Advancing regional heat load forecasting through sophisticated data-driven methodologies integrated with robust adversarial training strategies

Facing the severe challenges of global climate change and the continuous growth of building energy consumption, improving the accuracy and robustness of the prediction of district heating systems has become critical to achieving building energy efficiency. However, existing thermal load prediction models still exhibit significant limitations under complex climatic conditions and anomalous data scenarios. To address these challenges, this paper proposes a novel deep learning model, RobustTransBlock, which innovatively integrates a periodic feature extraction module, a self-attention mechanism, and adversarial training to resolve multi-timescale dependency modeling and data anomaly robustness. Based on real thermal load data from residential buildings in Denmark, the model's performance is validated through ablation studies, adversarial attack experiments, and visual analysis. Experimental results demonstrate that the proposed model achieves 15.7%, 16.7%, and 13.1% reductions in Mean Absolute Error (MAE) across different time steps compared to state-of-the-art methods, while exhibiting enhanced stability under adversarial attack scenarios. The research results provide technical support for dynamic optimization of district heating systems and reliable prediction in anomalous data environments, demonstrating significant engineering applicability. Notably, this work pioneers the introduction of adversarial training into building thermal load prediction and resolves the multi-timescale dependency modeling challenge through synergistic design of attention mechanisms and periodic feature decoders, offering novel insights for advancing building energy prediction technologies. The code is available at https://github.com/zhuhaoran-ai/RobustTransBlock.