A multi-task learning-based heat load prediction method of multi-use buildings: Integrating weather-heat load conditional learning and heat load time-delay prior constraint

Accurate prediction of multi-building heat loads in district heating systems is essential for optimizing heating scheduling strategies. Existing methods suffer from limited capability in representing the multi-source driving features of building heat loads, as well as their time-delay dynamics. To address these issues, a novel heat load prediction method was proposed. Temporal Convolutional Network (TCN) and Transformer models were adopted to extract the local and global features of the weather time-series. A weather-heat load conditional learning mechanism was introduced to adaptively identify the heat load temporal information that was strongly correlated with the extracted weather features. Finally, a multi-task learning-based prediction model was developed, which integrated a shared driving feature module and building-specific prediction branches. An autoregression-based heat load time-delay prior constraint mechanism was employed to improve the model training process. The proposed method was validated using the operational data from a district heating system in Hebei Province, China. The results showed that the proposed model achieved high prediction accuracy for multi-building heat loads, with R2 values ranging from 0.97 to 0.99. Compared with conventional and representative deep learning and machine learning methods, the proposed method reduced MAE and MAPE by an average of 65.73% and 74.89%, respectively. This study could provide reliable feedforward information for load allocation and operational optimization of district heating systems.