基于改进LSTM的低压配电网日线损率预测方法  

A daily line loss prediction method for medium and low voltage distribution networks based on improved LSTM

作　　者：边舒芳 张伟 BIAN Shufang;ZHANG Wei(State Grid Jibei Electric Power Company Limited,Tangshan Fengrun District Power Supply Branch Co.,Ltd.,Tangshan 063000,Hebei China;State Grid Jibei Electric Power Company Limited,Luanzhou Power Supply Branch Co.,Ltd.,Luanzhou 063700,Hebei China)

机构地区：[1]国网冀北电力有限公司唐山市丰润区供电分公司,河北唐山063000 [2]国网冀北电力有限公司滦州市供电分公司,河北滦州063700

出　　处：《粘接》2025年第1期188-192,共5页Adhesion

摘　　要：针对目前低压配电网日线损预测精度较低,原始电力数据缺失和异常值问题,提出了一种包含数据预处理和改进LSTM预测网络的双阶段线损率预测,及基于GAN扩充样本,增加样本多样性的方法。改进LSTM预测网络为一个融合多层LSTM的R-CNN深度学习网络架构,可提取电力数据特征以及时间维度信息。通过实验,与Bi-LSTM、LSTM自动编码器、CNN-GRU、BL-Seq2seq相比,所提预测网络的RMSE、MAE、RA2、训练时间指标综合性能最优。实验结果表明,所提预测网络在低压配电网日线损率预测中可以获得更好的预测精度,且模型训练时间最短。In order to solve the problems of low accuracy of daily line loss prediction,lack of original power data and outliers in the current low-voltage distribution network,a two-stage line loss rate prediction method including data preprocessing and improved LSTM prediction network and a method based on GAN to expand the sample and increase the sample diversity were proposed.The improved LSTM prediction network was an R-CNN deep learning network architecture with multi-layer LSTM,which can extract features and temporal dimension information from power data.Through experiments,compared with Bi LSTM,LSTM autoencoder,CNN-GRU,BL-Seq2seq,the pro⁃posed prediction network had the best comprehensive performance in terms of RMSE,MAE,RA2,and training time indicators.The experimental results showed that the proposed prediction network can achieve better prediction ac⁃curacy in predicting the daily line loss rate of low-voltage distribution networks,and the model training time is the shortest.

关 键 词：低压配电网 线路损失 深度学习 卷积神经网络 循环神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TM711[自动化与计算机技术—计算机科学与技术]