Multi-Task Learning Based Joint Source-Load Forecasting Model for Integrated Energy System

The efficient operation of integrated energy systems urgently requires solving the prediction challenges caused by the strong randomness of renewable energy and the coupling characteristics of multi-energy loads. This paper proposes a joint forecasting model that combines MTL-MHA-CNN-LSTMS, which constructs a shared representation space for source-load correlation features through multi-task learning, captures key temporal dynamics using a multi-head attention mechanism, and innovatively adopts a two-stage LSTM to model long-term trends and residual correction separately. Experiments show that the model achieves R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> scores of 90.4%, 97.7%, 96.9%, and 92.2% in electricity load, heat load, solar power, and wind power forecasting, respectively, outperforming baseline models by 4.9% to 15.8%. The study not only provides a high-accuracy, low-computation solution for source-load forecasting but also opens a new pathway for multi-energy collaborative forecasting research through its proposed dynamic feature interaction framework.