Performance predictions of ground source heat pump system based on random forest and back propagation neural network models

With rapid development of artificial intelligence, data-driven prediction models play an important role in energy prediction, fault detection, and diagnosis. This paper proposes an ensemble approach using random forest (RF) for hourly performance predictions of GSHP system. Two years of in situ data were collected in an educational building situated in severe cold area in China. Prediction models were established for performance indicators, and results indicate that the average error for COPs, COPu, EERs and EERu were all controlled within 5%. The model established by small amount of data can accurately predict long-term performance, thereby reducing time and difficulty of data collection. RF models, trained with different parameter settings were compared, results indicate that model accuracy was not very sensitive to variables numbers. The impact of input variables on prediction performance was analyzed, and importance ranking changed with period and performance indicators. By comparing the variable importance list, it was possible to establish which parameters were abnormal and lists of different periods can reflect whether the energy structure of building has changed. The overall superiority of RF was verified by comparing with back propagation neural network (BPNN) from robustness, interpretability, and efficiency. First, since GSHP system involving multiple indicators, the robustness, measured by average accuracy, was used to evaluate the accuracy level. According to CV-RMSE, robustness of RF is approximately 3.3% higher than that of BPNN. Second, RF is highly interpretive but BPNN is typical black box model. Finally, modeling complexity and training time of BPNN were much greater than RF.