为解决热负荷数据因受天气变化、季节波动、设备故障等多因素影响而呈现的非平稳性和噪声干扰问题，以及传统预测模型在该场景下精度不足的挑战，提出一种基于 CEEMDAN 分解、PatchTST 和 iTransformer 的混合模型框架，并结合 Bagging 融合策略。通过 CEEMDAN 将热负荷数据分解为多个本征模态函数（IMF）以抑制非平稳性和噪声；利用 PatchTST 的分块策略与自注意力机制捕捉数据多尺度依赖，处理长期周期性和短期局部波动；借助 iTransformer 调整 IMF 分量的注意力权重，减少预测误差累积并增强多分量信号融合能力；通过 Bagging 策略提升模型鲁棒性。实验结果表明，该模型在热负荷预测任务中 MSE 和 MAE 分别低至 0.0076 和 0.1956，显著优于传统方法和现有 SOTA 模型。该模型构建了多模态分解与双路径预测的协同框架，整合模态分解、多尺度依赖建模、注意力权重调整及集成学习，有效应对非平稳热负荷序列的预测难题，大幅提升了预测精度与稳定性。

To address the non-stationarity and noise interference in thermal load data caused by weather changes, seasonal fluctuations, equipment failures, and other factors, as well as the challenge of insufficient accuracy of traditional prediction models in this scenario, a hybrid model framework based on CEEMDAN decomposition, PatchTST, and iTransformer is proposed, combined with the Bagging fusion strategy. CEEMDAN is used to decompose the thermal load data into multiple intrinsic mode functions (IMFs) to suppress non-stationarity and noise; PatchTST's block strategy and self-attention mechanism are utilized to capture multi-scale dependencies in the data, handling long-term periodicity and short-term local fluctuations; iTransformer is employed to adjust the attention weights of the IMF components, reducing the accumulation of prediction errors and enhancing the fusion ability of multi-component signals; and the Bagging strategy is adopted to improve model robustness. Experimental results show that the model achieves MSE and MAE as low as 0.0076 and 0.1956 respectively in the thermal load prediction task, significantly outperforming traditional methods and existing SOTA models. This model constructs a collaborative framework of multi-modal decomposition and dual-path prediction, integrating modal decomposition, multi-scale dependency modeling, attention weight adjustment, and ensemble learning, effectively addressing the prediction challenges of non-stationary thermal load sequences and significantly improving prediction accuracy and stability.

国能宁夏供热有限公司项目（NXDL-2024-44）