Source-load joint forecasting of office building electro-thermal integrated energy systems based on multi-task learning and dynamic optimal integration

Accurate forecasting of the source–load supply–demand relationship in integrated energy systems (IES) is essential for optimizing energy dispatch and improving operational efficiency. Independent source–load forecasting disrupts the supply–demand integrity, while static integration strategies fail to adaptively adjust model weights, limiting prediction accuracy. To address these issues, this study proposes a joint forecasting method for electro-thermal IES. The Unified Information Coefficient (UIC) is first used to identify highly relevant input features and reduce redundant information. Then, an MFCNN-BiLSTM-CBAM hybrid framework is developed to extract multi-scale temporal features and enhance focus on critical patterns. On this basis, a multi-task learning (MTL) approach is introduced to enable low-level feature sharing and adaptive loss weighting for jointly forecasting electrical load, thermal load, and renewable energy output. Furthermore, a dynamic optimal ensemble strategy employs a Multi-Layer Perceptron (MLP) to adjust base learner weights in real time. Experiments on a solar-assisted combined heat and power system in an office building in Xi’an, China show that the proposed method outperforms independent forecasting and static integration strategies, achieving mean absolute percentage errors (MAPE) of 1.07%, 1.05%, and 1.21% in winter, summer, and transition seasons, respectively, demonstrating strong accuracy and robustness.