基于GRU神经网络的WGAN短期负荷预测模型  

Short-term Load Forecasting Model of WGAN Based on GRU Neural Network

作　　者：高翱 王帅 韩兴臣 张智晟 GAO Ao;WANG Shuai;HAN Xingchen;ZHANG Zhisheng(College of Electrical Engineering,Qingdao University,Qingdao 266071;Qingdao Power Supply Company,State Grid Shandong Electric Power Company,Qingdao 266001)

机构地区：[1]青岛大学电气工程学院,青岛266071 [2]国网山东省电力公司青岛供电公司,青岛266001

出　　处：《电气工程学报》2022年第2期168-175,共8页Journal of Electrical Engineering

基　　金：国家自然科学基金资助项目(52077108)。

摘　　要：为了提高短期负荷预测模型的精度,提出了一种基于门控循环单元(Gated recurrent unit,GRU)神经网络的Wasserstein生成对抗网络(Wasserstein generative adversarial network,WGAN)短期负荷预测模型。将Wasserstein距离作为生成对抗网络(Generative adversarial network,GAN)的损失函数,与传统GAN相比,可以解决其训练过程中存在的梯度消失、模式崩溃等问题。同时,其生成器和判别器模型采用GRU神经网络,用以解决循环神经网络中存在的梯度问题。通过与GRU神经网络模型、传统GAN模型、采用KL散度作为损失函数且生成器和判别器采用GRU的GAN模型进行对比试验,证明了所提出的新模型具有更好的预测精度和稳定性。To improve the accuracy of short-term load forecasting,a short-term load forecasting model of WGAN based on GRU neural network is proposed.The Wasserstein distance is taken as the loss function of GAN,compared with the traditional GAN,it can solve the problems of gradient disappearance and mode collapse in the training process.At the same time,its generator and discriminator model uses GRU neural network to solve the gradient problem in recurrent neural network.Compared with GRU neural network model,traditional GAN model,and GAN model with KL divergence as loss function and GRU as generator and discriminator,it is proved that the new model has better prediction accuracy and stability.

关 键 词：短期负荷预测 生成对抗网络 Wasserstein距离 GRU神经网络 电力系统 

分 类 号：TM715[电气工程—电力系统及自动化]