In a demand response (DR) market run by independent system operators (ISOs), load aggregators are important market participants who aggregate small retail customers through various DR programs. A load aggregator can minimize the allocation cost by efficiently allocating its demand response resources (DRRs) considering retail customers' characteristics. However, the uncertain response behaviors of retail customers can influence the allocation strategy of its DRRs, increasing the economic risk of DRR allocation. This paper presents a riskbased DRR allocation method for the load aggregator that takes into account not only the physical characteristics of retail customers but also the risk due to the associated response uncertainties. In the paper, a conditional valueatrisk (CVaR) is applied to deal with the risk due to response uncertainties. Numerical results are presented to illustrate the effectiveness of the proposed method.
1. Introduction
As demand response (DR)
[1]
has been attracting increasing attention as a means to control the electricity demand efficiently, load aggregators have begun to play an important role in the DR market. Load aggregators join together retail customers and participate in the DR market on their behalf. Furthermore, load aggregators operate DR programs for retail customers to help them manage their loads effectively during a power system emergency or a lack of power supply. The objective of load aggregators is to maximize their profits. In order to achieve this goal, it is imperative for them to establish a demand response resource (DRR) allocation strategy. When load aggregators set up such a strategy, they should consider not only retail customers’ physical characteristics but also the economic risk due to the uncertainty of retail customers’ response behaviors.
During the past years, various studies related to DR
[2

7]
have been conducted. DR is studied with new equipment in a smart grid environment, i.e., renewable energy resources
[2]
, electric vehicles
[3]
, smart appliances
[4]
, and so on. Nevertheless, these studies did not consider the physical characteristics of customers. Although reference
[5]
considers the physical characteristics of the DR participants, it does not take into account the economic risk due to the uncertainty of DR participants’ response behaviors. The risk associated with the uncertainties in load reductions is considered by load serving entities (LSEs) in
[6]
. However, the research in
[6]
is not related to the allocation of DRRs for retail customers from the load aggregators’ perspective. A concept of DRR allocation is described in
[7]
, but it depends only on DR participants’ opinion without taking into account retail customers’ physical characteristics and the economic risk due to their response uncertainties.
This paper presents a riskbased DRR allocation method that incorporates both retail customers’ physical characteristics and the economic risk associated with the uncertainties in load reductions. The risk makes it difficult to determine the allocation of DRRs because it increases the expected economic losses in a DRR allocation. In order to minimize the risk due to the uncertainty of load reductions in the DRR allocation, it is essential for the load aggregator to assess the risk. The conditional valueatrisk (CVaR) is employed in this study to deal with the economic risk due to the uncertain response behaviors of retail customers. In addition, the DR participation rate of the retail customer is proposed in this study to calculate the CVaRs of retail customers. The mixedinteger nonlinear programming (MINLP) is also applied in this paper to solve the DRR allocation problem of the load aggregator.
2. Problem Formulation
The DR market is usually opened by an independent system operator (ISO) the day before a power system emergency or a lack of power supply
[8]
. After the ISO informs DR market participants about the total duration of DR event, participants enter load reduction bids in the DR market. Some participants such as large customers can take part in the DR market directly, but the participation of retail customers is restricted by the ISO. This is because only a customer with a larger load reduction capacity than prespecified minimum load can participate in the DR market. In order to involve small retail customers in the DR market, new entities called load aggregators, who bring together retail customers and take part in the DR market on their behalf, appear in the DR market. Moreover, load aggregators operate DR programs for retail customers to help them manage their loads efficiently.
After the load reduction bidding ends, the ISO clears the DR market and notifies the DR event information on the scheduled DR amounts and DR market prices to participants. Based on the DR event information notified by the ISO, load aggregators allocate the scheduled DR amounts to retail customers; that is, load aggregators establish the DRR allocation strategy before the DR event takes place. Because the sum of the actual amounts of load reductions of retail customers expects to be equal to the overall load reduction achievement of load aggregators, the appropriate DRR allocation strategy is needed for load aggregators to meet the scheduled DR amounts in the DR event. However, the uncertain response behaviors of retail customers affect the DRR allocation of load aggregators because uncertainties in load reductions cause an economic risk for them. Thus, load aggregators need to assess the risk associated with the uncertainties in load reductions and consider both retail customers’ physical characteristics and the risk due to response uncertainties when they allocate DRRs to retail customers.
In the DR event, load aggregators indirectly reduce loads by operating DR programs for retail customers while large customers fulfill load reductions directly. After the DR event, a load aggregator gets paid from ISO based on the actual amounts of load reductions of each DR participant. Further, load aggregators calculate the financial incentives for the actual amounts of load reduction achieved by each retail customer.
In this study, the total amount of load reductions of the load aggregator expects to be equal to the DR amount scheduled by the ISO. Then, the profit maximization can be achieved by allocation cost minimization under the assumption that the scheduled DR amount expects to be met by the load aggregator and its revenue from ISO remains constant. The DRR allocation problem of the load aggregator can be formulated as the following Eqs. (19). The objective function minimizes the allocation cost of the load aggregator as follows:
where
T
is the total duration of the DR event;
N
is the number of retail customers;
is the incentive of the DR programs of the load aggregator at time
t
;
DR_{it}
is the DRR allocated to retail customer
i
at time
t
;
u_{it}
is the status of retail customer
i
at time
t
(i.e., its value is 1 if the DRR is allocated and 0 otherwise).
The various constraints on the characteristics of retail customers are as follows
[5
,
9
,
10]
:
a) The DRR balance equation:
where
DR_{ISO,t}
is the DR amount scheduled by the ISO at time
t
. Eq. (2) states that the sum of the DRRs allocated to each retail customer needs to be equal to the DR amount scheduled by the ISO at time
t
.
b) The minimum and maximum DRR limits:
where
is the minimum DRR limit of retail customer
i
and
is the maximum DRR limit of retail customer
i
. In (3), the minimum DRR limit defines the lowest limit of load reductions that the retail customer should fulfill in the DR event. Likewise, the retail customer should reduce loads under the maximum DRR limit in the DR event according to (4).
c) The up and down maximum ramping rate DRR limits:
where
RU_{i}
is the up maximum DRR ramping rate of retail customer
i
;
RD_{i}
is the down maximum DRR ramping rate of retail customer
i
;
Δt
is the dispatch time interval. The up and down maximum DRR ramping rates of retail customers represent the ramping capacity for reducing and consuming electricity, respectively. These measures identify the rate at which a retail customer would change its loads.
d) The minimum up and down time DRR limits:
where
MU_{i}
is the minimum up time DRR limit of retail customer
i
and
MD_{i}
is the minimum down time DRR limit of retail customer
i
. The minimum up time DRR limit defines the minimum consecutive hours for which the retail customer should continue to reduce its load in the DR event. Likewise, the minimum down time DRR limit is the minimum consecutive hours for which the retail customer should stop reducing its load in the DR event.
e) CVaR of the load aggregator:
where
CVaR^{max,β}
(
Cost_{LA,t }
) is the maximum CVaR of the allocation cost of the load aggregator within a given confidence level of
β
(
β
∈[0,1] ) at time
t
;
is the load aggregatordefined coefficient within a given confidence level of
β
at time
t
. The load aggregatordefined coefficient indicates the affordable risk level established by the load aggregator
[11]
. If the load aggregator’s risk preference is high, the load aggregatordefined coefficient is assigned a high value, whereas if the load aggregator is risk averse, the load aggregatordefined coefficient is assigned a low value.
3. RiskBased DRR Allocation Method
 3.1 CVaR constraint for DRR allocation
The economic risk in the DRR allocation strategy of the load aggregator is caused by the uncertainty of retail customers’ response behaviors. In order to handle the uncertainty associated with the DRR allocation, risk assessment techniques can be used for the load aggregator to determine the DRR allocation strategy. The valueatrisk (VaR) is widely employed as a simple risk assessment. The VaR, the expected maximum loss within the given confidence level
β
over the holding period, can be formulated by the following Eq. (12, 13):
where
Prob
(
R
≥
x
) is the probability distribution of the returns
R
in the highest bound.
The VaR tends to ignore potential losses exceeding the VaR value. As an alternative measure of VaR, the CVaR is a useful technique for accurate risk assessment. The CVaR is defined as the weighted average of the losses exceeding the VaR value
[14]
. Using the computed VaR, the expected CVaR is calculated by the following Eq. (13):
The CVaR has a relatively larger value than the VaR. The CVaR can also measure the tail of the distribution more accurately than the VaR can
[15]
.
In the DRR allocation strategy, the CVaR of the allocation cost of the load aggregator within the confidence level
β
at time
t
can be calculated as follows:
where
p_{i }
is the DR participation rate of retail customer
i
. With the subadditivity property of the CVaR, the righthandside term of (12) can satisfy the following equation:
Accordingly, the maximum CVaR of the allocation cost of the load aggregator within the given confidence level
β
at time
t
is equal to the righthandside term of (13). Using the (positive) homogeneity of the CVaR, the righthandside term of (13) can be expressed as follows:
Based on (12)(14), the CVaR constraint (9) can be rewritten as follows:
Therefore, the load aggregator needs to assess the CVaRs of retail customers for the solution of the DRR allocation problem, taking into account the economic risk due to response uncertainties.
 3.2 DR participation rate of retail customer
In this section, retail customers’ DR participation rates are introduced for solving the DRR allocation problem considering the economic risk. In this study, the DR participation rate of a retail customer is proposed to represent the uncertainty of response behaviors of retail customers. The DR participation rate of retail customer
i
is given by:
where
is the DR amount allocated by the load aggregator to retail customer
i
at time
t
in the past DR event and
is the actual amount of load reduction by retail customer
i
at time
t
in the past DR event. Based on the DR participation rate of the retail customer, the probability distributions of the DR participation rates of retail customers can be obtained to assess the CVaRs of retail customers.
Fig. 1
illustrates the load aggregator’s DRR allocation strategy based on the CVaR.
Load aggregator’s DRR allocation strategy based on the CVaR.
 3.3 CVaR constrained DRR allocation method
In this subsection, the MINLP is applied to solve the DRR allocation problem. The procedure of the proposed method can be summarized as follows:

Step 1) Calculate the DR participation rates of retail customers using the historical DR amounts scheduled by the load aggregator and the historical amounts of actual load reductions of retail customers.

Step 2) Obtain the probability distribution of the DR participation rates of retail customers.

Step 3) Calculate the CVaRs of retail customers within the given confidence level.

Step 4) Determine the load aggregatordefined coefficient within the given confidence level at timetin the DR event.

Step 5) Solve the DRR allocation problem subject to the constraints of the physical characteristics and CVaRs of retail customers.
The overall procedure of the proposed method is illustrated in
Fig. 2
.
Overall procedure of the proposed method.
4. Numerical Example
In this section, the results of a numerical example are presented to validate the effectiveness of the proposed method. Retail customers’ maximum load reduction capacities, the historical DR amounts scheduled by the ISO, and the historical amounts of actual load reductions of retail customers were obtained from the Korea Power Exchange (KPX). It is assumed in this study that the load aggregator operates two DR programs, and retail customers participate in only one DR program. Retail customers who take part in the DR programs of the load aggregator voluntarily reduce their loads in case of the DR event. The load aggregator pays incentives to these retail customers without imposing the financial penalties of DR programs.
In order to establish the optimal DRR allocation strategy, the load aggregator needs to consider the DR participation rates of retail customers as well as the physical characteristics of retail customers, which are listed in
Table 1
.
Physical characteristics and DR participation rates of retail customers
Physical characteristics and DR participation rates of retail customers
In
Table 1
, nine retail customers are assumed to participate in the DR programs of the load aggregator. All these retail customers have maximum load reduction capacities that are lower than 300 kW. Because only entities that can provide a load reduction capacity larger than 300 kW are assumed to participate directly in the DR market, the retail customers in
Table 1
should participate in the DR market indirectly through the DR programs of the load aggregator
[8]
. Physical characteristics of retail customers are arbitrarily selected in
Table 1
. It is also assumed in this study that the probability distributions of the DR participation rates of retail customers follow a normal distribution, and the standard deviations of probability distributions, i.e., the volatilities of the load reductions of retail customers, are 2% of the means of the DR participation rates of the retail customers.
Table 2
shows some of the information given to the participants when the ISO notifies a DR event in the DR market.
DR event information given by the ISO
DR event information given by the ISO
In
Table 2
, DR market prices are obtained from KPX data on May 22, 2012. The unit time of the DR event is 30 minutes, so the DR event in
Table 2
proceeds for 3 hours
[8]
. Furthermore, DR market prices differ depending on the times of the DR event in
Table 2
.
Based on the DR market prices in
Table 2
, the load aggregator can determine the incentives of the DR programs.
Table 3
shows the incentives of the DR programs of the load aggregator.
Incentives of load aggregator’s DR programs (won/kWh)
Incentives of load aggregator’s DR programs (won/kWh)
In
Table 3
, it is assumed that the load aggregator set the incentives of DR programs A and B as 50% and 40% of the DR market prices, respectively.
In order to evaluate the effectiveness of the proposed method, this paper introduces a baseline method and compares the outputs of the proposed method with those of the baseline method. In the baseline method, the load aggregator establishes the DRR allocation strategy considering the physical characteristics of retail customers, but the CVaR of the load aggregator is not taken into account. In addition, the proposed method is divided into two confidence levels: 97.5% and 99.0%. On the basis of
[16
,
17]
, this study estimates the CVaRs of the DR participation rates of retail customers when the confidence levels are 97.5% and 99.0%, individually.
In this study, the load aggregatordefined coefficient in (9) is set as follows:
where
ω_{LA}
is the risk tolerance level of the load aggregator and
MP_{t}
is the DR market price at time
t
. Eq. (17) implies that the average allocation cost of the load aggregator in
β
% of the worst cases must not exceed
ω_{LA}
times the income of the load aggregator at time
t
.
Table 4
shows the results of the DRR allocation of the load aggregator when the risk tolerance level is fixed as 0.3. In
Table 4
, regardless of the method type, retail customers 69 receive more allocated DRRs from the load aggregator than retail customers 15. This is because retail customers 69 participate in DR program B, whose incentives are lower than DR program A. Thus, the incentives of the DR programs of the load aggregator have a great influence on the allocation of DRRs to retail customers under various constraints. In
Table 4
, the total DRRs allocated to retail customers 2, 6, 7, and 8 with the proposed methods within all the confidence levels are greater than those with the baseline method. However, retail customers 3 and 9 receive less total DRRs with the proposed methods than with the baseline method, as shown in
Table 4
. This is why retail customers 6, 7, and 8 have lower CVaRs than retail customer 9 among the retail customers who participated in DR program B. Likewise, retail customer 2 has a lower CVaR than retail customer 3 among the retail customers who take part in DR program A. Consequently,
Table 4
implies that the load aggregator allocates DRRs to retail customers who have low CVaRs in the DRR allocation strategy
Results of DRR allocation (kW)
Results of DRR allocation (kW)
Fig. 3
illustrates the changes in the allocation costs as the risk tolerance level increases in the CVaR constraint of the proposed methods within 97.5% and 99.0% confidence levels.
Changes in the allocation costs as the risk tolerance level increases.
Moreover, in addition to the allocation costs of the load aggregator with the proposed methods, the difference percentages of the allocation costs between the baseline method and proposed methods when the risk tolerance level increases are listed in
Table 5
.
Allocation cost with proposed method and difference percentage of allocation cost between baseline method and proposed method
Allocation cost with proposed method and difference percentage of allocation cost between baseline method and proposed method
In
Fig. 3
and
Table 5
, the allocation cost of the load aggregator in the baseline method is fixed at 2,163,751 won.
Fig. 3
and
Table 5
show that all the allocation costs with the proposed methods are equal to or higher than the allocation costs with the baseline method. This result implies that the CVaR constraint influences in allocation costs of the load aggregator in the DRR allocation strategy. In addition, higher risk tolerance levels tend to lower the allocation costs of the load aggregator with the proposed method within all the confidence levels, as shown in
Fig. 3
and
Table 5
. This is why the higher the risk tolerance level the load aggregator sets up, the greater the economic risk the load aggregator can take. In
Fig. 3
and
Table 5
, the allocation costs when the confidence level is 99.0% are generally higher than the allocation costs when the confidence level is 97.5%. Because the CVaRs of retail customers increase as the confidence level increases, the allocation costs within 99.0% confidence level are more influenced by the CVaR constraint than the allocation costs within 97.5% confidence level. Unexpectedly, the allocation costs within 97.5% confidence level are higher than the allocation costs within 99.0% confidence level for the risk tolerance levels 0.6 and 1.0. The reason for this difference is that other physical constraints in addition to the CVaR constraint also influence the DRR allocation strategy of the load aggregator. In
Fig. 3
and
Table 5
, the risk tolerance level begins from 0.3 because the load aggregator cannot establish the DRR allocation strategy when the risk tolerance levels are 0.1 and 0.2. This result implies that risk tolerance levels that are too low to restrict the DRR allocation of the load aggregator in the DR event. The results of this numerical example show that the risk preference of the load aggregator has an influence on the allocation of DRRs. The proposed riskbased DRR allocation method can be useful as a decisionmaking tool to help the load aggregator allocate DRRs to retail customers based on its risk preference
5. Conclusion
This paper presents a riskbased DRR allocation method that considers CVaRs as well as the physical characteristics of retail customers. In the paper, the DR participation rate of the retail customer is proposed to evaluate the uncertainties of load reductions. Further, CVaR is used for assessing the economic risk associated with the response uncertainties of retail customers. The numerical example shows that the proposed riskbased DRR allocation method can be used to help the load aggregator establish a DRR allocation strategy that considers the risk due to the response uncertainties of retail customers.
Acknowledgements
This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology (No. 2013R1A1A2057969). This work was also supported by the Human Resources Development program (No. 20134030200340) of the Korea Institute of Energy Technology Evaluation and Planning(KETEP) grant funded by the Korea government Ministry of Trade, Industry and Energy.
BIO
JiHui Kim She received B.S. degree from Korea University, Seoul, Republic of Korea, in 2008. She is currently pursuing Ph.D. degree in Electrical Engineering at Korea University.
Jaehee Lee He received Ph.D. degree from Korea University, Seoul, Republic of Korea, in 2014. He is currently an employee of the Kepco Economic & Management Research Institute in Korea.
SungKwan Joo He received the M.S. and Ph.D. degrees from the University of Washington, Seattle, in 1997 and 2004, respectively. From 2004 to 2006, he was an Assistant Professor in the Department of Electrical and Computer Engineering at North Dakota State University, Fargo, U.S. He is currently an Associate Professor in the School of Electrical Engineering at Korea University, Seoul, Korea.
2012
“Assessment of Demand Response and Advanced Metering: Staff Report,”
Dietrich K.
,
Latorre J. M.
,
Olmos L.
,
Ramos A.
2012
“Demand Response in an Isolated System with High Wind Integration,”
IEEE Trans. Power Syst.
27
(1)
20 
29
DOI : 10.1109/TPWRS.2011.2159252
Soares J.
,
Morais H.
,
Sousa T.
,
Faria P.
2013
“Day Ahead Resource Scheduling Including Demand Response for Electric Vehicles,”
IEEE Trans. Smart Grid
4
(1)
596 
605
Yi P.
,
Dong X.
,
Iwayemi A.
,
Zhou C.
,
Li S.
2013
“RealTime Opportunistic Scheduling for Residential Demand Response,”
IEEE Trans. Smart Grid
4
(1)
227 
234
DOI : 10.1109/TSG.2012.2225155
Khodaei A.
,
Shahidehpour M.
,
Bahramirad S.
2011
“SCUC with Hourly Demand Response Considering Intertemporal Load Characteristics,”
IEEE Trans. Smart Grid
2
(3)
564 
571
Joo J. Y.
,
Ilic M.
2012
“Distributed MultiTemporal Risk Management Approach to Designing Dynamic Pricing,”
IEEE Power & Energy Society General Meeting
California, USA
Shao S.
,
Pipattanasomporn M.
2011
“An Approach for Demand Response to Alleviate Power System Stress Conditions,”
IEEE Power & Energy Society General Meeting
Michigan, USA
2013
“Demand Response Market Operation Rules,”
Ministry of Knowledge Economy in Korea
Fu Y.
,
Shahidehdour M.
,
Li Z.
2005
“Security Constrained Unit Commitment with AC Constraints,”
IEEE Trans. Power Syst.
20
(3)
1538 
1550
DOI : 10.1109/TPWRS.2005.854375
Shahidehpour M.
,
Fu Y.
2005
“Benders decomposition,”
IEEE Power Energy Mag.
3
(2)
20 
21
Bruno S. V.
,
Sagastizabal C.
2008
“Optimization of Real Asset Portfolio Using a Coherent Risk Measure: Application to Oil and Energy Industries,”
Optimization Online
1 
8
Cuthbertson K.
,
Nitzsche D.
2004
Financial Engineering: Derivatives and Risk Management
Wiley
559 
601
Mazaheri M.
2008
“Risk Budgeting Using Expected Shortfall (CVaR): An Overview,”
SSRN
Hull John C.
2006
Options, Futures, and Other Derivatives
PEARSON
New Jersey
435 
437
Fabrian C. I.
2008
“Handling CVaR Objectives and Constraints in TwoStage Stochastic Models,”
Eur. J. Oper. Res.
191
(3)
888 
911
DOI : 10.1016/j.ejor.2007.02.052
Huang Y. H.
2006
“A Comparison of Value at Risk Approaches and a New Method with Extreme Value Theory and Kernel Estimator,” Ph.D. dissertation
Dept. of Economics, CUNY
New York, USA
Tsay R. S.
2010
Analysis of Financial Time Series
(3rd ed.)
Wiley
333 
338