Evaluating Feature Selection Methods for Short-term Load Forecasting

Identifying an appropriate set of predictors is important for making efficient and accurate forecasting models. In this paper, we study the application of some feature selection methods for prediction of household energy consumption. This study follows a two-stage framework. First, it identifies candidate features based on literature study and data characteristics of a load profile and then it selects a subset of relevant features using four different feature selection methods; F-regression, Mutual Information, Recursive Feature Elimination and Elastic Net. We evaluate the effectiveness of these methods, in conjunction with an ensemble-based prediction algorithm (Gradient Boosted Regression Tree), using smart meter data of 23 houses in Norway. To study the performance of these methods for different load profiles, we grouped households into clusters of similar consumption behaviour and computed average performance of each mechanism over clusters' members. Test results show that all feature selection methods could identify a custom-made subset of highly relevant features for each household. Across all clusters, building predictive models utilizing feature selection techniques led to considerable improvements in training speed and simplicity, as well as comparable prediction accuracy with models without feature engineering.