Efficient Feature Selection Strategy for Accurate Electricity Demand Forecasting

Accurate electric load prediction offers an important input information for various smart decisions in energy systems. This paper focuses on short-term forecasting of the electric load of small-scale decentralized energy systems (buildings, energy communities, microgrids, virtual power plants, local energy internets, etc.), which are newly evolving energy system models. Quite few researchers have done feature selection before training forecast models, which is an important preprocessing task of data mining and broadly practiced for knowledge exploration in expert and intelligent systems. This paper proposes a feature selection strategy to find the most significant and non-repetitive variables for accurate short-term electric load prediction in distributed energy systems in general and buildings in particular. In the devised strategy, Binary Genetic Algorithm (BGA) is employed for the variable selection task and Gaussian Process Regression (GPR) is applied for quantifying the fitness score of the variables. The proposed feature selection strategy is implemented and validated using actual electricity consumption data of various buildings located in Otaniemi area of Espoo, Finland. The results are compared with those obtained by other predictor selection methods and show outperformed performances.