Variability in automated responses of commercial buildings and industrial facilities to dynamic electricity prices☆
Highlights
► Demand response (DR) baseline models have error so shed estimates have error. ► Buildings exhibit variability in automated responses to DR signals. ► We present a metric to determine the source of observed DR variability. ► Most observed DR variability results from model error, not response variability. ► Response variability in aggregate populations may affect power system performance.
Introduction
Buildings are becoming increasingly important as active resources that support power system operations. Though buildings have played a small role in power systems operations in the past – either with relays that interrupt power to air conditioners and water heaters [1], [2], or by “voice dispatch” of large commercial and industrial loads [3] – recent Smart Grid investments are demonstrating the potential for buildings to become grid-interactive resources that are just as controllable as – or even more controllable than – electricity generators [4].
In “demand response” (DR) programs, power system operators can achieve system-wide demand reductions by providing financial incentives for buildings to change their electricity consumption patterns through both “shifts” in energy use and load reductions, or “sheds.” Buildings generally participate in DR by enrolling in dynamic electricity pricing programs or demand/capacity bidding programs. In dynamic pricing programs, buildings face high electricity prices during hours when the grid is stressed, encouraging them to shed load or shift energy use to less expensive hours. In capacity/demand bidding programs, buildings bid load reductions and, if called upon, shed load at certain times in exchange for payment. In this paper, we focus on commercial buildings and industrial facilities (C&I facilities) participating in a dynamic electricity pricing program. These facilities use the Open Automated Demand Response (OpenADR) Communication Specification [5] to receive DR event notifications from the utility, and during events they automatically execute pre-programmed DR strategies [6].
The central challenge we address in this paper is that DR parameters, such as Average Demand Shed, Rebound, Daily Peak Demand, and Daily Energy (which we define in Table 2), must be measured relative to an estimate of how much electricity a facility would have consumed in the absence of the DR event. DR parameters are computed by subtracting a counterfactual baseline from the actual power consumption of the facility. Therefore, DR parameters will exhibit variability due to both baseline model error and real variability in the facility’s response. We will use the following terms throughout the paper:
- 1.
Unmodeled load variability, or baseline model error, is load variability that is not captured by a baseline model and not due to a DR signal. Unmodeled load variability complicates DR programs that use baselines for financial settlement (e.g., demand/capacity bidding programs and programs in which loads participate in wholesale ancillary services markets). Moreover, even DR programs that do not use baselines for settlement (e.g., dynamic electricity pricing programs) use baselines for Measurement and Verification (M&V) and to calculate the cost-effectiveness of the DR programs [7].
- 2.
Real DR variability is event-to-event variability in a facility’s actual response, for example, due to building managers and/or occupants overriding pre-programmed DR strategies; broken equipment; and variability in responses as a function of occupancy, weather, and other variables.
- 3.
Observed DR variability occurs as a result of the combination of unmodeled load variability and real DR variability.
Fig. 1 illustrates the concepts of baseline model error and observed DR variability. In this figure, we plot the actual and baseline-predicted load for an office building on two DR days and one normal day. The left and middle plots show that responses to DR signals can seem variable – and may, in fact, be variable. The right plot demonstrates baseline model error.
The purpose of this paper is to understand the variability of C&I facility responses to DR events. The question is important for two reasons. First, in order to efficiently allocate generating resources, power system operators must predict how aggregations of facilities will respond on demand response days. If all observed DR variability resulted from unmodeled load variability, a power system operator could expect consistent DR behavior and would only need to deal with the usual amount of demand-side variability. However, if real DR variability is present, the DR program may create an extra burden of variability for the system operator to manage. This could require additional power system services (e.g., reserves). In extreme cases, real DR variability could result in significant deviations in grid frequency or expected power flow.
The second reason variability is important is because DR programs are evaluated on the basis of whether or not facilities (individually or in sum) appear to have reliable responses on DR days. M&V of utility DR programs, including those that do not use baselines for settlement, often include analyses of the DR performance (versus a baseline) of both individual facilities and aggregations of facilities [6]. Variability may affect the evaluation of the DR program and determinations about whether or not a facility is suitable for DR [8]. Moreover, observed DR variability in an individual facility affects how the facility perceives its own DR performance from event-to-event. A facility whose performance seems inconsistent from one event to another may be tempted to modify its DR strategy; however, the perceived inconsistency may have been caused by baseline model error.
Real DR variability is the most relevant measure for power system operators and DR program evaluators. However, real DR variability can only be estimated indirectly, by thoroughly characterizing unmodeled load variability and relating it to observed DR variability. Therefore, in this paper, we first compute the error associated with DR parameter estimates (e.g., demand shed estimates) for 38 C&I facilities that participated in an automated dynamic electricity pricing program in California. We then construct a variability metric that captures the relative importance of real DR variability versus unmodeled load variability, and compute this metric for all of the facilties. We find that most observed DR variability is the result of baseline model error.
A note on terminology: The DR community uses several different terms to denote the counterfactual power usage on DR days: baselines, predictions, and forecasts. In this paper, we use the term ‘baseline predictions’ to refer to ex-post estimates of counterfactual power usage computed with regression parameters (identified with historical demand/temperature data) and actual temperature data for the purpose of M&V. We reserve the term ‘forecast’ for ex-ante estimates computed with forecasted temperature data, which we do not discuss in this paper. We use the term ‘DR parameter estimates’ to refer to values, such as demand sheds, computed with actual demand data and baseline predictions. The DR community often refers to these values as ‘DR calculations’; however, we prefer our terminology because it makes clear that the values are uncertain. The term ‘DR parameter estimates’ should not be confused with ‘DR estimates,’ engineering estimates of expected demand sheds.
The rest of this paper is organized as follows: In Sections 2 and 3, we describe our data and baseline model. In Section 4, we explain our error analysis. Then, in Section 5, we present our results and discussion with respect to baseline model error and DR variability. Lastly, in Section 6, we conclude.
Section snippets
Data
We use 15-min interval whole building electric load data from 38 large C&I facilities (peak demand >200 kW) in California that participated in Pacific Gas and Electric Company’s (PG&E’s) Automated Critical Peak Pricing (CPP) Program between 2006 and 2009. PG&E called CPP DR events on up to 12 summer business days (non-holiday, weekdays) per year when system-wide load was expected to be high, which, in California, usually occurs on hot summer days as a result of commercial and residential air
Baseline model and DR parameters
Electric utilities generally use simple models to determine baseline electric load on DR days for financial settlement and/or M&V. Many of these models involve averaging the daily electric demand over several days (e.g., those with the highest energy usage) before the DR day [11], [12]. Unfortunately, baseline models built by averaging can be biased. Regression-based baseline models, which are less likely to suffer from bias, have long been used for M&V by the energy efficiency community [13],
Error analysis
Most error analyses on regression-based baseline models use the standard errors associated with the regression coefficients [13], [16], [18]. However, these errors underestimate the true error due to a number of issues. First, the regression parameters are correlated. Specifically, time-of-week is correlated to temperature: the highest temperatures tend to occur in the afternoon and the lowest temperatures occur overnight. Second, the regression residuals are autocorrelated. In Fig. 2, we show
DR parameter errors
The error analysis method presented in Section 4.1 allows us to assign error estimates to DR parameter estimates. In Fig. 4, we show DR parameter and error estimates for all 2009 facility-years and the 2009 aggregate population. In most cases, the error estimates are large relative to the DR parameter estimates. For example, on average, across all years, the error associated with Shed 1 is approximately ±120% of the parameter values and the error associated with Shed 2 is approximately ±180% of
Conclusions
We have developed a method to determine the error associated with DR parameter estimates. We find that this error is often large and so DR parameter estimates reported without error estimates may be misleading. For example, we may classify a steady shedder as a variable shedder and, therefore, judge the facility to be poorly controlled when, in fact, baseline model error simply prevents us from measuring consistent sheds. Since DR parameter estimates have error, all calculations derived with
Acknowledgments
We thank Phillip Price, Mary Ann Piette, and Ashok Gadgil for great advice and feedback. We also thank PG&E Company for the electric load data. Johanna Mathieu was funded by a UC Berkeley Chancellor’s Fellowship. Some of this work was conducted at the Lawrence Berkeley National Laboratory under U.S. Department of Energy Contract No. DE-AC02-05CH11231. Sila Kiliccote was funded by the California Energy Commission (CEC) under Contract No. 500-03-026.
References (23)
- et al.
Statistical analysis of baseline load models for non-residential buildings
Energy and Buildings
(2009) PRISM: an introduction
Energy and Buildings
(1986)- et al.
Communication systems for load management
IEEE Transactions on Power Apparatus and Systems
(1985) - et al.
Demand Response Spinning Reserve Demonstration
(2007) - ERCOT Wholesale Market Subcommittee Demand-Side Working Group, Load Participation in the ERCOT Nodal Market, Technical...
- et al.
Achieving controllability of electric loads
- et al.
Open Automated Demand Response Communications Specification (Version 1.0)
(2009) - G. Wikler, I. Bran, S. Prijyandonda, S. Yoshida, K. Smith, M. Piette, S. Kiliccote, G. Ghatikar, D. Hennage, C. Thomas,...
2009 Load Impact Evaluation and Cost Effectiveness Tests of California Statewide Automated Demand Response Programs
(2010)- C.A.E. Consulting, Highly Volatile-Load Customer Study for Southern California Edison, Pacific Gas and Electric...
Introduction to Commercial Building Control Strategies and Techniques for Demand Response
Cited by (0)
- ☆
Based on “Examining Uncertainty in Demand Response Baseline Models and Variability in Automated Responses to Dynamic Pricing” by J.L. Mathieu, D.S. Callaway, and S. Kiliccote which appears in the proceedings of the IEEE Conference on Decision and Control and European Control Conference (CDC-ECC) 2011. ©2011 IEEE.