A Holistic Feature Selection Method for Enhanced Short-Term Load Forecasting of Power System

Short-term load forecasting (STLF) is important for the operational security and economics of power system. However, most of the STLF methods lack an efficient feature selection method to model the time series nonlinearities and feature interaction. In this article, a new holistic feature selection method is presented. The feedforward long short-term memory (F-LSTM) network is proposed to learn the nonlinear mapping function between features and load. Then, a feature importance matrix is designed to reflect relevancy, redundancy, and interaction among the candidate features. Moreover, a hybrid filter-wrapper approach is developed to select suitable features efficiently. The filter part separates useless information for the trained F-LSTM output. The wrapper part selects the optimal subset by fine-tuning the threshold. The results from an empirical study in Switzerland suggest that: 1) the selected subset of features shows high relevancy, low redundancy, and high interaction, which is also consistent with the features selected by other feature selection methods, and 2) the proposed method has good prediction performance and can be applied to various artificial neutral network-based STLF models, which delivers an average 12.1% improvement.