Multi-criteria comprehensive study on predictive algorithm of hourly heating energy consumption for residential buildings

The increasing of building energy necessitates reliable energy consumption prediction. Certain research work is necessary to thoroughly illustrate and compare advantages and disadvantages of various models. Therefore, this study investigated comprehensive trade-off between performances of commonly used forecasting models based on multiple performance metrics. Considering the requirements of actual building energy system, the objectives included accuracy, interpretability, robustness, and efficiency. With actual heating energy, prediction models were established by applying extreme gradient boosting (XGBoost), random forest (RF), artificial neural network (ANN), gradient boosting decision tree (GBDT), and support vector regression (SVR). A comparison revealed the following: 1) RF exhibits optimal average accuracy (under different training datasets), whereas ANN exhibits contrary properties. 2) The robustness of RF is the highest from adaptation to different training datasets with minimum standard deviation of error; XGBoost and ANN exhibit contrary properties. 3) RF, GBDT, and XGBoost are rendered effectively interpretable. 4) At equivalent accuracy level, ANN and SVR require auxiliary algorithms, whereas other models can achieve reasonable accuracy no tuning required. BPNN's calculation time is of an order magnitude higher than those of other models. Overall, XGBoost exhibits the optimal efficiency. This study can provide guidance for effectively selecting prediction models for energy management.