Faculty of Economics and Clare College, University of Cambridge, UK.
Causal Tree Estimation of Heterogeneous Household Response to Time-Of-Use Electricity Pricing
Schemes Jinhua Wang and Melvyn Weeks
We examine the distributional effects of the introduction of Time-of-Use (tou) pricing schemes where the price per kWh of electricity usage depends on the time of consumption. These pricing schemes are enabled by smart meters, which can regularly (i.e. half-hourly) record consumption. Using causal trees, and an aggregation of causal tree estimates known as a causal forest (Athey & Imbens 2016, Wager & Athey 2017), we consider the association between the effect of TOU pricing schemes on household electricity demand and a range of variables that are observable before the introduction of the new pricing schemes. Causal trees provide an interpretable description of heterogeneity, while causal forests can be used to obtain individual-specific estimates of treatment effects. Given that policy makers are often interested in the factors underlying a given prediction, it is desirable to gain some insight to which variables in this large set are most often selected. A key challenge follows from that fact that partitions generated by tree-based methods are sensitive to subsampling, while the use of ensemble methods such as causal forests produce more stable, but less interpretable estimates. To address this problem we utilise variable importance measures to consider which variables are chosen most often by the causal forest algorithm. Given that a number of standard variable importance measures can be biased towards continuous variables, we address this issue by including permutation-based tests for our variable importance results. Keywords: Machine learning, TOU tariffs, Smart metering, Household electricity demand.
Melvyn Weeks is Assistant Professor in Economics at the University of Cambridge and a Fellow of Clare College Cambridge. His PhD is from the University of Pennsylvania, with fields in microeconomics and econometrics. Melvyn’s research interest spans both theoretical and applied microeconometrics including policy evaluation, understanding behaviour over discrete choice; modelling demand systems in empirical industrial organisation; revealed and stated preference models; model testing and evaluation; and computationally intensive methods including machine learning, simulation-based inference and the bootstrap. Recently, Melvyn has used a combination of Machine Learning approaches and econometrics to understand relationships that exist within large and complex datasets, avoiding the well known multiple testing problem. He is currently pursuing a number of projects examining the impact of time-of-use pricing in the electricity market. Assisted by a number of PhD students, this work combines aspects of econometric and machine learning tools to explore the extent to which households respond to intra-day variation in prices, and specifically how demand response varies according to both demographic characteristics and attributes of load demand proles. As an extension of this work, Melvyn is an advisor to a Cambridge-based company, Smart Meter Analytics Platform, which seeks to accelerate the utilisation of customer-level energy data by providing a cutting edge and ready-to-deploy data analytics service to energy companies. In addition his research involves the development and application of models of discrete choice.