A deep-learning framework for predicting building heat load with hyperparameter optimisation and physics-constrained post-processing

Driven by global energy transition and carbon neutrality goals, accurate building heat load prediction is of great significance for intelligent heating control, demand-side energy management, and fault detection. This study focuses on a single-apartment located in Bologna and predicts the heat load for the next hour based on the previous 24 h of data. A dataset of 18 typical operating scenarios was constructed, encompassing variations in outdoor meteorological conditions, occupancy, thermostat settings, and equipment heat gains. Subsequently, the Optuna framework was employed to systematically optimize the hyperparameters of five models—LSTM, TCN, MLP, XGBoost, and LR—and their prediction performance was comparatively analyzed. Results show that that the deep learning models LSTM and TCN perform best, with LSTM slightly outperforming (LSTM: MAE = 0.082 kW, RMSE = 0.118 kW, R2 = 0.863; TCN: MAE = 0.0864 kW, RMSE = 0.1196 kW, R2 = 0.8606). Whereas the LR model exhibits the poorest performance, achieving only 0.1202 kW, 0.1504 kW, and 0.7793 on the respective metrics. After applying a post-processing (PC) correction mechanism, the MAE and RMSE of LSTM decreased to 0.058 kW and 0.100 kW, corresponding to reductions of 28.2% and 15.3%, respectively. This framework effectively integrates data-driven methods with physical constraints while maintaining low computational cost, providing a scalable and practical solution for intelligent building energy management and low-carbon operation.