A Feature Engineering Strategy for Data-Driven Predictive Models of Educational Building's Electricity Consumption

This paper proposes a feature engineering strategy for data-driven models that predict hourly electricity consumption of educational buildings. The proposed strategy constructs raw features based on datetime and outdoor weather data and then uses Recursive Feature Elimination (RFE) to select improved features. The number of selected features in RFE is determined using ensemble techniques, based on the R2 scoring of K Nearest Neighbours (KNN), Random Forest (RF) and eXtreme Gradient Boosting (XGB) models. The performance of the strategy is evaluated using the four-month electricity consumption data from a building in ASHRAE dataset. Four data-driven models are trained using both raw features and improved features to make predictions. Comparison of their predictive performance, as measured by MAPE, RMSE, MAE and R2, demonstrates that the proposed feature engineering strategy provides the data-driven models with 1) higher predictive accuracy, 2) better elimination of redundant features, and 3) improved sample efficiency.