摘要
准确的短期多元负荷预测是确保综合能源系统可靠经济运行的必要前提。针对现有模型预测精确度不高的问题,本文提出一种基于改进最大信息系数相关性分析和MMoE-TCN多任务学习的负荷预测方法。首先,采用改进的最大信息系数相关性分析方法筛选目标预测负荷的特征序列集。然后,建立基于参数软共享机制的MMoE多任务学习模型,通过专家子网和门控单元合理分配子任务的共享特征信息,挖掘多元负荷间的耦合特性,进而使用时间卷积神经网络构建子任务模型,用于负荷预测。最后,使用IES公开数据集进行算例分析,其误差均低于MTL-TCN、MTL-LSTM和LSTM模型,验证了本文所提方法有较高的预测准确度。
The accurate short-term multiple load forecasting is a necessary prerequisite to ensure reliable and economical operation of the integrated energy system.To solve the problem of low accuracy of current model,this paper proposes a load forecasting method based on improved maximum information coefficient(IMIC)correlation analysis and multi-gate mixture-of-experts fused into temporal convolutional neural network(MMoE-TCN)multi-task learning.Firstly,the improved maximum information coefficient correlation analysis method is used to screen the feature sequence set of the target forecast load.Then,the MMoE multi-task learning model based on parameter soft sharing mechanism is constructed,and the coupling characteristics among multiple loads are dug out through expert subnets and gating units,and the sub-task model is constructed using temporal convolution network.Finally,the IES open data are used for example analysis,and the error is lower than those of MTL-TCN,MTL-LSTM and LSTM models,which verifies the effectiveness and feasibility of the proposed method.
作者
王定美
张睿骁
赵龙
WANG Dingmei;ZHANG Ruixiao;ZHAO Long(Electric Power Research Institute of State Grid Gansu Electric Power Company,Lanzhou 730070,Gansu,China)
出处
《电气传动自动化》
2023年第1期39-45,38,共8页
Electric Drive Automation
基金
国网甘肃省电力公司科技项目(522722200032)。
