Benchmarking performance of machine-learning methods for building energy demand modelling

Abstract

The relevance, relative importance and co-linearity of input parameters to the results of building energy demand forecasts were investigated. Two calendar years of historical data including weather variables and days of week were used. The study also aimed to assess the performance of multiple-linear-regression, support-vector-machine and artificial-neural-network models for predicting daily heating, ventilation and air-conditioning energy consumption of a commercial building in France. Mean absolute error, root mean square error and coefficient of variation of root-mean squared error were selected as the performance criteria. The results showed that the best performance was achieved via the artificial-neural-network model according to all performance measures. In addition, the other two models were not able to meet the predicting requirements for energy consumption in a building since their coefficient-of-variation-o-root-mean-squared error values were not below 30%. The results also indicated that there was multiple co-linearity between the number of degree days and outdoor temperature. Furthermore, the most significant parameter on daily energy consumption was found to be the number of degree days, followed by global radiation, sunshine rate and the day of the week, respectively.