Joint source-load forecasting for large shopping centers based on hierarchical multi-task learning

To address the low prediction accuracy caused by the complex coupling between energy sources and loads in large shopping centers, a multi-task joint forecasting model is proposed. First, correlation analysis is performed on three uncertain variables, namely, heating load, electricity load, and photovoltaic (PV) generation, to reveal the relationships among forecasting tasks and their influencing factors. Based on the correlation characteristics, a multi-channel convolutional neural network (MCNN) is employed to extract temporal features from heating load, electricity load, PV generation, and environmental variables. Subsequently, a multi-head attention (MHA) mechanism is introduced to dynamically assign feature weights and suppress noise interference from weakly correlated information. To further exploit the coupling relationships among forecasting tasks, a hierarchical multi-task learning (HMTL) architecture based on long short-term memory networks is developed. Hard parameter sharing is adopted for strongly correlated tasks, while soft parameter sharing is employed for weakly correlated but coupled tasks. By integrating MCNN, MHA, and HMTL, the proposed framework achieves high-precision joint forecasting of energy sources and loads. Experimental results demonstrate that the proposed method reduces training time by 29.52% compared with single-task forecasting. The mean absolute percentage errors of heating load, electricity load, and PV generation are 2.66%, 1.98%, and 3.42%, respectively.