Han Haotong, Peng Jishen, Ma Jun, Liu Shang Lin, Liu Hao
Faculty of Electrical and Control Engineering, Liaoning Technical University, Huludao, 125000, China.
Sci Rep. 2025 Apr 25;15(1):14530. doi: 10.1038/s41598-025-98224-4.
In the context of increasing electricity consumption and the growing complexity of energy usage patterns, accurate power load forecasting faces more significant challenges. Given the widespread demand for precise load curve prediction, this paper proposes a CEEMDAN-SE-VMD + SelfAttention-TCN Fusion model. Load curves contain rich information, and comprehensive information mining contributes to prediction accuracy. Specifically, the methodology first employs Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose the load curve, where the resulting Intrinsic Mode Functions (IMFs) are adaptively categorized into high and low-frequency components using Sample Entropy (SE). The high-frequency components undergo further extraction through Variational Mode Decomposition (VMD), while the low-frequency components are directly utilized as model inputs without additional processing. Through this process, the original load curve is reconstructively expressed as sequences of high and low-frequency components. These reconstructed sequences are then fed into Self-Attention Temporal Convolutional Network (TCN) for prediction, and the individual predictions are integrated to generate the final forecast. Using historical German power load data from ENTSO for case analysis, the model achieves an Root Mean Square Error (RMSE) of 24.1293 (MW), Mean Absolute Error (MAE) of 17.268 (MW), and R-square of 0.9838, surpassing the prediction accuracy of other comparative models. Experimental results reveal that the signal decomposition and reconstruction process enables more effective expression of deep characteristic information inherent in load data, thereby enhancing model learning performance. Furthermore, the self-attention mechanism strengthens TCN's ability to capture data dependencies. The proposed model demonstrates both a high tolerance for raw load data and superior prediction accuracy, with experimental validation confirming its excellent performance in real-world short-term power load forecasting applications.
在电力消耗不断增加以及能源使用模式日益复杂的背景下，准确的电力负荷预测面临着更为严峻的挑战。鉴于对精确负荷曲线预测的广泛需求，本文提出了一种CEEMDAN-SE-VMD + 自注意力-时间卷积网络（SelfAttention-TCN）融合模型。负荷曲线包含丰富信息，全面的信息挖掘有助于提高预测精度。具体而言，该方法首先采用带自适应噪声的完全集合经验模态分解（CEEMDAN）对负荷曲线进行分解，然后使用样本熵（SE）将得到的本征模态函数（IMF）自适应地分类为高频和低频分量。高频分量通过变分模态分解（VMD）进一步提取，而低频分量则直接作为模型输入，无需额外处理。通过这一过程，原始负荷曲线被重构为高频和低频分量序列。然后将这些重构序列输入自注意力时间卷积网络（TCN）进行预测，并将各个预测结果进行整合以生成最终预测。使用来自ENTSO的德国历史电力负荷数据进行案例分析，该模型的均方根误差（RMSE）为24.1293（MW），平均绝对误差（MAE）为17.268（MW），决定系数（R平方）为0.9838，超过了其他对比模型的预测精度。实验结果表明，信号分解和重构过程能够更有效地表达负荷数据中固有的深度特征信息，从而提高模型的学习性能。此外，自注意力机制增强了TCN捕捉数据依赖性的能力。所提出的模型对原始负荷数据具有较高的容忍度和卓越的预测精度，实验验证证实了其在实际短期电力负荷预测应用中的优异性能。
