Data-Driven Based Prediction of the Energy Consumption of Residential Buildings in Oshawa

Abstract

Buildings consume about 40% of the global energy. Building energy consumption is affected by multiple factors, including building physical properties, performance of the mechanical system, and occupants’ activities. The prediction of building energy consumption is very complicated in actual practice. Accurate and fast prediction of the building energy consumption is very important in building design optimization and sustainable energy development. This paper evaluates 24 energy consumption models for 83 houses in Oshawa, Canada. The energy consumption, social and demographic information of the occupants, and the physical properties of the houses were collected through smart metering, a phone survey, and an energy audit. A total of 63 variables were determined, and based on the variable importance, three groups with different numbers of variables were selected, i.e., 26, 12, and 6 for electricity consumption; and 26, 13, and 6 for gas consumption. A total of eight data-driven algorithms, namely Multiple Linear Regression (MLR), Stepwise Regression (SR), Support Vector Machine (SVM), Backpropagation Neural Network (BPNN), Radial Basis Function Neural Network (RBFN), Classification and Regression Tree (CART), Chi-Square Automatic Interaction Detector (CHAID), and Exhaustive CHAID (ECHAID), were used to develop energy prediction models. The results show that the BPNN model has the best accuracies in predicting both the annual electricity consumption and gas consumption, with mean absolute percentage errors (MAPEs) of 0.94% and 0.94% for training and validation data for electricity consumption, and 2.63% and 0.16% for gas consumption, respectively.