Inverse Reinforcement Learning Reveals Electricity Consumption Behavior Shifts

Understanding the complex ways households consume electricity, particularly how they react to socioeconomic changes and climatic events, is crucial for effective energy policy design. Traditional models often oversimplify this behavior, especially the non-linear responses to thermal stress influenced by income, habits, and the built environment. This research introduces a novel approach by treating households as intelligent agents interacting within their environment, applying Inverse Reinforcement Learning (IRL) to infer their consumption behavior through model-implied reward functions. The study specifically investigates how these inferred reward functions evolve when households experience significant shocks, such as the energy crisis and heatwaves between 2021 and 2023. Applied to different electricity consumption clusters in Italy, the framework revealed that cooling behaviors were reshaped heterogeneously across consumer groups, with responses conditioned by their existing habits and living conditions. The findings identified a spectrum of responses, from temporary adjustments to durable shifts, and even groups showing negligible change. Furthermore, the research highlighted that the timing of consumption (intradaily scale) is an independent dimension of behavioral heterogeneity, suggesting that energy policies should consider not just who consumers are and where they live, but also when they consume and the persistence of their behavioral changes.