How many watts: A data driven approach to aggregated residential air-conditioning load forecasting

Abstract

Due to the significant contribution of air-conditioning load towards total energy consumption in residential buildings, accurate modelling and forecasting of such load is key to effective demand-side energy management programmes. This paper suggests a data driven framework for 15 min-ahead AC load forecasting based on modern machine learning techniques that includes Support Vector Regression, Ensemble Trees, and Artificial Neural Network. To the end, it utilizes a correlation based feature selection method to identify information that is relevant for machine learning modelling. The effect of spatio-temporal features selection on prediction output and the effect of training data quantity on convergence characteristics were analysed and discussed. The effectiveness of the proposed approach is evaluated using a 20-household, half-year data set from an ongoing research testbed set up at the faculty housing units of Singapore University of Technology and Design. An linear combination method was proposed to combine models and the resulting model gave a mean absolute percentage error of 11.27%.