Evaluation of Machine Learning Algorithms for Demand Response Potential Forecasting

Abstract

This paper focuses on the ability of machine learning algorithms to capture the demand response (DR) potential when forecasting the electrical demand of a commercial building. An actual sports-entertainment centre is utilised as a testbed, simulated with Energy-Plus, and the strategy followed during the DR event is the modification of the chiller water temperature of the cooling system. An artificial neural network (ANN) and a support vector machine (SVM) predictive model, are utilised to predict the DR potential of the building, due to the significant amount of execution time of the EnergyPlus model. The data-driven models are trained and tested based on synthetic databases. Results demonstrate that both ANN and SVM models can accurately predict the building electrical power demand for the scenarios without or with daily DR events, whereas both predictive models are not accurate in forecasting the electrical demand during the rebound effect.