基于SHAP重要性排序和时空双流的多风场超短期功率预测  

Ultra-short-term power forecasting for multiple wind fields based on SHAP importance ranking and spatio-temporal dual flow

作　　者：付波[1] 李昊 权轶[1] 李超顺[2] 赵熙临[1] 杨远程 FU Bo;LI Hao;QUAN Yi;LI Chaoshun;ZHAO Xilin;YANG Yuancheng(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;College of Civil and Hydraulic Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;State Grid Shishou City Power Supply Company,Shishou 434400,China)

机构地区：[1]湖北工业大学电气与电子工程学院,武汉430068 [2]华中科技大学土木与水利工程学院,武汉430074 [3]国网石首市供电公司,湖北荆州434400

出　　处：《重庆理工大学学报（自然科学）》2024年第5期249-258,共10页Journal of Chongqing University of Technology：Natural Science

基　　金：湖北省重点研发计划项目(2021BAA193)。

摘　　要：针对多风场风功率预测中时空特征提取不充分的问题,提出一种基于空间、时间双流特征提取的功率预测方法。采用沙普利加性解释(SHAP)方法分析原始高维数值天气预报(NWP)中各变量的重要性,选择贡献度高的变量子集作为预测模型输入,降低模型复杂度。构建基于自适应动态邻接矩阵的改进图注意力网络(IGAT)提取多风场的动态空间特征;同时将多头注意力机制(MHA)与时间卷积网络(TCN)结合,加强关键时序特征的学习。使用前馈神经网络输出多风场功率预测结果。以西北某十风场的数据进行案例研究,结果表明所提模型的预测效果优于其他模型。To address the insufficient spatio-temporal feature extraction in multi-wind field wind power prediction,this paper proposes a power prediction method based on spatial and temporal dual-stream feature extraction.First,the Shapley Additive Explanations(SHAP)method is employed to analyze the importance of each variable in the original high-dimensional Numerical Weather Prediction(NWP),and a subset of variables with high contribution is selected as the input to the prediction model,reducing the complexity of the model.Second,an Improved Graph Attention Network(IGAT)based on an adaptive dynamic adjacency matrix is built to extract the dynamic spatial features of multiple wind fields.Meanwhile,Multi-Head Attention Mechanism(MHA)is integrated with Temporal Convolutional Network(TCN)to enhance the learning of key temporal features.Then,a feed-forward neural network is used to output the power prediction results of multiple wind fields.Finally,a case study is conducted with data from ten wind fields in Northwest China,and our results show the proposed model performs better in prediction than other models.

关 键 词：多风场功率预测 变量选择 图注意力网络 多头注意力机制 时间卷积网络 

分 类 号：TM614[电气工程—电力系统及自动化]