Improved Data-Driven Building Daily Energy Consumption Prediction Models Based on Balance Point Temperature

Abstract

The data-driven models have been widely used in building energy analysis due to their outstanding performance. The input variables of the data-driven models are crucial for their predictive performance. Therefore, it is meaningful to explore the input variables that can improve the predictive performance, especially in the context of the global energy crisis. In this study, an algorithm for calculating the balance point temperature was proposed for an apartment community in Xiamen, China. It was found that the balance point temperature label (BPT label) can significantly improve the daily energy consumption prediction accuracy of five data-driven models (BPNN, SVR, RF, LASSO, and KNN). Feature importance analysis showed that the importance of the BPT label accounts for 25%. Among all input variables, the daily minimum temperature is the decisive factor that affects energy consumption, while the daily maximum temperature has little impact. In addition, this study also provides recommendations for selecting these model tools under different data conditions: when the input variable data is insufficient, KNN has the best predictive performance, while BPNN is the best model when the input data is sufficient.