基于序列成分重组与时序自注意力机制改进TCN-BiLSTM的短期电力负荷预测  

Short-term Power Load Forecasting Based on Sequence Component Recombination and Temporal Self-attention Mechanism Improved TCN-BiLSTM

作　　者：易雅雯 娄素华[1] YI Yawen;LOU Suhua(School of Electrical and Electronic Engineering,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]华中科技大学电气与电子工程学院,武汉430074

出　　处：《电力系统及其自动化学报》2025年第4期78-87,共10页Proceedings of the CSU-EPSA

基　　金：国家电网公司科学技术项目(5108-202218280A-2-429-XG)。

摘　　要：针对区域级电力负荷预测精度较低的问题,提出一种基于序列成分重组与时序自注意力机制改进时间卷积网络-双向长短期记忆网络(TCN-BiLSTM)的短期负荷预测方法。首先,通过中心频率法确定最佳初始分解数目,进而采用变分模态分解算法将原始负荷序列分解为多个不同频率的成分序列;其次,基于各成分序列的样本熵对多个成分序列进行K均值聚类,以获得最佳聚类数量的重组负荷序列分量;接着,将各重组分量输入所提出的负荷预测模型,获得各重组分量预测结果;最终,线性叠加各重组成分序列预测结果以获得最终负荷预测结果。算例分析表明,该方法与其他相关对比模型相比,预测均方根误差降低46.37%、模型拟合效果平均提升3.24%,表明该方法负荷预测精度高、模型拟合效果好,适用于区域级电力负荷预测。To address the issue of low accuracy in regional power load forecasting,a short-term load forecasting method based on sequence component recombination and temporal self-attention mechanism improved temporal convolutional network-bidirectional long short-term memory network(TCN-BiLSTM)is proposed in this paper.First,the optimal ini-tial decomposition number is determined through the central frequency method,and then the original load sequence is decomposed using the variational mode decomposition algorithm to obtain multiple sequences of different frequency components.Second,the multiple component sequences are clustered by K-means based on the sample entropy of each component sequence,thus obtaining the recombined load sequence components with the optimal number of clusters.Third,each recombined component is input into the load forecasting model proposed in this paper to obtain the predic-tion results of each recombined component.Finally,the prediction results of recombined components are linearly su-perimposed to obtain the final load prediction results.The analysis of a case study shows that compared with the average value of other related contrast models,the prediction root mean square error of the proposed method is reduced by 46.37%,and the model fitting effect is improved by 3.24%on average.This result indicates that the proposed method possesses a high prediction accuracy and a better model fitting effect,which is suitable for regional power load forecasting.

关 键 词：负荷预测 变分模态分解 样本熵 K均值聚类 时序自注意力机制 时间卷积网络 双向长短期记忆网络 

分 类 号：TM715[电气工程—电力系统及自动化]