Prediction of building electricity usage using Gaussian Process Regression

Abstract

The prediction of building energy use is the basis for smart building operation, which optimizes building performance through control and low-energy strategy. For reducing computation complexity and improving calculation accuracy, a comparative study of online electricity data predictions for different types of buildings was conducted. This study is also intended to assess the capability and accuracy of the supervised machine learning methods, with which the kernel algorithms of predictions were developed. Specifically, in this study, large-scale real data collected from the building energy management system were used in the online energy consumption forecasting, which is specially designed for optimized control, real-time fault detection, diagnosis and abnormality alarms. Firstly, the characteristics of building energy profiles and data reliability were addressed. Mathematical algorithms were introduced and their previous applications in building energy usage prediction were summarized, including the evaluation criteria that are effective for energy use predictions in buildings. The reliability and efficiency of the proposed algorithms were then demonstrated through the comparison between the monitored actual data and the predicted results. It is found that Gaussian Process Regression (GPR) can give acceptable predictions on the energy consumption of office buildings with an equilibrium of data prediction accuracy with the average deviations of below 15% and low computation time. Additionally, the statistical evaluation criteria proposed by ASHRAE can also be satisfied. For hotels and shopping malls where complex functions were applied in these buildings, their accuracy are not better or even the same as those of simplified models, due to the significant effects of the factors involving occupant's activities and schedules as well as data reliability on building energy usage. Our result revealed that GPR is a reliable method and can still generate highly accurate predictions when a large data set with a small time interval and complex energy use patterns obtained from real building measurements rather than simulated data are involved.