation model and utility-customer interactions would also electric vehicles for V2G frequency regulation: case studies for market partic-
ipation, IEEE Trans. Smart Grid 2 (4) (2011) 632–641.
be of interest. This might include investigations of the necessary [18] E.C. Kara, J.S. Macdonald, D. Black, M. Bérges, G. Hug, S. Kiliccote, Estimating
market policies, technological requirements, social promotions, the benefits of electric vehicle smart charging at non-residential locations: a
and financial aspects needed to support and operate such a model. data-driven approach, Appl. Energy 155 (2015) 515–525.
[19] H. Lund, W. Kempton, Integration of renewable energy into the transport and
electricity sectors through V2G, Energy Policy 36 (9) (2008) 3578–3587.
Acknowledgements [20] D. Dallinger, S. Gerda, M. Wietschel, Integration of intermittent renewable
power supply using grid-connected vehicles – a 2030 case study for California
and Germany, Appl. Energy 104 (2013) 666–682.
The work described in this study was coordinated by the Grid [21] C. Goebel, D.S. Callaway, Using ICT-controlled plug-in electric vehicles to sup-
Integration Group of Lawrence Berkeley National Laboratory and ply grid regulation in California at different renewable integration levels, IEEE
was supported by the State Grid Corporation of China Project Trans. Smart Grid 4 (2013, June) 729–740.
[22] J.L. Mathieu, M. Dyson, D.S. Callaway, Using residential electric loads for fast
(DZN17201300197, Study on Key Technologies for Power and Fre-
demand response: the potential resource and revenues, the costs, and policy
quency Control of System with ‘Source-Grid-Load’ Interactions). recommendations, in: Proceedings of the ACEEE Summer Study on Buildings,
2012, pp. 189–203.
[23] J.L. Mathieu, M. Kamgarpour, J. Lygeros, D.S. Callaway, Energy arbitrage with
References
thermostatically controlled loads, in: European Control Conference, 2013, pp.
2519–2526.
[1] Ahmad Faruqui, Ryan Hledik, S.S. George, J. Bode, P. Mangasarian, I. Rohmund, [24] J.L. Mathieu, M.E. Dyson, D.S. Callaway, Resource and revenue potential of
G. Wikler, D. Ghosh, S. Yoshida, A National Assessment of Demand Response California residential load participation in ancillary services, Energy Policy 80
Potential. Tech. Rep., Federal Energy Regulatory Commission, 2009. (2015) 76–87.
[2] S. Kiliccote, M.A. Piette, J. Mathieu, K. Parrish, Findings from seven years of field [25] H. Hao, Y. Lin, A.S. Kowli, P. Barooah, S. Meyn, Ancillary service to the grid
performance data for automated demand response in commercial buildings, in: through control of fans in commercial building HVAC systems, IEEE Trans.
2010 ACEEE Summer Study on Energy Efficiency in Buildings, 2010. Smart Grid 5 (4) (2014) 2066–2074.
[3] M.A. Piette, D. Watson, N. Motegi, S. Kiliccote, P. Xu, D. Watson, Automated [26] E.C. Kara, M.D. Tabone, J.S. MacDonald, D.S. Callaway, S. Kiliccote, Quantifying
Critical Peak Pricing Field Tests: Program Description and Results. Tech. Rep. flexibility of residential thermostatically controlled loads for demand response:
August, Lawrence Berkeley National Laboratory, Berkeley, CA, 2007. a data-driven approach, in: Proceedings of the 1st ACM Conference on Embed-
[4] P. Cappers, J. MacDonald, C. Goldman, O. Ma, An assessment of market and pol- ded Systems for Energy-Efficient Buildings, 2014, pp. 140–147.
icy barriers for demand response providing ancillary services in U.S. electricity [27] H. Hao, B.M. Sanandaji, K. Poolla, T.L. Vincent, Aggregate flexibility of thermo-
markets, Energy Policy 62 (March) (2013) 1031–1039. statically controlled loads, IEEE Trans. Power Syst. 30 (1) (2015) 189–198.
[5] O. Ma, N. Alkadi, P. Cappers, P. Denholm, J. Dudley, S. Goli, M. Hummon, S. [28] E.C. Kara, M. Berges, G. Hug, Impact of disturbances on modeling of thermo-
Kiliccote, J. MacDonald, N. Matson, D. Olsen, C. Rose, M.D. Sohn, M. Starke, B. statically controlled loads for demand response, IEEE Trans. Smart Grid 6 (5)
Kirby, M. O’Malley, Demand response for ancillary services, IEEE Trans. Smart (2015) 2560–2568.
Grid 4 (4) (2013) 1988–1995. [29] Z.T. Taylor, K. Gowri, S. Katipamula, GridLAB-D Technical Support Document:
[6] G. Heffner, C. Goldman, B. Kirby, Loads Providing Ancillary Services: Review of Residential End-Use Module Version 1.0. Tech. Rep., Pacific Northwest National
International Experience, Lawrence Berkeley National Laboratory. Tech. Rep., Laboratory, USA, 2008.
Berkeley, CA, 2007, May. [30] A. Ghasemi, S.S. Mortazavi, E. Mashhour, Hourly demand response and bat-
[7] D.J. Olsen, N. Matson, M.D. Sohn, C. Rose, J. Dudley, S. Goli, S. Kiliccote, M. Hum- tery energy storage for imbalance reduction of smart distribution company
mon, D. Palchak, J. Jorgeson, P. Denholm, S. Ma, Grid Integration of Aggregated embedded with electric vehicles and wind farms, Renew. Energy 85 (2016)
Demand Response, Part 1: Load Availability Profiles and Constraints for the 124–136.
Western Interconnection. Tech. Rep., Lawrence Berkeley National Laboratory, [31] X. Geng, P.P. Khargonekar, Electric vehicles as flexible loads: algorithms to opti-
Berkeley, CA, 2013, September. mize aggregate behavior, in: 2012 IEEE 3rd International Conference on Smart
[8] B.J. Kirby, Frequency Regulation Basics and Trends, Oak Ridge National Labora- Grid Communications, SmartGridComm 2012, 2012, pp. 430–435.
tory. Tech. Rep., 2004, December. [32] J. Duffie, W. Beckman, Solar Engineering of Thermal Processes, vol. 42, 4th ed.,
[9] D.S. Watson, M.A. Piette, O. Sezgen, N. Motegi, Machine to machine (M2M) tech- John Wiley and Sons, 2013.
nology in demand responsive commercial buildings, in: 2004 ACEEE Summer [33] A.M. Tompkins, Impact of temperature and humidity variability on cloud cover
Study on Energy Efficiency in Buildings, Pacific Grove, CA, 2004. assessed using aircraft data, Q. J. R. Meteorol. Soc. 129 (2003) 2151–2170.
[10] S. Katipamula, N. Lu, Evaluation of residential HVAC control strategies for [34] Y.-H. Wan, E. Ela, K. Orwig, Development of an equivalent wind plant power-
demand response programs. Part 1, ASHRAE Trans. 112 (2006) 535–546. curve, in: Proceedings of the WindPower 2010, Dallas, Texas May 23–26,
[11] S. David, M. Ann, D.S. Watson, S. Kiliccote, N. Motegi, M.A. Piette, Strategies for 2010.
demand response in commercial buildings, in: 2006 ACEEE Summer Study on [35] R. Pratt, C. Conner, B. Cooke, E. Richman, Metered end-use consumption and
Energy Efficiency in Buildings, 2006, pp. 287–299. load shapes from the ELCAP residential sample of existing homes in the Pacific
[12] M.A. Piette, D.S. Watson, N. Motegi, N. Bourassa, Findings From the 2004 Fully Northwest, Energy Build. 19 (3) (1993) 179–193.
Automated Demand Response Tests in Large Facilities. Tech. Rep., Lawrence [36] W. Sandusky, E. Pearson, N. Miller, R. Crowder, G. Parker, R. Mazzucchi, G.
Berkeley National Laboratory, Berkeley, CA, 2005, October. Stokes, J. Thomas, R. Pratt, G. Schuster, M. Halverson, J. Stoops, F. Peterson,
[13] R. Yin, X. Peng, K. Sila, M.A. Piette, Study on auto-DR and pre-cooling of commer- R. Gillman, R. Stokes, S. Hauser, ELCAP operational experience, Energy Build.
cial buildings with thermal mass in California, Energy Build. (2010) 965–975. 19 (3) (1993) 167–178.
[14] W. Zhang, J. Lian, C.-Y. Chang, K. Kalsi, Aggregated modeling and control of air [37] S. Habib, M. Kamran, U. Rashid, Impact analysis of vehicle-to-grid technology
conditioning loads for demand response, IEEE Trans. Power Syst. 28 (4) (2013) and charging strategies of electric vehicles on distribution networks – a review,
4655–4664. J. Power Sources 277 (2015) 205–214.
[15] P. Denholm, J. Jorgenson, T. Jenkin, D. Palchak, B. Kirby, M.O. Malley, The Value [38] B. Kirby, M. Milligan, A method and case study for estimating the ramping
of Energy Storage for Grid Applications. Tech. Rep., National Renewable Energy capability of a control area or balancing authority and implications for moder-
Laboratory, USA, 2013, May. ate or high wind penetration, in: Proceedings of the WindPower 2005, 2005,
[16] J. Lopes, F.J. Soares, P. Almeida, Integration of electric vehicles in the electric pp. 1–16.
power system, Proc. IEEE 99 (1) (2011) 168–183.

Please cite this article in press as: R. Tulabing, et al., Modeling study on flexible load’s demand response potentials for providing ancillary
services at the substation level, Electr. Power Syst. Res. (2016), http://dx.doi.org/10.1016/j.epsr.2016.06.018