Comparative study of data driven methods in building electricity use prediction

The energy use prediction of building systems is crucial to design a high-efficiency building and maintain low energy consumption operation, which is also important in optimizing building system control and retrofitting. This paper demonstrates a comparative study of four data-driven methods used in online building energy predictions involving large-scale data extracted from several types of buildings. The characteristics of building electricity use and data reliability were addressed through the data pre-treatment process including visualization, cleaning, parsing and filtering. Mathematical algorithms and their applications in previous studies were summarized and compared, and evaluation methods were developed. The performance and suitable application scenarios of the proposed algorithms were conducted via the comparison of monitoring data and predicted results. The study indicates that the most complex method which requires the highest computation ability, i.e., the Artificial Neural Network (ANN), does not lead to the highest accuracy, while as the fastest computation method, Gaussian Process Regression (GPR) usually has the results with the lowest accuracy. Support Vector Machine (SVM) and Multivariate Linear Regression (MLR) methods usually perform better in the case scenarios studied. All the prediction accuracies can meet the requirements of RMSE <30% and NMBE <10% proposed by ASHRAE, and the computation time varies from less than 1 s to 22 s per prediction. All these methods/algorithms worked well for buildings with stable energy use patterns. For buildings with complex and unstable occupancy schedules and energy use patterns, MLR and SVM methods have the ability to achieve a high accuracy with fast computation speed. (C) 2019 Elsevier B.V. All rights reserved.