Accuracy of different machine learning algorithms and added-value of predicting aggregated-level energy performance of commercial buildings

As with many other sectors, to improve the energy performance and energy neutrality requirements of individual buildings and groups of buildings, built environment is also making use of machine learning for improved energy demand predictions. The goal of achieving energy neutrality through maximized use of on-site produced renewable energy and attaining optimal level of energy performance at building-cluster level requires reliable short term (resolution shorter than one day) energy demand predictions. However, the prediction and analysis of the energy performance of buildings is still focused on the individual building level and not on small neighborhood scale or building clusters. In a smart grid context, to better understand electricity consumption at different spatial levels, prediction should be at both individual as well as at building-cluster levels, especially for neighborhoods with definite boundaries (such as universities, hospitals). Therefore, in this paper, using data from 47 commercial buildings, a number of machine learning algorithms were evaluated to predict the electricity demand at individual building level and aggregated level in hourly intervals. Predicting at hourly granularity is important to understand short-term dynamics, yet most of the neighborhood scale studies are limited to yearly, monthly, weekly, or daily data resolutions. Two years of data were used in training the model and the prediction was performed using another year of untrained data. Learning algorithms such as; boosted-tree, random forest, SVM-linear, quadratic, cubic, fine-Gaussian as well as ANN were all analysed and tested for predicting the electricity demand of individual and groups of buildings. The results showed that boosted-tree, random forest, and ANN provided the best outcomes for prediction at hourly granularity when metrics such as computational time and error accuracy are compared.