摘要
为提高配电网超短期负荷预测精度,从特征构建与模型优化两个角度出发,提出一种基于Prophet和双重多头自注意力-时间卷积网络的超短期负荷预测框架。首先,通过Prophet提取负荷序列中隐含的多时间尺度时序特征。然后,基于最大信息系数选择预测模型的输入特征,并采用最佳滑动窗口构建输入矩阵。最后,在时间卷积网络的基础上,引入特征和时序双重多头自注意力,用于挖掘负荷特征矩阵中不同输入特征、不同时间步之间的内部相关性,并为特征、时间步自适应赋权以突出重要信息的影响。基于湖南省某配电网台区负荷数据开展算例分析,消融实验结果表明所构建预测模型的有效性;与多种传统机器学习和深度学习预测模型的对比测试结果表明,所提方法具有更高的负荷预测精度。
To improve the accuracy of ultra-short-term load forecasting in the distribution network,this paper proposes an ultrashort-term load forecasting framework based on Prophet and dual muti-head self-attention-temporal convolutional network from two perspectives:feature construction and model optimization.First,the multi-timescale temporal features implied in the load sequence are extracted by Prophet.Then,the input features of the forecasting model are selected based on maximum information coefficient,and the input matrix is constructed by means of the optimal sliding window.Finally,on the basis of temporal convolutional network,the dual multi-head self-attention of features and time-steps are introduced to mine the internal correlation between different input features and different time-steps in the feature matrix,respectively,and the features and time-steps are adaptively weighted to highlight the influence of important information.Case analysis is carried out based on the load data of transformer areas in Hunan Province,China.The results of the ablation experiment demonstrate the effectiveness of the proposed forecasting model.The comparison test results with various traditional machine learning and deep learning forecasting models verify that the proposed method has higher load forecasting accuracy.
作者
周思思
李勇
郭钇秀
乔学博
梅玉杰
邓威
ZHOU Sisi;LI Yong;GUO Yixiu;QIAO Xuebo;MEI Yujie;DENG Wei(School of Electrical and Information Engineering,Hunan University,Changsha 410082,China;Electric Power Research Institution of China Southern Power Grid,Guangzhou 510663,China;Electric Power Research Institute of State Grid Hunan Electric Power Co.,Ltd.,Changsha 410007,China)
出处
《电力系统自动化》
EI
CSCD
北大核心
2023年第18期193-205,共13页
Automation of Electric Power Systems
基金
国家重点研发计划政府间国际科技创新合作重点专项资助项目(2022YFE0129300)
国家自然科学基金重点支持项目(U22B20104)。
关键词
配电网
负荷预测
特征提取
多头自注意力
时间卷积网络
配电变压器台区
distribution network
load forecasting
feature extraction
multi-head self-attention
temporal convolutional network
distribution transformer area