Modeling and Forecasting Hourly Electricity Demand by SARIMAX with Interactions

This paper presents an interaction forecasting framework with a focus on short-term load forecasting. It proposes a seasonal autoregressive integrated moving average model with the inclusion of exogenous variables (SARIMA: main effects) and interaction variables (cross effects) to forecast short-term electricity load using hourly load data from Tokyo Electric Power Company. The main effects and cross effects are measured through an iterative process of plotting, interpreting, and testing. Interactions of weather variables and calendar variables, as well as interactions of seasonal patterns and intraday dependencies, are analyzed, tested, and added to the model. We compare the SARIMAX model, which contains only main effects, with the Interaction-SARIMAX model, which includes cross effects in addition to the main effects. Our proposed SARIMAX-with-interactions model is shown to produce smaller errors than its competitors. Thus, including interaction effects of the exogenous variables into the SARIMAX model can potentially improve the model forecasting performance. Although the model is built using data of a specific region in Japan, the method is completely generic and therefore applicable to any load forecasting problem.