Applied Energy 224 (2018) 659–670

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

A bi-level scheduling model for virtual power plants with aggregated T

thermostatically controlled loads and renewable energy
⁎
Congying Wei, Jian Xu , Siyang Liao, Yuanzhang Sun, Yibo Jiang, Deping Ke, Zhen Zhang,
Jing Wang
The School of Electrical Engineering, Wuhan University, Wuhan 430072, China

H I GH L IG H T S

• Thermostatically controlled loads are scheduled to decrease imbalanced power of virtual power plant.
• Proportion static and dynamic aggregation not affected by parameters heterogeneity are proposed.
• AThebi-level optimal scheduling model and its two-step simplified algorithm are established.
• performance is verified under different regulating ranges, heterogeneity and forecast errors.

A R T I C LE I N FO A B S T R A C T

Keywords: With the penetration of renewable energy increasing, the power system requires higher flexibility of power
Virtual power plant regulation. Virtual power plants can aggregate distributed flexible loads to improve the utilization of distributed
Distributed energy resource renewable energy. In this paper, a bi-level scheduling model for virtual power plants with a large number of
Thermostatically controlled loads distributed thermostatically controlled loads and intermittent renewable energy is established to reduce the net
A bi-level scheduling
exchange power deviation caused by the forecast error of renewable energy. The upper level optimizes the
Forecast error
exchange power curve and reduces the imbalance costs in intraday, while the lower level tracks the optimized
power curve in real-time to complete the regulation target. Static and dynamic aggregation method reflecting the
regulation characteristics of aggregated thermostatically controlled loads is proposed and applied in lower/
upper level, respectively. In addition, a two-step simplified strategy is proposed to solve the mixed integer
nonlinear programming in upper level. Simulation results show that the proposed method can reduce the
maximum imbalance power, and it is not affected by parameters heterogeneity, which is suitable for virtual
power plants with diversified users.

1. Introduction other technologies, VPP aggregates distributed energy resources (DER)

including distributed generations (DG), energy storage systems (ESS)
With the rapid depletion of fossil fuels, developing renewable en- and flexible loads (FL), and then take part in the operation of the power
ergy sources (RES) such as photovoltaic generation (PV) and wind system as an agent [4]. VPP can coordinate each regional controllable
power has become a trend of energy revolution. However, the un- object to utilize distributed renewable energy locally, or provide an-
certainty of RES output may lead to power imbalance and increase the cillary services for centralized renewable energy power plants.
power regulating burden [1]. RESs are integrated into power system
either in a centralized way or a distributed way. In the centralized way, 1.1. Optimal operation for VPP
the stochastic model is often used to model and schedule the large-scale
power plants [2,3]; while in the distributed way, with the character- Nowadays, VPP has been studied in many literatures, and the RESs
istics such as small capacity, large number and wide distribution region, are always contained to improve the economic efficiency and en-
distributed RESs are difficult to be directly dispatched one by one. vironmental benefit.
Therefore, virtual power plant (VPP) is proposed to coordinate these Some focus on the cooperation between various DERs. Ref. [5] of-
distributed RESs. With advanced measurement, communication and fering an optimal operation for VPP with RESs, ESSs and conventional

⁎
Corresponding author.
E-mail address: [email protected] (J. Xu).

https://doi.org/10.1016/j.apenergy.2018.05.032
Received 31 October 2017; Received in revised form 5 April 2018; Accepted 5 May 2018
0306-2619/ © 2018 Elsevier Ltd. All rights reserved.
C. Wei et al. Applied Energy 224 (2018) 659–670

Nomenclature Tcin,min minimum allowable indoor temperature

Tcin,max maximum allowable indoor temperature
Matrix Tcout
,t outdoor ambient temperature
Tcdb temperature dead band
Ptstatic static power regulation characteristic of cluster TCLs uc∗,t on/off boundary of an individual HVAC
Pdynamic
c,t dynamic power regulation characteristic of an individual ρeup ,ρedown upward/downward imbalance energy price
TCL ρcup ,ρcdown upward/downward imbalance capacity price
Ptdynamic dynamic power regulation characteristic of cluster TCLs σimax max
,t ,δ j,t maximum length of each longitudinal/transverse seg-
Ut sequence of the threshold value uc∗,t ments
C set of the users in VPP
I set of Ii,j,t in each intraday optimization Variables

Indices Ii,j,t binary lookup table auxiliary 0–1 variable

SUt accumulative proportion regulation signal
c user index, c = 1,2,…,Ncus Pttarget target power of real-time power tracking
i longitudinal index in the interpolation, i = 1,2,…, Ntes1 PcHVAC
,t electric power of an individual HVAC
j transverse index in the interpolation, j = 1,2,…, Ntes2 PtTCL total power demand of TCLs
k iteration times of Step 2, k = 0,1,…, kmax Ptnet net exchange power between VPP and the power gird
t, τ time index, t = 1,2,…,T, τ ≤ t Tcin,t user’s room temperature
Tcset
,t indoor temperature set point
Parameters ΔTc,t temperature set point change
ut proportion regulation signal
kmax maximum iteration times of Step 2 Zc,t on/off status of HVAC system
K i,Xj,t , K iY,j,t slope of each longitudinal/transverse segments σi,j,t ,δi,j,t longitudinal/transverse segments variable.
KK i,XY j,t second order partial derivative in lookup table
Ncus number of customers in VPP ETP Parameters
Ntes1,Ntes2 number of discrete points of dynamic aggregation
Pcrate rated power of an individual HVAC P rate typical rated power of HVACs
Ptbasic basic loads excluding TCLs COP typical efficiency of HVACs
ptnet ,ahead VPP day-ahead planned trading curve R typical thermal resistance
PtPV ,intra total intraday forecast power of PVs C typical heat capacity

power plants in both day-ahead and balance markets. Refs. [6–8] been launched.
conducts day-ahead scheduling for an electricity-thermodynamics First, there are kinds of accurate models for an individual thermo-
coupled VPP. Ref. [9] proposes a forecast method for RESs based on statically controlled system: For example, Ref. [16] uses the thermal mass
meteorological data to guide the output of other units. Ref. [10] focuses of a building to defer power consumption from electric space heating;
on industrial VPP and tests its performance under different kinds of Ref. [17] establishes a high-order differential thermostatically con-
demand response (DR) programs. Ref. [11] identifies the operation of trolled model for large buildings; Ref. [18] deems TCLs as virtual ESS to
price-based DR in VPP day-ahead dispatch, and the incentive-based DR help micro-grids to consume distributed renewable energy; Ref. [19]
is further applied in intraday. Ref. [12] takes the coordinated operation presents a practical case of a modern non-residential building with
of electricity and natural gas networks base on VPP. controllable heating, ventilation and air-conditioning system (HVAC).
On the other hand, the bidding strategies of commercial VPP in Refs. [16–19] offer a HVAC model for a building. But for VPP consisting
electricity market are also discussed. Ref. [13] provides an optimal of residential TCLs with small capacity, the contribution that an in-
bidding strategy including the VPP benefit maximization and the day- dividual user can make to power system is minimal. Therefore, a load
ahead market clearing. Ref. [14] provides a non-equilibrium bidding aggregation program should be applied to make total capacity large
strategy for VPP based on deterministic price-based unit commitment. enough, so that their effect of DR can be fully developed.
Ref. [15] proposes a profit allocation framework with less calculation Second, based on this motivation, some research about the aggregation of
work. distributed TCLs has been carried out. For example, Ref. [20] proposes a
Most of the above literatures make use of flexible loads to deal with framework where TCLs provides ancillary services to power grid under
the forecast error of RESs [7–15], and some further consider VPP pro- direct load control. Ref. [21] builds models for HVACs in EnergyPlus,
viding multiple services [5,7,11,14,15]. Luckily, relatively complete and calculates their aggregated power regulation capacity. Ref. [22]
theories about VPP day-ahead optimization has been developed, but constructs the power density function for both the active and inactive
they need to be further corrected by intraday rolling optimization and machine-states of TCLs in real-time. Ref. [23] describes the state of air
verified by real-time power tracking. Besides, their work is based on the conditioners by means of the temperature priority list. Based on the
premises that the aggregated regulation characteristic of distributed demand bidding curve, Ref. [24] aggregates response capabilities of
flexible loads is known, which is not easy to obtain directly in practice. smart homes. The above literatures quantitatively analyze the regula-
Therefore, the regulation model and aggregation method of distributed tion capacity of TCLs, but they just evaluate the performance of TCLs in
flexible loads are needed. a certain time section and make load control decisions step by step,
which is referred as the “static” control in this paper. However, the
thermodynamic model of TCL is usually described as differential
1.2. Model and aggregation for distributed flexible loads equations of temperature versus time, which means the regulation be-
tween each time section influences others. Hence, because the optimi-
As for modeling the distributed flexible loads in VPP, many scholars zation period contains more than one time section in day-ahead and
focus on the analysis of thermostatically controlled loads (TCL) due to intraday optimization, it is necessary to establish a “dynamic”
their great power regulation potential, and much related work have

660
C. Wei et al. Applied Energy 224 (2018) 659–670

aggregation model considering the mutual influence of the control level model: the upper level mainly discusses VPP as the agent of DERs
signals in multi-time sections. to participate in the electricity market and minimize the imbalance cost
Third, some scholars have attempted to use dynamic aggregation to solve in intraday, while the lower level focuses on VPP allocating power
the above problem: Ref. [25] applies an energy storage model with time- regulation task to internal DERs and testing the effect of power tracking
varying parameters to aggregate TCLs. Refs. [26,27] use typical para- in real-time.
meters to replace distributed parameters due to the difficulties in col- In the upper level, after VPP determines the day-ahead planned
lecting accurate parameters of each user: Ref. [26] employs the average curve for purchasing electricity from the power grid (usually called net
value while Ref [27] uses tracer devices after parameter optimization. exchange power curve, which is shown as black dotted line in Fig. 1),
Refs. [25–27] make dynamic aggregation of TCLs, but they are all VPP makes rolling optimization based on intraday information about
model-based, which means: (1) the setting of model parameters is of PVs and loads. To be specific, in intraday, VPP make rolling forecast on
great importance. The online parameters setting is required in Ref. [25] the total output of PVs, which is closer to the actual values. On account
because of the time-varying parameters, and Ref. [26] even clearly of the inevitable gap between the intraday forecast and the day-ahead
specifies that excessive parameter heterogeneity reduces the effective- forecast, net exchange power imbalance may occur in VPP if the day-
ness. (2) They need temperature information, or require users to have ahead scheme is still implemented, further resulting in corresponding
the same kind of thermostatically controlled devices and similar tem- costs (imbalanced costs that should be paid for seeking balance service
perature set points [23–27]. But in practice, VPP contains diverse [26] or the penalty for the breach of the plan [28]). In order to reduce
thermostatically controlled devices under different temperature, and these costs, the terminal users calculate their own dynamic regulation
the operating information of these devices is unattainable because of characteristics locally and send the results to the VPP agent in the re-
the privacy protection. quired data format. The VPP agent calculates the aggregated regulation
characteristics after receiving these data packets, and run intraday
1.3. Contributions and paper layout rolling optimization model with the objective of minimizing the power
imbalance cost. Finally, the expected net exchange power curve is
In summary, VPP with flexible loads (represented by TCLs) has been corrected and updated (as blue line in Fig. 1).
proved to be efficient in improving the utilization of renewable energy, In the lower level, even if the intraday forecast is more accurate
and many aggregation methods are proposed to describe their power than the day-ahead, there is still a gap with the actual output, resulting
regulation characteristics. However, most of the existing dynamic ag- in forecast error. Moreover, considering the complexity about the TCLs
gregation methods of residential TCLs are model-based and directly models, it is inevitable to simplify them in dynamic aggregation, re-
take the temperature as the control signal, which are greatly affected by sulting in model error. Thus, it is necessary to perform real-time power
parameters and rely on the assumption that the type of devices and tracking after the upper level model has optimized the corrected net
temperature range are approximately the same. These two issues limit power curve. Each TCL only needs to upload its accurate static reg-
the application of their method. ulation characteristics, and VPP uses the static aggregation algorithm to
In this paper, a bi-level scheduling model for VPP with a large evaluate the current load regulation capacity. Finally, Base on the
number of distributed TCLs and intermittent renewable energy is con- lasted information about the PVs and FLs in real-time, VPP can further
structed, and the load aggregation methods are also established. The calculate its actual net exchange power (as read line in Fig. 1) and
main contributions are summarized as follows: works out the corresponding control signals, which can guide the reg-
ulation of TCLs.
(1) A static and dynamic aggregation method for TCLs is proposed, There are two points worth mentioning: (1) VPP participates in
in which static aggregation evaluates the accurate regulation ca- the day-ahead market and determines the planned net exchange
pacity at one time section, and the dynamic characteristics further power curve. But this part is not discussed in this paper. Here, the
consider the mutual regulation between multiple time sections. day-ahead planned curve is calculated based on the day-ahead fore-
Based on the proportional signal, the aggregation model is in- cast of RESs and flexible loads curve without regulation; (2) since the
sensitive to the heterogeneity of parameters, and the physical in- TCL model is a differential equation, the choice of time interval is
formation of the load models are also not updated due to the critical. In order to ensure the accuracy of the basic TCL model, the
privacy protection. time interval in the lower level is chosen to be 1 min. In intraday, the
(2) Based on the aggregated characteristics obtained by the above period of rolling optimization is chosen to be 2 h, while its time in-
methods, this paper establishes a bi-level scheduling model to de- terval is 15 min. The cubic spline interpolation method is used to
crease the VPP imbalance power caused by the forecast error of make up the time scale mismatching between the upper level and
RESs output. In upper level, dynamic regulation characteristics of lower level.
TCLs are applied to rolling correct the planned exchange power
curve in intraday, while the lower level uses the static regulation
characteristics to track target power curve in real-time. For time-
liness, a two-step simplified strategy is proposed to transform the
MINLP into a MILP and iterative NLP.

The remaining content is organized as follows: Section 2 describes

the framework; Section 3 introduces the control signal form and elicits
the static and dynamic aggregation method respectively; Section 4 fo-
cuses on the intraday optimization model, and gives the corresponding
two-step solving method; Section 5 is a simulation and discusses the
regulation performance of TCLs and the scope of application; and
Section 6 provides conclusions.

2. Framework

This paper focuses on the VPP consisting of distributed PVs and

TCLs. The framework is shown in Fig. 1. The research is based on a bi- Fig. 1. Framework of the bi-level scheduling.

661
C. Wei et al. Applied Energy 224 (2018) 659–670

3. Modeling of aggregated TCLs temperature set point proportionally according to their own allowed
temperature ranges.
This section describes the model of thermostatically controlled de-
ΔTc,t = ut ·ΔTcmax (5)
vices and establishes aggregation algorithms in both static and dynamic
scenarios. Here the HVAC is analyzed, and the similar method can be t
applied to others such as water heaters and refrigerators. In addition, Tcset set
,t = Tc,0 + ∑ ΔTc,τ = Tcset max
,0 + ΔTc ·SUt
some thermostatically controlled devices have two modes: cooling and τ=1 (6)
heating. For convenience, they are assumed to be working in the where ut is the proportional control signal issued by the VPP control
cooling mode. center, and each user receives the same signal; SUt is the cumulative
control signal; Tcset
,0 is the temperature set point at initial time, and it is
3.1. Modeling of single TCL also assumed to be comfortable indoor temperature set by the user,
meeting Tcset
,0 = (Tc
in,min
+ Tcin,max )/2 .
3.1.1. Basic model
Many literatures have studied the HVAC. Although the models are Compared to directly using ΔTc,t as the control signal, this method
sometimes different (Ref. [18] considers solar radiation, Ref. [25] has the following advantages: (1) Proportional adjustment means the
considers model error, and Ref. [29] considers the occupancy of people sharing equally, and ensures the fairness in allocation; (2) Users can
and equipment), the mathematical expression can be abstract into a change the allowed temperature range according to their own pre-
difference equation: ference and then decide the degree of load regulation, which follows
the voluntariness for users when they participate in DR; (3) Different
Tcin,t = fc (Tcin,t − 1,Zc,t ,Tcout
,t ) (1) users allow different temperature ranges. It is not highly practical to
judge with actual value.
where Tcin,t is the indoor temperature of user c at time t; Tcout
, t is the
In addition, since the indoor temperature cannot be uploaded to the
outdoor ambient temperature; Zc,t is the on/off status of the HVAC
control center (see Principle 2), Eq. (4) should be replaced by Eq. (7):
system. When the HVAC is working, Zc,t = 1, otherwise it is 0. fc (·)
reflects its thermodynamic function. Different users have different
⎧− 1 ⩽ SUt ⩽ 1
parameters, and even expressions may also be different. ⎩ − 1 ⩽ ut ⩽ 1
⎨ (7)
Since the power of the fixed-frequency air condition cannot be
regulated continuously, it is difficult to accurately maintain the indoor
temperature at a certain set point. It is common to sometimes open and 3.2. Aggregation of TCLs
sometimes close the compressor, and makes the indoor temperature
fluctuate periodically within a range. The logic is shown in Eq. (2): As mentioned in Section 1, only when many users coordinate with
each other can the total capacity be large enough. Therefore, it is in-
⎧1, if Tcin,t − 1
⩾ + Tcset
,t Tcdb/2 evitable to aggregate distributed flexible load.
⎪
Zc,t = 0, if Tc,t − 1 ⩽ Tc,t −Tcdb/2
in set
Before the derivation of aggregation model, it is assumed that there
⎨
⎪ Zc,t − 1, otherwise is a strong two-way communication network between users and the VPP
⎩ (2)
agent. In fact, although the fully decentralized control proposed in Ref.
where Tcset
is the temperature set point;
,t Tcdb
is the temperature dead [31] can also be adopt, we used the centralized control here for brevity.
band. In model-based aggregation, the physical concept is clear and it is
The set point change method is always used to adjust Zc,t and further easy to operate. However, the parameters contained in fc (·) and the
affect the power of the HVAC. In this case, Tcset
,t is composed of two parts: user's indoor temperature can often reflect the user's living conditions.
Tcset set Uploading them to the unified platform will involve privacy issues. For
,t = Tc,t − 1 + ΔTc,t (3)
this, the principle 2 is followed:
where ΔTc,t is the change of the temperature set point according to the
Principle 2. The communication between users and the VPP control
control instruction, and the original set point is Tcset
,t − 1.
center is only limited to information directly related to electricity (load
As for user’s comfort, it is guaranteed from two aspects: (1) the room
power, regulation ability, etc.). Other non-electricity information
temperature should be acceptable to the user (range constraint); (2) the
(HVAC parameters, room temperature, etc.) should not be uploaded.
change of temperature set point should not be too fast (rate constraint),
namely:
in,min 3.2.1. Static aggregation
⎧Tc ⩽ Tcin,t ⩽ Tcin,max
The static aggregation is used to evaluate the regulation capacity or
⎨|ΔTc,t | ⩽ ΔTcmax (4)
⎩ assign regulation tasks to each user. In Eq. (2), since the indoor
where [Tcin,min,Tcin,max ]
is the indoor temperature range set by the user;
ΔTcmax is maximum allowable change rate, and for convenience, it is
assumed that ΔTcmax = (Tcin,max−Tcin,min−Tcdb)/2 . Tcin ,max
Finally, in a HVAC, the relationship between temperature para-
meters or variables is shown in Fig. 2.
ΔTcmax Tcset
,t
3.1.2. Selecting control signals
There are two ways when TCLs participate in DR: one is to calculate Tcset
,0
ΔTc ,t
power regulation with frequency or voltage signal received from the
power system, which is often used in second-level load control [30]; the Tcset
, t -1
Tcdb
other way is at minute-level or hour-level, where the control center
directly issues power regulation signals to terminal equipment. Besides, Tcin ,min
two signal forms are common: actual value (such as directly using ΔTc,t )
[23,24,26] or relative proportion. The latter is used here, specifically: 0

Principle 1. When allocating power regulation task, users change their Fig. 2. Temperature set point change.

662
C. Wei et al. Applied Energy 224 (2018) 659–670

temperature and on/off status in previous time (τ < t) are known in coupling.
real-time, the on/off status at time t is uniquely determined by Tcset ,t if the The HVAC model established in Section 3.1 considers the on/off
forecast error of outdoor temperature is neglected. From Eq. (5) and status in a short time interval (1 min). However, when the time scale is
(6), with the known of the Tcset set
,t − 1 (or SUt−1), Tc,t (or SUt) can be replaced long and the time resolution is low (15 min), the on/off status is often
by ut. Finally, the static aggregated regulation characteristics is trans- difficult to be distinguished accurately, and the average power of the
formed to the relationship between the power demand of TCLs and ut, device in the unit time interval is not either 0 or Pcrate . Therefore, Eq. (1)
which is need to update real-time. Define a threshold uc∗,t meeting Eq. is changed to Eq. (13) with a relative long time interval:
(8), and it can be proved that Eq. (9) is equivalent to Eq. (2).
Tcin,t = fc (Tcin,t − 1,PcHVAC
,t ,Tcout
,t ) (13)
⎛Tcin,t − 1−Tcset 1
uc∗,t = ,t − 1 + ⎛Zc,t − 1− ⎞ Tcdb⎞/ΔTcmax where PcHVAC
is the power of a HVAC, and it is a continuous variable
,t
⎝ ⎝ 2⎠ ⎠ (8)
whose value range from 0 to Pcrate .
1 if u ⩽ uc∗,t In addition, if the fluctuation of the indoor temperature is ignored, it
Zc,t = ⎧ means Tcin,t ≈ Tcset out
⎨ ,t . Further in Eq. (13), since Tc,t is obtained by tem-
⎩ 0 otherwise (9)
perature forecast, it just needs to adjust Tcin,t − 1 and Tcin,t to change the
Accordingly, users only need to locally calculate uc∗,t and upload it power of HVAC. Combining Eqs. (5) and (6), Eq. (1) can be simplified as
online. VPP control center operate static aggregation program Eq. (14):
(Algorithm 1) according to the rated power reported offline. Then the
relationship between total power demand and control signal (ut) is PcHVAC
,t = gc (SUt ,SUt − 1,Tcout
,t ) (14)
obtained, as shown in Fig. 3. However, according to Principle 2, Eq. (14) can only used on the
Algorithm 1 (Static aggregation). user side (which can be done by terminal energy management systems
[24,32]), and users do not upload Eq. (14) to the VPP control center.
1: input {u1,∗t ,u2,∗ t ,⋯,u N∗cus,t } Hence, Eq. (14) is transformed into a numerical table locally, and then
it is sent to the VPP control center. The aggregation algorithm is as
2: C1 = {c|uc∗,t ⩽ 1} , C2 = {c|uc∗,t ⩾ −1} , C3 = C1 ∩ C2
follows:
3: Ptstatic (1) = ∑c ∈ C2 Pcrate , Ut (1) = −1, i = 2
Algorithm 2 (Dynamic aggregation).
4: while C3 ≠ ∅, do
5: find the user with min c ∈ C3 {uc∗,t } , and assign its index to c∗ 1: input (T1,out out out
t , T2,t ,⋯,TNcus,t )
6: Ut (i) = u c∗∗,t , Ptstatic (i) = Ptstatic (i−1)−Pcrate
∗
2: for each users c = 1 to Ncus, do
7: C3 = C3−{c∗} , i = i + 1 3: for each longitudinal segment i = 0 to Ntes1, do
8: end while 4: for each transverse segment j = 0 to Ntes2, do
9: Ptstatic (i) = Ptstatic (i−1) , Ut (i) = 1 5: SUt = −1 + (2i)/ Ntes1; SUt − 1 = −1 + (2j )/ Ntes2
10: output Ptstatic , Ut 6: Pdynamic
c,t (i,j ) = gc (SUt ,SUt − 1,Tcout
,t )
7: end for
8: end for
In Algorithm 1, Ncus is the number of users. Step 2 filters the data
9: end for
within [−1,1], while Step 3 and Step 9 expand the sequences to im-
10: output Ptdynamic = ∑ Pdynamic
prove their integrity on the regulation boundary. c ∈ C c,t
As for real-time power tracking (the lower level), the control signal
can be calculated in VPP control center with the following equation:
In Algorithm 2, two-dimensional indexes are required in the dis-
Pttarget = Ptnet ,real−PtPV ,real (10) cretization as shown in Step 3–8, and Ntes1, Ntes2 are the numbers of
discretization segments. Steps 2–9 are calculated by each user locally,
⎧1, if Pt
target
< min{Ptstatic} and Step 10 is made in the VPP control center. Pdynamic
c,t is the dynamic
⎪ Pt
target
− Ptstatic (k ) power regulation characteristic of an individual user while Ptdynamic is the
∼ = ⎪ Ut (k ) + (Ut (k + 1)−Ut (k )) Ptstatic (k + 1) − Ptstatic (k )
u total value obtained by aggregating cluster TCLs. Fig. 4 shows the result
t
⎨ static target of the Algorithm 2 at a specific time.
⎪ if Pt (k ) ⩽ Pt ⩽ Ptstatic (k + 1)
⎪− 1, if P target > max{P static}
⎩ t t (11)
4. VPP intraday rolling optimization
∼ ,1−SU }}
ut = min{−1−SUt − 1, max{u (12)
t t−1
VPP can participate in the electricity market in med-term [33], day-
where Ptnet ,real is the corrected VPP net exchange power converted from
ahead [11] and intraday [26]. In day-ahead, VPP determines the
Ptnet in Section 4; PtPV ,real is the real-time output of PVs. Eq. (10) cal-
∼ is the linear in- planned purchasing curve. However, since VPP contains RESs inside,
culates the real-time power regulation target, and u t
the uncertainty of their output leads to a gap between the actual and the
terpolation result, which is further limited by Eq. (12) according to Eq.
planned net exchange power. Thus it is considered that VPP uses the
(7).
Finally, after the VPP control center works out ut, it will be issued to
Ptstatic
each user, and users only need to perform accordingly. All model
parameters are translated into uc∗,t , and only one data (uc∗,t or ut) is
PcHVAC
transmitted for each user in each interaction, which can reduce the data ,t Pt target
quantity. Pcrate

3.2.2. Dynamic aggregation 0

-1 uc ,t 1 Ut -1 ut 0 1 Ut
Dynamic aggregation is used in intraday rolling optimization for
providing power regulation characteristics of TCLs. Since the power of Update the Boundary uc ,t Calculate ut through interpolation

HVAC is determined by the temperatures in several time sections, the

dynamic aggregation model should reflect the multi-time section Fig. 3. The static aggregation.

663
C. Wei et al. Applied Energy 224 (2018) 659–670

Eq. (19).
In Eq. (16), PtTCL is a decision variable, and the regulation char-
acteristic is given by Ptdynamic in Section 3.2.2. This process is expressed
Total power of TCLs (MW)

as:
PtTCL = interpolation(SUt − 1,SUt ,Ptdynamic ) (20)
where SUt−1 and SUt are used to conduct two-dimensional linear in-
terpolation for the numerical table (matrix) Ptdynamic .

4.2. Transformation of the model

The model established in section 4.1 is nonlinear programming: (1)

There is a judgment for max{·} in Eq. (15); (2) Eq. (18) is quadratic
SU equation constraints; (3) Eq. (20) is a two-dimensional interpolation.
t-1
(p.u For fast solving the problem, these three parts are equivalently trans-
.) p.u.)
SUt ( formed.

Fig. 4. A sample for the result of dynamic aggregation.

4.2.1. Transformation of the objective function
Two time-independent variables P c,up and P c,down are introduced,
lasted intraday forecast information to re-optimize the planned net meeting:
exchange power curve, so as to minimize the costs for this power im-
balance. P c,up ⩾ Ptup,P c,down ⩾ Ptdown (21)
Further, Eq. (15) is modified to Eq. (22).
4.1. Modeling
mincost = Δt ∑ (ρeup Ptup + ρedown Ptdown) + ρcup P c,up + ρcdown P c,down
t (22)
The goal is to make the net exchange power curve as close as pos-
sible to the planned curve by regulating its controllable units. In gen- Eq. (21) implies P c,up ⩾ max t {Ptup} and P c,down ⩾ max t {Ptdown} . So
eral, ESSs, generators and FLs are applied. In order to verify the reg- when Eq. (22) obtains optimal solution, it must meet P c,up = max t {Ptup}
ulation performance of aggregated TCLs, it is assumed that only cluster and P c,down = max t {Ptdown} . Namely, after introducing Eq. (21), Eq. (15)
TCLs are used. and Eq. (22) are equivalent.

4.1.1. Objective function 4.2.2. Omission of mutually-exclusive constraints

The objective function is to minimize the imbalance costs of VPP, Ref. [34] testifies that battery charge/discharge mutually-exclusive
including the imbalance energy cost (accumulated value) and im- equation constraints can be omitted somewhere. It is also can be proved
balance capacity cost (maximum value) [26], calculated by: that this method is applicable to our problem. For convenience, the
optimization problem in 4.1 is called “the original problem”, and after
mincost= Δt ∑ (ρeup Ptup + ρedown Ptdown) omitting Eq. (18), it is called “simplified problem”.
t
∼up ∼down ∼net ∼TCL
+ ρcup max t {Ptup} + ρcdown max t {Ptdown} Proof. Set (Pt ,Pt ,Pt ,Pt ) as the optimal solution to the simplified
(15)
problem with the objective function value cost1. Construct
where Ptup and Ptdown are respectively the upward/ downward deviation up down net TCL up ∼up down ∼down
(Pt ̂ ,Pt ̂ ,Pt ̂ ,Pt ̂ ) meeting Pt ̂ = Pt −ΔPt , Pt ̂ = Pt −ΔPt ,
between the intraday and the day-ahead planned net exchange power; ∼up ∼down
where ΔPt = min{Pt ,Pt } , and its objective function value is cost2.
ρeup and ρedown are respectively the upward/downward imbalance energy
up down
price. ρcup and ρcdown are respectively the upward /downward im- Obviously, (Pt ̂ ,Pt ̂ ) meets Eq. (18) and (19) and also meets
up down ∼ up ∼down net ∼net
balanced capacity price. All prices above are positive and a constant ̂
Pt −Pt ̂ = Pt −Pt . From Eq. (17), Pt ̂ = Pt due to the same
price mode is selected to reduce the interference caused by electricity TCL ∼TCL net TCL
Ptnet ,ahead . Further, from Eq. (16), Pt ̂ = Pt . So (Pt ̂ ,Pt ̂ ) must meet
price fluctuation. up down net TCL
Eqs. (16) and (17), (20). Namely (Pt ̂ ,Pt ̂ ,Pt ̂ ,Pt ̂ ) is not only a
feasible solution to the simplified problem, but also a feasible solution
4.1.2. Constraints
to the original problem.
Since VPP more focuses on the integration of DERs as well as power ∼up up ∼down down
From Eq. (19), ΔPt ⩾ 0 , so Pt ⩾ Pt ̂ , Pt ⩾ Pt ̂ , which means
exchange with the grid, the requirements on network are not high ∼up ∼down ∼net ∼TCL
(sometimes do not own grid operation rights). So the total power bal- cost1 ⩾ cost 2 . But in the simplified problem, (Pt ,Pt ,Pt ,Pt ) is its
up down net TCL
ance is considered without the network constraints. optimal solution while (Pt ̂ ,Pt ̂ ,Pt ̂ ,Pt ̂ ) is just a feasible solution, so
cost1 ⩽ cost 2 . Combine the above conclusions, cost1 = cost 2 , and the
Ptnet + PtPV ,intra = PtTCL + Ptbasic (16) ∼up ∼down
condition of equality is ΔPt = 0 , in other words, Pt = 0 or Pt = 0,
∼up ∼down ∼net ∼TCL
Ptnet = Ptnet ,ahead + Ptup−Ptdown (17) meeting Eq. (18). Namely (Pt ,Pt ,Pt ,Pt ) must be a feasible so-
lution to the original problem (∗).
Ptup·Ptdown = 0 (18) As for any feasible solution to the original problem (set the corre-
sponding objective function value is cost3), it must also be a feasible
Ptup ⩾ 0,Ptdown ⩾ 0 (19) solution to the simplified problem, and meets cost1 ⩽ cost 3. Combine
∼up ∼down ∼net ∼TCL
where Eq. (16) is the balance constraint; is the net power exchangePtnet with (∗), in the original problem, (Pt ,Pt ,Pt ,Pt ) is the feasible
between VPP and the grid; PtPV ,intra is the intraday rolling forecast of solution with the minimum objective function value. Namely
∼up ∼down ∼net ∼TCL
total output of PVs; Ptbasic is the basic uncontrollable loads; PtTCL is the (Pt ,Pt ,Pt ,Pt ) is the optimal solution to the original problem.
aggregated power of TCLs. Eq. (17) calculates the power deviation, and
Ptnet ,ahead is the day-ahead planned curve. Eq. (18) is the mutually-ex- 4.2.3. Mathematical expression of the interpolation
clusive constraint between upward/downward deviations, limited by As mentioned above, Eq. (19) is a two-dimensional linear

664
C. Wei et al. Applied Energy 224 (2018) 659–670

interpolation where the logical judgment is contained and need to in- 4.3.2. Step 2: Solving the precise problem
troduce 0–1 variables. The schematic diagram is shown in Fig. 5, and After Step 1, the cell where the operating point of each time section
the corresponding mathematical expression is as follows: is located has been determined, which means i, j has been known. So
start
σi,j,t , δi,j,t , K i,Xj,t , K iY,j,t , KK i,XY
j,t and Pi,j,t regress only dependent of time t
0 ⩽ σi,j,t ⩽ Ii,j,t ·σimax max
,t ,0 ⩽ δi,j,t ⩽ Ii,j,t · δ j,t (23)
(respectively abbreviated as σt , KtX , KtY , KKtXY , and Ptstart ). At the same
∑∑ Ii,j,t = 1 time, the quadratic term is incorporated in Step 2 to obtain more ac-
i j (24) curate results:
0 ⩽ σt ⩽ σtmax,0 ⩽ δt ⩽ δtmax (29)
SUt − 1 = ∑∑ (Xistart
,j,t · Ii,j,t + σi,j,t )
i j (25) SUt − 1 = σt + σtstart ,SUt = δt + δtstart (30)
SUt = ∑∑ (Yistart
,j,t · Ii,j,t + δi,j,t ) PtTCL = KKtXY σt δt + KtX σt + KtY δt + Ptstart (31)
i j (26)
where σtstart = ∑i ∑j σistart
,j,t Ii,j,t , δtstart = ∑i ∑j δistart
,j,t Ii,j,t , Ptstart =
PtTCL = ∑∑ KK i,XY
j,t · σi,j,t · δi,j,t + ∑∑ K i,Xj,t ·σi,j,t + ∑∑ K iY,j,t ·δi,j,t ∑i ∑j Pistart
,j,t Ii,j,t .
i j i j i j
The optimization in Step 2 can be summarized as follows:
+ ∑∑ Pistart
,j,t · Ii,j,t
i j (27) Objective: Eq. (22).
where Eq. (23) limits the length in each segment; Ii,j,t is the auxiliary Subject to: Eq. (7), (16), (17), (19), (21), (29)–(31).
0–1 variable; σi,j,t and δi,j,t are respectively the longitudinal/transverse This is a simple NLP (especially QCP), which is solved with TOMLAB
segment variables with their maximum value σimax ,t and δ jmax
,t . Eq. (24) / SNOPT.
indicates that only one cell can be selected in each time section. Eqs.
(25) and (26) are used to calculate SU. Xistart start
,j,t and Yi,j,t are the origins of 4.3.3. Adjusting the solution
σi,j,t and δi,j,t , respectively. Eq. (27) is used to calculate PtTCL , and Pistart ,j,t is
Since the quadratic term is omitted in Step 1, I needs to be cor-
the origin meeting Pistart dynamic
(i,j ) ; K i,Xj,t and K iY,j,t is the longitudinal rected. Locally adjust I according to σt , δt obtained in Step 2: when σt
,j,t = Pt
and transverse slope of each cell, meeting and δt come near to 0 or maximum, it indicates that I may directly limit
K i,Xj,t = (Ptdynamic (i + 1,j )−Ptdynamic (i,j ))/ σimax,t , K iY,j,t = (Ptdynamic (i,j + 1) the value of σt and δt , and there may be a better scheme after adjusting
−Ptdynamic (i,j ))/ δ jmax KK i,XY dynamic
(i + 1,j + 1)−Ptdynamic I. Thus the following method is used:
,t , and j,t = (Pt
(i + 1,j )−Pta (i,j + 1) + Pta (i,j ))/(σimax δ max
) can be analogy to the second-
,t j,t
⎧ 1, σt ⩾ (1−η) σtmax
order partial derivative.
Ntσ = 0, ησtmax < σt < (1−η) σtmax
⎨
4.3. Summarizing and solving ⎩−1, σt ⩽ ησtmax

⎧ 1, δt ⩾ (1−η) δtmax
Combine Sections 4.1 and 4.2, the intraday rolling optimization Ntδ = 0, ηδtmax
< δt < (1−η) δtmax
model can be expressed as: ⎨
⎩−1, δt ⩽ ηδtmax

Objective: Eq. (22). 1, i = (it∗ + Ntσ ),j = (jt∗ + Ntδ )

Subject to: Eq. (7), (16), (17), (19), (21), (23)–(27). Ii,j,t = ⎧
⎨
⎩ 0, otherwise
Obviously, the model is transformed into a MINLP (especially where (it∗,jt∗) is the location of the operating point, and η = 5%.
MIQQ). problem, but in this paper: Ntes1 and Ntes2 should not be too Apply the adjusted I into Step 2 for calculation again, and then
small because of the discretization error, resulting in the large number continue to correct I based on the σt , δt re-output by Step2, until there is
of 0–1 variables; Besides, since the quadratic term in Eq. (27) is blended no new I combination scheme appearing. The best result in the iterative
with the 0–1 variables, it cannot be transformed into a second-order process is selected as the final result.
cone constraint, which leads to the failure of some MIQQ solving
methods. Therefore, a two-step simplified algorithm is proposed to 5. Case study
quickly obtain the approximate optimal solution of the problem here.
The basic idea of simplification is to separate quadratic term from in- The effectiveness of the method proposed in Sections 3 and 4 is
tegers, shown in Fig. 6. discussed. First, the effect of TCLs on reducing the imbalance cost of
Where obj is the objective function value, stored in the set obj, and VPP caused by the forecast error of PVs is verified. Second, the sensi-
its solution sol is stored in the set sol. kmax is maximum iteration times, tivity to the homogeneity of TCL parameters is tested. Third, simulation
whose value is related to the number of discretization points and the under different kinds of forecast errors is considered. Finally, the sol-
length of the optimization period. ving speed of the two-step simplified algorithm is examined.

4.3.1. Step 1: Solving the approximate problem

The first step is to solve the logical judgment, that is, to determine
the value of Ii,j,t . Here, if the second-order is ignored, Eq. (27) can be
reduced to Eq. (28):

PtTCL = ∑∑ (K i,Xj,t σi,j,t + K iY,j,t δi,j,t + Pistart

,j,t · Ii,j,t )
i j (28)
The optimization in Step 1 can be summarized as follows:

Objective: Eq. (22).

Subject to: Eq. (7), (16), (17), (19), (21), (23)–(26), (28).
This is a large-scale MILP, and is solved by TOMLAB/CPLEX. Fig. 5. Two-dimensional linear interpolation.

665
C. Wei et al. Applied Energy 224 (2018) 659–670

ahead forecast, causing that SU is negative and continues declining in

Fig. 8(c). It means that HVACs continuously reduces the temperature
set point to increase loads. However, the regulation capacity of TCLs is
limited, and increasing loads for a long time means that HVACs should
keep on-status as long as possible, leading to a high synchronization. In
Scheme 2, if SU is maintained at -1 for a long time, HVACs must be off
to avoid too low indoor temperature. The shutdown of a large number
of HVACs in a short period results in a steep drop in VPP net exchange
power during t = 660–700 min. This confirms that there are the pos-
sibility of reducing load diversity and increasing load fluctuation when
TCLs participate in DR [38].
Nevertheless, if Scheme 3 is used, the SU is higher than that of
Scheme 2 during t = 600–690 min. It means that after the intraday
rolling optimization, the VPP control center integrates the intraday
forecast to optimize the net exchange curve in 2 h advance. It reduces
the load regulation pressure for leaving the reserve during
t = 669–690 min.
Moreover, the imbalance costs are checked, and the corresponding
results are shown in Table 2:
In Scheme 1, the difference between the actual PV output and the
day-ahead forecast is directly calculated as imbalance power when pay
imbalance cost to the grid. These costs are up to $745.40, indicating
that VPP should have regulation ability if high-permeability PVs are
Fig. 6. Flowchart of the two-step simplified algorithm. contained. Otherwise, too large imbalance costs can reduce the eco-
nomics of VPP. In Scheme 2, after TCLs participate, the imbalance cost
5.1. Parameter settings is reduced, demonstrating that on a typical summer day when TCLs

Assume that the VPP serves residential or commercial customers

with 2100 HVACs on a typical summer day, and the curves of the un-
controllable load and the outdoor temperature are shown in Fig. 7(a)
and (b), respectively. Refs. [23,24] present the ETP model which is
applied to Eq. (1), and its parameters are generated randomly with the
uniform distributions. The mean value of these distributions is shown in
Table 1 [25,35]. It can be seen in Fig. 7(c), based on the recorded data
in Walloon-Brabant [36], the total capacity of PV is assumed as 5 MW.
In practice, imbalance prices are closely related to the regulation ability
and operation status of the grid, and the price-fixing is beyond the re-
(a) Uncontrollable load curve.
search of this paper. Thus, under the constant price mode, the im-
balance energy/capacity prices are set 100 $/MWh and 200 $/MWh,
respectively [37]. As for the discretization and iteration i,
Ntes1 = Ntes2 = 20 and kmax = 30 are set here, respectively.

5.2. Effects of TCLs

This section tests the effects of TCLs in a typical case. Additionally,

more analyses are done from two aspects: imbalance prices and tem-
perature ranges. Besides, the following three strategies are compared:

(b) Outdoor temperature.

Scheme 1: TCLs do not participate in DR. The VPP net exchange
power deviation between the actual and the day-ahead demand is all
solved by seeking the balance service from the grid.
Scheme 2: TCLs participate in DR. After the day-ahead planned
curve is determined, TCLs directly track it at real-time, which only
considers the minimum of the current power deviation.
Scheme 3: TCLs participate in DR. After the day-ahead planned
curve is determined, VPP uses the aggregation algorithm proposed
in this paper to calculate the regulation characteristics of cluster
TCLs which are applied into the intraday optimizing model to
rolling correct the planned curve. Subsequently, TCLs track the
latest optimized curve in real-time.

5.2.1. Typical case

The power tracking results are shown in Fig. 8. (The forecast error is (c) PV output data.
calculated as: Forecast error = PtPV ,ahead−PtPV ,intra )
During t = 480–720 min, the PV output is higher than the day- Fig. 7. Basic input information.

666
C. Wei et al. Applied Energy 224 (2018) 659–670

Table 1 Table 3
Thermal parameters of the HVACs. Comparisons under different prices.
ETP parameters P rate (kW) COP (p.u.) R (°C/kW) C (kWh/°C) Imbalance Imbalance energy Imbalance cost ($) Relative
5.6 2.5 2 2 capacity price price improve (%)
Scheme 2 Scheme 3
Temperature setting T0set (°C) T min (°C) T max (°C) Tcdb (°C)
0 100 59.19 61.91 −4.59
23 22 24 1
50 100 107.90 107.78 0.11
100 100 156.62 125.10 20.13
200 100 254.05 158.90 37.45
400 100 488.90 226.50 49.54

More details
in Fig.8 (b) account for a large proportion, they can contribute to deal with the
forecast error of PVs output. In Scheme 3, the imbalance cost is further
reduced to 37.45% compared with Scheme 2. More specifically, the
decrease in total cost is mainly resulted by the reduction of the max-
imum imbalance power (nearly 65.30%). However, it should be pointed
out that the energy costs increase as the capacity costs reduce, which
means that reducing the maximum imbalance power often leads to the
increase in the imbalance energy.

(a) VPP actual net exchange power.

5.2.2. Imbalance prices
For the ratio of imbalance capacity price to imbalance energy price,
it supposes that ρeup ≈ ρedown and ρcup ≈ ρcdown , and the imbalance energy
price is fixed to $100/MWh, while the imbalance capacity price is ad-
justed.
In Table 3, without the imbalance capacity cost, Scheme 3 is slightly
inferior to Scheme 2, because TCLs try their best to make up the de-
viation currently, which is an efficient measure [22]. When the im-
balance capacity cost is considered, Scheme 3 is better than Scheme 2.
Furthermore, the higher imbalance capacity price, the more reduction
of the imbalance cost. Hence, it can be concluded that this method is
applicable where the imbalance capacity cost is high.

(b) Power tracking deviation. 5.2.3. Regulating range

In this section, the power regulating of TCLs is changed by adjusting
[Tcin,min,Tcin,max ], which can further affect ΔTcmax . The corresponding
maximum power deviation and the imbalance costs are also calculated,
which are shown in Fig. 9.
In Fig. 9, when the temperature regulation range is too small, the
maximum imbalance power in Scheme 2 is larger than Scheme 1 (when
ΔTcmax = 0 ). This is because the load regulation capacity reaches the
Reduce regulation
to leave reserve limit soon if the regulating range is too small, which increases the load
fluctuation as the analysis in Section 5.2.1. It can be concluded that
TCLs not certainly reduce the maximum imbalance power when parti-
cipating in DR. However, even if the maximum imbalance power in-
creases, the total imbalance cost still decrease, because TCLs can reduce
imbalance energy before the regulation limit is reached. Therefore, it
also benefits to have TCLs participate in regulation for the cost saving.
(c) The curve of SU. The method is better when the temperature regulating range is
small. For example, when ΔTcmax = 0.5 °C , it can reduce the imbalance
Fig. 8. Results in typical case. costs by 53.06% compared with Scheme 2, but when ΔTcmax = 1 °C, the
improvement is reduce to 37.45%. Therefore, the conclusion can be
Table 2 drawn that the method can be better applied when the regulation ca-
Imbalance costs in typical case. pacity of FLs is insufficient.
Nevertheless, it can be found in Fig. 9(b) that although Scheme 3
Imbalanced capacity Imbalanced energy Total imbalance
has advantages, the degree of improvement to Scheme 2 is reduced
cost ($) cost ($) cost ($)
compared with that in Fig. 9(a), which is due to the increase of im-
Scheme 1 283.07 462.33. 745.40 balance energy as mentioned before. Additionally, when
Scheme 2 194.85 59.19 254.04 ΔTcmax = 1.25 °C , Scheme 3 is not as good as Scheme 2, because the
Scheme 3 67.60 91.29 158.89
error of TCLs model used in intraday still exist, and the regulation
characteristic obtained by the dynamic aggregation method is com-
paratively simplified.

667
C. Wei et al. Applied Energy 224 (2018) 659–670

Hence, the PV data for the other two days in the Walloon-Brabant is
selected, shown in Fig. 10. To demonstrate clearly, the PV data in Fig. 7
is named PV 1 and the PV data in Fig. 10 is named PV 2 and PV 3,
respectively. In Table 5, it can be found that the statistical characteristic
of PV 1 and PV 2 is similar, while the forecast error of PV 3 is larger
than that of PV 1.
Power tracking results are examined when PV 2 and PV 3 are input,
as shown in Fig. 11.

(1) Comparison between PV 1 and PV 2

In Fig. 11(a), the VPP net exchange power deviation under PV 2 is

less than 0.1 MW, showing much better results than that under PV 1.
More specifically, the real-time measured values in PV 1 maintain
(a) Maximum deviation. larger than day-ahead forecast values for a long time, while in PV 2 they
are sometime higher and sometimes lower than the day-ahead forecast
values with changeable positive-negative polarity of forecast error. This
leads two different results: in PV 2, when t = 580 min, the forecast error
begins to decrease and the required load regulation is reduced, which
corresponds to a small u and slow the speed of SU; when t = 720 min,
the forecast error turns from negative to positive, and the load reg-
ulation is changed from increase to decrease, which makes u become
positive and SU start to approach 0. Both effectively alleviate the power
regulation pressure, which avoids the adjustment of the planned curve
due to the lack for power regulation. This shows that when TCLs make
up the power deviation, changeable polarity of forecast error is more
conducive to the completion of target tracking.

(2) Comparison between PV 1 and PV 3

In Fig. 11(b), PV 3 is an extreme case relative to PV 1. Although the

forecast error polarity is similar to that of PV 1 (the actual value
(b) Imbalance costs. maintains higher than the forecast values for a long time), the forecast
error of PV 3 is obviously larger than that of PV 1, which changes the
Fig. 9. The performance under different regulating ranges. results significantly, as shown in Table 6.
In Table 6, each indicator under PV 3 in both methods is inferior to
5.3. Heterogeneity of parameters that under PV 1, which demonstrates that with similar error polarity,
the bigger the forecast error, the worse the tracking effect. However, in
As mentioned in Section 1, the model-base aggregation method is extreme case, Scheme 3 is still superior to Scheme 2, saving 15.60%
affected by the heterogeneity of the model parameters. For this, Scheme imbalance costs compared.
4 is introduced referring to the method in Ref. [26].
Scheme 4: TCLs participate in DR. After the day-ahead planned 5.5. Calculation performance
curve is determined, VPP uses the TCL model with the typical para-
meters (average values are used here) to replace the actual models with As the intraday rolling optimization guides real-time power
distributed parameters, and applies it into intraday optimization to tracking, the timeliness of the algorithm is critical to the practical ap-
rolling correct the planned curve. At real-time, TCLs track the latest plication. Moreover, it should be pointed out that compared with linear
target curve.In addition, X represents Pcrate , Qcrate , R c , Cc and ΔTcmax ; X programming in Refs. [25–27], the optimization model is non-linear
represents their average value; s represents the degree of heterogeneity programming, and complexity and solving speed are the main diffi-
quantitatively, meeting X ∈ [(1−s ) X ,(1 + s ) X ]; Tcset
,0 meets culties. Thus, the timeliness of the simplified algorithm is verified.
Tcset
,0 ∈ [T set
basic + (1−s )(T set
0 −T set
basic ),T set
basic + (1 + s )(T set
0 −T set
basic ,
)] where Statistics is made on the calculation time for 440 times optimiza-
set
Tbasic = 16°C . Simulation results are shown in Table 4: tions in 5 days (because the number of variables during the last 2 h in
In Table 4, when there is no heterogeneity in model parameters, the each day is significantly less than other periods, so their statistics data is
only error in Scheme 4 is the transition of the intraday and real-time. not included). The results are shown in Fig. 12.
Compared to Scheme 2, Scheme 4 reduces the imbalance costs by In Fig. 12, the longest time consumption of a single intraday opti-
nearly 45.91%, while Scheme 3 is inferior to Scheme 4. However, when mization is less than 8 s, and 90% of the intraday optimizations are
s = 0.1 or 0.2, the parameters show heterogeneity, Scheme 3 is better
than Scheme 4. When s = 0.3, the high heterogeneity of the parameters Table 4
makes Scheme 4 fail, and Scheme 3 still performs well, reducing im- Comparisons under different ETP parameters.
balance costs by 37.88% compared to Scheme 2. This shows that the
s Total imbalance cost ($)
method is not sensitive to the heterogeneity of parameters, and can be
used where user parameters are widely distributed. Scheme 2 Scheme 3 Scheme 4

0 317.35 171.64 163.68

0.1 265.48 168.83 179.22
5.4. Forecast error
0.2 254.05 158.90 208.36
0.3 242.21 150.47 275.84
This section examines the performance under different PV data.

668
C. Wei et al. Applied Energy 224 (2018) 659–670

(a) Power tracking deviation under PV 2.

(a) PV 2

(b) Power tracking deviation under PV 3.

(b) PV 3
Fig. 10. PV output data.

Table 5
Comparisons between PV 1, PV 2 and PV 3.
PV 1 PV 2 PV 3

Maximum power (MW) 4.12 4.25 3.05

Maximum forecast error (MW) 1.42 1.36 1.48
Cumulative forecast error (MWh) 4.62 4.83 8.32

completed within 6 s. This shows that the simplified algorithm has good (c) SU under three sets of PV data.
timeliness, and it can guide real-time power tracking.
Furthermore, the average time consumption for solving the MILP in Fig. 11. The performance under different forecast error.
Step 1 and the NLP in Step 2 is shown in Table 7.
In Table 7, the calculation time mainly goes to solving MILP in Step Table 6
1, taking 62.26% on average. Although Step 2 is solved several times in Imbalance cost under PV 3.
iterations, on the one hand, the determination of I is completed with the Imbalance capacity Imbalance energy Total imbalance cost
average iteration of 2.989 times, and the longest one only has 7 times; cost ($) cost ($) ($)
on the other hand, without integer variables, the average solving time
Scheme 2 134.35 325.70 460.05
of the simple QCP in Step 2 is only 0.59 s. Therefore, the algorithm does
Scheme 3 244.40 300.72 545.12
not take too much time.
In addition, since the power regulation characteristics of individual
users is locally calculated in parallel by users, the VPP control center
just integrates these calculation results. In short, the quantity size of level, thermostatically controlled loads are used to track target curve,
HVACs only affects the redundancy of information transmission and so as to reduce the imbalance power caused by the forecast error of
does not cause the curse of dimensionality in the optimization model. renewable energy. In addition, proportional control signals are applied
How to build a strong communication network is beyond the scope of instead of temperature control signals to protect the privacy of users
this study. participating in demand response. With random sampling, 2100 sam-
ples are selected, and their ability to participate in the operation of
6. Conclusion virtual power plant after of aggregation is verified. The results show
that in typical case, the method in this paper will reduce the maximum
This paper proposes a dynamic and static aggregation method for imbalance power and save the imbalance cost. Moreover, this method is
thermostatically controlled loads, and establishes a bi-level scheduling not affected by the heterogeneity of parameters, and it is more suitable
algorithm for virtual power plant. At intraday level, rolling optimiza- for virtual power plant with diverse users. It also proves the conclusion
tion is made to correct total net exchange power curve. At real-time that the changeable polarity of forecast error of renewable energy is

669
C. Wei et al. Applied Energy 224 (2018) 659–670

industrial virtual power plant using the best demand response strategy. Appl Energy
2016;164:590–606.
[11] Ju L, Tan Z, Yuan J, et al. A bi-level stochastic scheduling optimization model for a
virtual power plant connected to a wind–photovoltaic–energy storage system con-
sidering the uncertainty and demand response. Appl Energy 2016;171:184–99.
[12] Cui H, Li F, Hu Q, et al. Day-ahead coordinated operation of utility-scale electricity
and natural gas networks considering demand response based virtual power plants.
Appl Energy 2016;176:183–95.
[13] Kardakos EG, Simoglou CK, Bakirtzis AG. Optimal offering strategy of a virtual
power plant: a stochastic bi-level approach. IEEE Trans Smart Grid
2015;7(2):794–806.
[14] Mashhour E, Moghaddas-Tafreshi SM. Bidding strategy of virtual power plant for
participating in energy and spinning reserve markets—part I: problem formulation.
IEEE Trans Power Syst 2011;26(2):949–56.
[15] Rahmani-Dabbagh S, Sheikh-El-Eslami MK. A profit sharing scheme for distributed
energy resources integrated into a virtual power plant. Appl Energy
2016;184:313–28.
[16] Thavlov A, Bindner HW. Utilization of flexible demand in a virtual power plant set-
up. IEEE Trans Smart Grid 2015;6(2):640–7.
Fig. 12. Time consumption of a single intraday optimization. [17] Bianchini G, Casini M, Vicino A, et al. Demand-response in building heating sys-
tems: A Model Predictive Control approach. Appl Energy 2016;168:159–70.
[18] Jin X, Mu Y, Jia H, et al. Dynamic economic dispatch of a hybrid energy microgrid
Table 7 considering building based virtual energy storage system. Appl Energy
Time consumption in Step 1 and Step 2. 2016;194:386–98.
[19] Michailidis IT, Schild T, Sangi R, et al. Energy-efficient HVAC management using
Total (s) Step 1 (s) Step 2 (s) Iteration (times) cooperative, self-trained, control agents: a real-lifebuilding case study. Appl Energy
2018;211:113–25.
Average 4.696 2.924 1.767 2.989 [20] Ruiz N, Cobelo I, Oyarzabal J. A direct load control model for virtual power plant
Maximum 7.234 4.292 3.810 7 management. IEEE Trans Power Syst 2009;24(2):959–66.
[21] Yin R, Kara EC, Li Y, et al. Quantifying flexibility of commercial and residential
loads for demand response using setpoint changes. Appl Energy 2016;177:149–64.
[22] Williams T, Wang D, Crawford C, et al. Integrating renewable energy using a smart
more favorable to the regulation of aggregated thermostatically con-
distribution system: Potential of self-regulating demand response. Renew Energy
trolled loads. Finally, the timeliness of the two-step simplified strategy 2013;52(2):46–56.
is proposed and validated by making statistics on 440 times intraday [23] Lu N, Zhang Y. Design considerations of a centralized load controller using ther-
optimizations. mostatically controlled appliances for continuous regulation reserves. IEEE Trans
Smart Grid 2014;4(2):914–21.
The future work will focus on the coordination method of load [24] Jiang Y, Xu J, Sun Y, et al. Day-ahead stochastic economic dispatch of wind in-
models with different time scales. Besides, VPP can provide multiple tegrated power system considering demand response of residential hybrid energy
services, so after we finished the model constructing, how it profit from system. Appl Energy 2017;190:1126–37.
[25] Mathieu JL, Kamgarpour M, Lygeros J, et al. Arbitraging intraday wholesale energy
a practical electricity market should be analyzed. market prices with aggregations of thermostatic loads. IEEE Trans Power Syst
2015;30(2):763–72.
Acknowledgements [26] Zhou Y, Wang C, Wu J, et al. Optimal scheduling of aggregated thermostatically
controlled loads with renewable generation in the intraday electricity market. Appl
Energy 2017;188:456–65.
This work was supported in part by the National Key R&D Program [27] Iacovella S, Ruelens F, Vingerhoets P, et al. Cluster control of heterogeneous ther-
of China [2017YFB0902900], in part by the National Natural Science mostatically controlled loads using tracer devices. IEEE Trans Smart Grid
2017;8(2):528–36.
Foundation of China [51477122].
[28] Liang Z, Alsafasfeh Q, Jin T, et al. Risk-constrained optimal energy management for
virtual power plants considering correlated demand response. IEEE Trans Smart
References Grid 2017; PP(99): p. 1–1.
[29] Aswani A, Master N, Taneja J, et al. Reducing transient and steady state electricity
consumption in HVAC using learning-based model-predictive control. Proc IEEE
[1] Verzijlbergh RA, Vries LJD, Dijkema GPJ, et al. Institutional challenges caused by 2012;100(1):240–53.
the integration of renewable energy sources in the European electricity sector. [30] Liao S, Xu J, Sun Y, et al. Load-damping characteristic control method in an isolated
Renew Sustain Energy Rev 2017;75:660–7. power system with industrial voltage-sensitive load. IEEE Trans Power Syst
[2] Li X, Zhang R, Bai L, et al. Stochastic low-carbon scheduling with carbon capture 2016;31(2):1118–28.
power plants and coupon-based demand response. Appl Energy 2018;210:1219–28. [31] Chen C, Wang J, Kishore S. A distributed direct load control approach for large-
[3] Tang C, Xu J, Sun Y, et al. Look-ahead economic dispatch with adjustable con- scale residential demand response. IEEE Trans Power Syst 2014;29(5):2219–28.
fidence interval based on a truncated versatile distribution model for wind power. [32] Ji HY, Baldick R, Novoselac A. Dynamic demand response controller based on real-
IEEE Trans Power Syst 2018;33(2):1755–67. time retail price for residential buildings. IEEE Trans Smart Grid 2014;5(1):121–9.
[4] Kieny C, Berseneff B, Hadjsaid N, et al. On the concept and the interest of virtual [33] Pandžić H, Kuzle I, Capuder T. Virtual power plant mid-term dispatch optimization.
power plant: Some results from the European project Fenix. Power & Energy Society Appl Energy 2013;101(1):134–41.
General Meeting, 2009. PES '09. IEEE. IEEE; 2009. p. 1–6. [34] Li Z, Guo Q, Sun H, et al. Storage-like devices in load leveling: Complementarity
[5] Pandžić H, Morales JM, Conejo AJ, et al. Offering model for a virtual power plant constraints and a new and exact relaxation method. Appl Energy 2015;151:13–22.
based on stochastic programming. Appl Energy 2013;105(5):282–92. [35] Mathieu TL, Dyson M, Callaway DS. Using residential electric loads for fast demand
[6] Zapata J, Vandewalle J, D'Haeseleer W. A comparative study of imbalance reduction response: The potential resource and revenues, the costs, and policy re-
strategies for virtual power plant operation. Appl Therm Eng 2014;71(2):847–57. commendations. ACEEE Summer Study on Buildings; 2012.
[7] Zamani AG, Zakariazadeh A, Jadid S. Day-ahead resource scheduling of a renewable [36] Elia. http://www.elia.be/en/grid-data/power-generation/Solar-power-generation-
energy based virtual power plant. Appl Energy 2016;169:324–40. data/Graph; [accessed 2018.04.04].
[8] Giuntoli M, Poli D. Optimized thermal and electrical scheduling of a large scale [37] Elia. http://www.elia.be/en/grid-data/balancing/imbalance-prices; [accessed
virtual power plant in the presence of energy storages. IEEE Trans Smart Grid 2018.04.04].
2013;4(2):942–55. [38] Zhang W, Kalsi K, Fuller J, et al. Aggregate model for heterogeneous thermo-
[9] Tascikaraoglu A, Erdinc O, Uzunoglu M, et al. An adaptive load dispatching and statically controlled loads with demand response. Power and Energy Society
forecasting strategy for a virtual power plant including renewable energy conver- General Meeting. IEEE; 2012. p. 1–8.
sion units. Appl Energy 2014;119(15):445–53.
[10] Nosratabadi SM, Hooshmand RA, Gholipour E. Stochastic profit-based scheduling of

670