基于Inception和注意力机制的双分支日前电价预测  

Double-branch Day-ahead Electricity Price Forecasting Based on Inception and Attention Mechanism

作　　者：毕贵红[1] 孔凡文 黄泽 陈冬静 骆钊[1] 杨毅[1] BI Guihong;KONG Fanwen;HUANG Ze;CHEN Dongjing;LUO Zhao;YANG Yi(School of Electric Power Engineering,Kunming University of Science and Technology,Kunming 650500,China;Power China Guiyang Engineering Corporation Limited,Guiyang 550081,China)

机构地区：[1]昆明理工大学电力工程学院,云南省昆明市650500 [2]中国电建集团贵阳勘测设计研究院有限公司,贵州省贵阳市550081

出　　处：《电力系统自动化》2025年第5期128-144,共17页Automation of Electric Power Systems

基　　金：国家重点研发计划资助项目(2022YFB2703500)。

摘　　要：在电力市场化的背景下,开放电力市场受需求端负荷、新能源出力和市场间耦合关系等复杂因素影响,其电价波动变得愈发强烈且难以预测。为合理选择影响电价波动的综合因素,降低原始电价序列非稳定性、强波动性对电价预测所产生的负面影响,提出了一种基于双模式分解与Inception、注意力机制组合的双分支日前电价预测方法。首先,将最大信息系数筛选和与日前电价相关性较高的影响因素进行组合,作为模型相关变量特征矩阵输入;然后,通过变分模态分解和群分解将原始电价序列分解为多个更能反映电价波动规律的子序列,将不同分解方法得到的子序列按高频到低频进行排序,再组合构造多尺度电价分量矩阵作为模型电价分支输入,以提高模态分量的规律性和信息的丰富性;最后,将改进的Inception模块与并行多维注意力(PMDA)、自注意力机制分别进行组合,搭建双分支输入的日前电价预测模型,以提取不同分支输入数据的重要特征并进行融合,输出次日电价预测结果。以北欧电力市场历史数据为例进行验证,并与传统注意力机制进行对比,实验结果表明所提PMDA机制能够更有效地提取电价序列重要特征,以提高日前电价预测精度。In the background of electricity marketization,the open electricity market is affected by demand-side load,renewable energy output and coupling relationship between markets,and its electricity price fluctuation becomes more and more intense and difficult to forecast.In order to rationally select the comprehensive factors affecting the fluctuation of electricity price and reduce the negative impact of the instability and strong fluctuation of the original electricity price series on the electricity price forecasting,this paper proposes a double-branch day-ahead electricity price forecasting method based on the combination of dual-mode decomposition,Inception and attention mechanism.First,the maximum information coefficient screening and the influence factors with high correlation with the day-ahead electricity price are combined as the input of the model correlation variable feature matrix.Then,through variational mode decomposition and swarm decomposition,the original electricity price series are decomposed into multiple subseries that can better reflect the electricity price fluctuation law.The subseries obtained by different decomposition methods are sorted from high frequency to low frequency,and then the multi-scale electricity price component matrix is constructed as the input of the electricity price branch of the model to improve the regularity and information richness of the modal components.Finally,the improved Inception module is combined with parallel multidimensional attention(PMDA)and self-attention mechanism respectively to build a double-branch day-ahead electricity price forecasting model to extract and fuse important features of input data from different branches and output the electricity price forcasting results for the next day.By using the historical data of Nordic electricity market as a case for verification,and comparing with the traditional attention mechanism,the experimental results show that the proposed PMDA mechanism can extract important features of electricity price series more eff

关 键 词：电价预测 注意力 最大信息系数 Inception网格 电力市场 

分 类 号：TM73[电气工程—电力系统及自动化] F426.61[经济管理—产业经济]