Hybrid thermal load prediction model for the regional built environment based on random forest feature screening

With the Chinese government’s strong call for energy conservation and emission reduction, the regulatory requirements for optimizing the energy-carbon relationship in buildings are increasing. Accurate thermal load prediction in the built environment is essential for optimizing heating system design and operational strategies. To correctly forecast the heating energy load of regional buildings, this research proposes a hybrid prediction model that uses random forest (RF) feature screening as input variables and the bi-directional long short-term memory (Bi-LSTM) network for extracting the relationships between the features and the loads. The comparison results demonstrate that the hybrid model enhanced by RF has dramatically improved the calculation speed and prediction accuracy compared to the original model. The proposed RF-Bi-LSTM model showed significant improvements in prediction accuracy and computational efficiency compared with traditional models (exemplified by LSTM). Specifically, the root mean square error (RMSE) decreased by 12.7% from 1.057 for the conventional model to 0.923, the coefficient of variation of the root mean square error (CV-RMSE) was reduced by 12.7% (from 0.055 to 0.048), and the mean absolute percentage error (MAPE) decreased by 9.5% (from 4.54% to 4.11%). Furthermore, the computational time was significantly shortened by 37.8%, decreasing from 111 s to 69 s. These results highlight the superior performance of the hybrid model in balancing accuracy and efficiency for building load forecasting tasks. The model can contribute to energy saving and carbon reduction by reducing building energy consumption through precision load forecasting.