摘要
高效准确的短期电力负荷预测能帮助电力部门合理制定生产调度计划,减少资源浪费。深度学习中以循环神经网络(recurrent neural network,RNN)为主体构建的预测模型是短期负荷预测方法中的典型代表,但存在难以有效提取历史序列中潜在高维特征且当时序过长时重要信息易丢失的问题。提出了一种基于Attention机制的卷积神经网络(convolutional neural network,CNN)-GRU(gated recurrent unit)短期电力负荷预测方法,该方法将历史负荷数据作为输入,搭建由一维卷积层和池化层等组成的CNN架构,提取反映负荷复杂动态变化的高维特征;将所提特征向量构造为时间序列形式作为GRU网络的输入,建模学习特征内部动态变化规律,并引入Attention机制通过映射加权和学习参数矩阵赋予GRU隐含状态不同的权重,减少历史信息的丢失并加强重要信息的影响,最后完成短期负荷预测。以美国某公共事业部门提供的公开数据集和中国西北某地区的负荷数据作为实际算例,该方法预测精度分别达到了97.15%和97.44%,并与多层感知机(multi-layer perceptron,MLP)、径向基神经网络(radial basis function neural network,RBF)、支持向量回归(support vector regression,SVR)、GRU、CNN、自编码器(autoencoder,AE)-GRU和未引入Attention机制的CNN-GRU进行对比,实验结果表明所提方法具有更高的预测精度。
Efficient and accurate short-term load forecasting can help power utilities to rationally formulate production scheduling plans and reduce resource waste.RNN(recurrent neural network)-based predictive model in deep learning is a typical method of short-term load forecasting,but it is difficult to effectively extract potential high-dimensional features in historical sequences and important information is easily lost when the time series is too long,resulting in decrease of prediction accuracy.This paper proposes a short-term load forecasting method of CNN(convolutional neural network)-GRU(gated recurrent unit)based on attention mechanism.This model takes historical load data as input and builds a CNN architecture composed of one-dimensional convolution layers and pooling layers to extract complex dynamic changes of load.A high-dimensional feature is constructed by constructing the proposed feature vector into time series as GRU input.It models the internal dynamic change of the feature,and the attention mechanism is introduced to the GRU implicit state through mapping weight and learning parameter matrix,reducing the loss of historical information and enhancing the impact of important information,thus completing short-term load forecasting.Taking the public data set provided by a public utility department in the United States and the load data of a certain region in China as practical examples,the prediction accuracy of the method reaches 97.15%and 97.44%respectively.Comparing with multi-layer perceptron(MLP),radial basis function(RBF)neural network,support vector regression(SVR),GRU,CNN,AE(autoencoder)-GRU and CNN-GRU,the proposed method has higher prediction accuracy showed in experimental results.
作者
赵兵
王增平
纪维佳
高欣
李晓兵
ZHAO Bing;WANG Zengping;JI Weijia;GAO Xin;LI Xiaobing(School of Electrical and Electronic Engineering,North China Electric Power University,Changping District,Beijing 102206,China;China Electric Power Research Institute,Haidian District,Beijing 100192,China;School of Automation,Beijing University of Posts and Telecommunications,Haidian District,Beijing 100876,China)
出处
《电网技术》
EI
CSCD
北大核心
2019年第12期4370-4376,共7页
Power System Technology
基金
国家重点研发计划项目((2016YFF0201201)~~
关键词
短期负荷预测
卷积神经网络
门控循环单元
注意力机制
short-term load forecasting
convolutional neural network
gated recurrent unit
attention mechanism