基于Attention-LSTM与多模型集成的短期负荷预测方法  

Short-term load forecasting method based on Attention-LSTM and multi-model integration

作　　者：朱继忠 苗雨旺 董朝阳 董瀚江 陈梓瑜 李盛林 ZHU Jizhong;MIAO Yuwang;DONG Zhaoyang;DONG Hanjiang;CHEN Ziyu;LI Shenglin(School of Electric Power,South China University of Technology,Guangzhou 510641,China;School of Electrical Engineering,Nanyang Technological University,Singapore 639798,Singapore)

机构地区：[1]华南理工大学电力学院,广东广州510641 [2]南洋理工大学电气工程学院,新加坡639798

出　　处：《电力工程技术》2023年第5期138-147,共10页Electric Power Engineering Technology

基　　金：国家自然科学基金资助项目(52177087);国家外国专家资助项目(G2022163018L)。

摘　　要：目前深度学习技术发展快速,针对其在短期负荷预测任务中处理离散数据效果较差以及泛化性不佳的问题,提出一种基于注意力机制的长短期记忆网络(long short-term memory network with attention mechanism,Attention-LSTM)与Stacking多模型集成的负荷预测方法,可以兼顾二者优势。首先,利用均值编码的方式处理离散特征,接着应用Attention-LSTM对负荷数据进行特征提取,再将处理后的数据一同输入到基于Stacking的多模型集成预测模型中,通过3种基学习器对输入特征进行分析处理,最终通过元学习器完成预测。算例使用2个数据集中的实际负荷数据进行分析,对2个数据集中的负荷数据分别进行预测,并与门控学习单元、轻量级梯度提升机、支持向量机方法进行对比。仿真结果表明,所提方法在2个数据集的预测精度均能够超过98%,比其他3种方法的预测精度更高。Lately,deep learning techniques have developed rapidly.To address the problem of the poor processing capability of discrete data and the problem of poor generalization in short-term load forecasting tasks,a load forecasting method based on a long short-term memory network with attention mechanism(Attention-LSTM)and Stacking multi-model integration is proposed,which can take into account their respective advantages.Firstly,the discrete features are processed by mean coding,and then Attention-LSTM is used to extract the features of the load data.Next,the processed data are input into the multi-model integrated prediction model based on Stacking.The input features are analyzed and processed through three basic learners.Finally,the prediction is completed through the meta-learner.Actual load data from two datasets are used for analysis in the case study,and the load data in the two datasets are predicted separately,and compared with gated recurrent units,light gradient boosting machine and support vector machine.The simulation results show that the prediction accuracy based on the proposed method in the two datasets can exceed 98% in terms of accuracy,which is higher than the other three methods.

关 键 词：电力系统 短期负荷预测 长短期记忆网络 Stacking集成学习 注意力机制 均值编码 

分 类 号：TM743[电气工程—电力系统及自动化]