Abstract
We evaluate the causal effect of hour-ahead price interventions on the reduction in residential electricity consumption using a data set from a large-scale experiment on 7,000 households in California. By estimating user-level counterfactuals using time-series prediction, we estimate an average treatment effect of ~0.10 kWh (11%) per intervention and household. Next, we leverage causal decision trees to detect treatment effect heterogeneity across users by incorporating census data. These decision trees depart from classification and regression trees, as we intend to estimate a causal effect between treated and control units rather than perform outcome regression. We compare the performance of causal decision trees with a simpler, yet more inaccurate k-means clustering approach that naively detects heterogeneity in the feature space, confirming the superiority of causal decision trees. Lastly, we comment on how our methods to detect heterogeneity can be used for targeting households to improve cost efficiency.
Introduction
Grid operators increasingly rely on Demand Response (DR) to maintain grid stability. By incentivizing consumers to reduce power usage during peak demand, operators can avoid activating expensive and carbon-intensive "peaker" plants. However, two primary challenges persist: the Baseline Problem—accurately estimating what a user's consumption would have been without intervention (the counterfactual)—and the Targeting Problem—identifying which specific households are most responsive to signals.
This study utilizes data from a large-scale Randomized Controlled Trial (RCT) conducted by OhmConnect, Inc., involving approximately 7,000 California households serviced by PG&E, SDG&E, and SCE. The dataset spans November 2016 to December 2017 and includes hourly smart meter consumption, local ambient air temperature, and household census data.
Methodology
The research framework consists of a two-step non-experimental estimation process:
- Counterfactual Estimation: To address the baseline problem, the authors benchmarked several regression models (OLS, LASSO, Ridge, KNN, Random Forests) to predict a user's consumption using historical data and weather covariates. Random Forest (RF) emerged as the most accurate, providing the tightest confidence intervals.
- Causal Decision Trees: To identify patterns in who responds, the study utilizes causal trees. Unlike standard decision trees that predict an outcome, causal trees partition the population into groups to maximize the difference in the treatment effect (the reduction) between groups. This approach was validated against a naive $k$-means clustering method, proving that demographic similarity alone does not guarantee similar behavioral responses to incentives.
Results and Analysis
The Average Treatment Effect (ATE) was found to be approximately -0.10 kWh per event per household, representing an average reduction of 11.5%. The Random Forest estimator significantly outperformed industry-standard baselines, which often over-estimate reductions.
Key Insights
- Automation is Critical: Households with at least one smart home device (e.g., a smart thermostat) reduced consumption by 36.7%, whereas non-automated households only reduced by 9.6%.
- Weather Sensitivity: There is a strong positive correlation between ambient temperature and the magnitude of reduction, suggesting that air conditioning is the primary flexible load in residential settings. Reductions were most concentrated in summer months and inland regions.
- Weak Price Elasticity: While higher incentives lead to higher reductions, the correlation is surprisingly weak. This suggests that factors like "moral suasion" or the simple nudge of a notification might be as important as the exact dollar amount.
Conclusion
This work demonstrates that causal decision trees provide a superior method for detecting treatment effect heterogeneity compared to traditional clustering. By identifying highly responsive subpopulations, DR providers can optimize their bidding in wholesale markets and improve cost-efficiency. Future work will focus on integrating these models into real-time dispatch systems and exploring non-monetary drivers of consumer behavior.
Summary updated