基于BiLSTM-MhSa-ResNet的储能电站SOC预测  

SOC prediction of energy storage power station based on BiLSTM-MhSa-ResNet

作　　者：廖圣瑄 李林[2] 丁伟 唐起超 唐志军 杨继盛 LIAO Shengxuan;LI Lin;DING Wei;TANG Qichao;TANG Zhijun;YANG Jisheng(Longyuan(Beijing)Wind Power Engineering&Consulting Co.,Ltd.,Beijing 100034,China;University of Electronic Science and Technology of China,Chengdu 610000,China;Longyuan Qinghai New Energy Development Co.,Ltd.,Geermu 816099,China)

机构地区：[1]龙源(北京)风电工程设计咨询有限公司,北京100034 [2]电子科技大学,四川成都610000 [3]龙源(青海)新能源开发有限公司,青海格尔木816099

出　　处：《电子设计工程》2023年第22期78-82,共5页Electronic Design Engineering

基　　金：龙源电力集团股份有限公司科技创新项目(GJNY-20-15)。

摘　　要：储能电站荷电状态(SOC)评估对储能电站安全稳定运行起到重要作用。针对当前传统神经网络预测精度不足的问题,提出一种双向长短期记忆多头残差网络(BiLSTM-MhSa-ResNet)进行SOC预测。该模型使用多头自注意力机制提高了特征值的获取能力,通过残差神经网络解决了梯度异常问题,应用双向长短期记忆网络捕获了长期依赖关系,从而提高了预测能力。实验结果表明,采用BiLSTM-MhSa-ResNet进行充电SOC预测时,平均绝对误差为1.02%,均方根误差为1.31%,决定系数为0.998,提高了SOC预测的准确性。进行放电SOC预测实验时,该模型也具有较好的训练效果。State-of-Charge(SOC)assessment of energy storage power stations is critical to the safe and stable operation of the plants.For the problem of insufficient prediction accuracy of traditional neural network,a Bi-directional Long Short-Term Memory-Multi-headed Self-attention-Residual Network(BiLSTM-MhSa-ResNet)is proposed for SOC prediction.The model uses a multi-head self-attention mechanism to improve the acquisition ability of eigenvalues,solves the gradient anomaly problem through a Residual Neural Network,and applies a Bi-directional Long Short-Term Memory Network to capture long-term dependencies,thereby improving the prediction ability.The experimental results show that when BiLSTM-MhSa-ResNet is used for charging SOC prediction,the average absolute error is 1.02%,the root mean square error is 1.31%,and the determination coefficient is 0.998.It means that the proposed method improves the accuracy of SOC prediction efficiently.Furthermore,in the discharge SOC prediction experiment,the model achieve great training result as well.

关 键 词：荷电状态 深度学习 锂电池 神经网络 储能电站 

分 类 号：TM912[电气工程—电力电子与电力传动]