IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2017 1 Aggregating a Large Number of Residential Appliances for Demand Response Applications Fadi Elghitani, and Weihua Zhuang, Fellow, IEEE Abstract—Current demand response (DR) programs focus on industrial consumers as they can provide a large magnitude of demand modification. In order to extend DR programs to the residential sector, aggregating service demands from a large number of residential consumers is necessary in order to achieve a sensible benefit to the power network. In this paper, we propose a methodology for residential demand aggregation, based on a multi-class queueing system. Each class represents demand blocks of a specific power level, time duration, and a service delay requirement. We use this model to minimize the cost of the appliances’ aggregated power consumption under day-ahead pricing (DAP). Using realistic appliances’ data, we show that the proposed framework achieves a cost reduction that is close to the best achievable one. Index Terms—Demand response, Residential appliances, Ag- gregation Model, Cutting-plane method. I. I NTRODUCTION T HE operation of electric power systems is traditionally based on demand following. The role of supply-demand matching of the electric energy is assigned to the generation- side of a power network, while the demand-side is assumed to be non-controllable and has to be satisfied regardless of its cost. Demand following can be very expensive, especially during periods of peak demand, when the least efficient gen- erators have to be activated to satisfy the increase in demand [1]. Therefore, utility companies began to realize the potential benefits of controlling consumers’ demands, referred to as demand-side management (DSM), which contrasts traditional generation-side decisions. A specific type of DSM is demand response (DR), which seeks demand modification via different financial incentives for the consumers. The residential sector represents a significant part in electric energy demand (e.g., 32.9% of total demand in Ontario [2]). As a result, it is desired to incorporate the residential sector to various DR applications. Residential consumers can contribute to DR via two categories of appliances: deferrable loads and thermostatically-controlled loads (TCLs). Deferrable loads can be controlled to defer their energy consumption to a future time, but should be satisfied within a certain deadline to avoid consumer’s inconvenience. Examples of deferrable loads in- clude washing machines and dish washers. On the other hand, TCLs represent appliances which control the temperature of some compartments. TCL examples include refrigerators, air- conditioners, and electric water heaters. Many studies on residential DR are related to Home Energy Management (HEM) applications, where appliances within a The authors are with the Department of Electrical and Computer En- gineering, University of Waterloo, Waterloo, ON, Canada. (e-mail: fel- ghita@uwaterloo.ca; wzhuang@uwaterloo.ca) single household are scheduled to minimize the resident’s elec- tricity bill [3]–[8]. Since on average a residential consumer has a relatively few number of controllable appliances, individual demand models for appliances can be used in the optimization problem. However, applying algorithms developed for HEM to schedule the appliances of several consumers is difficult due to increased computational complexity. Hence, aggregation de- mand models are needed to represent the demands of different groups of appliances, in order to reduce the DR problem size. In existing studies, separate aggregation models are used for both deferrable and TCL appliances. For deferrable appliances, aggregation model is based on only the energy requirements of different appliances, without taking account of power consumption profiles [9]–[14]. The aggregation is done by simply summing up all the energy requirements from all appliances. The limitations of this model come from two assumptions: 1) all appliances require the same class of delay performance, which is not accurate due to the heterogeneity of the appliance types and consumer preferences; and 2) the demand is perfectly elastic, which can be satisfied within a small time or over a long time as desired. The model is not suitable for scheduling TCLs as there is no information about the controlled temperature. On the other hand, aggregation models for TCLs in existing studies are based on classifying TCLs according to their current temperature and its (ON/OFF) operation status [15]–[17]. The system state is represented by the probability mass function (PMF) over these different bins. An advantage is that the accuracy of the model depends on the resolution of the PMF, but not on the number of aggregated TCLs. The control signal represents the switching probabilities (turning ON or OFF) that are submitted to each group of TCLs residing in the same bin. Since each bin represents a state variable, the system state-space can be huge for a given temperature range. Therefore, it is difficult to use the aforementioned aggregation model for optimization. Nevertheless, the model is used for evaluating the performance of a large number of TCLs. Different from the existing studies, our proposed methodol- ogy works for both deferrable and TCL appliances. Demands from a heterogeneous population of appliances is aggregated into a set of pre-specified classes of demand. This multi-class approach allows an efficient utilization of DR resources while ensuring consumers’ convenience, in contrast to restricting demand aggregation of controllable appliances into a single class. Also, demands can be differentiated according to their energy consumption durations, which allows us to deduce with high accuracy the controllable power consumption profile resulting from DR decisions.