融合注意力机制的CNN-GRU动车组蓄电池SOC估算方法  

CNN-GRU Battery SOC Estimation Method Fused with Attention Mechanism for Electric Multiple Units

作　　者：王升晖 田庆 刘力豪 冯恩来 于天剑[2] WANG Shenghui;TIAN Qin;LIU Lihao;FENG Enlai;YU Tianjian(CRRC Qingdao Sifang Rolling Stock Co.,Ltd.,Qingdao,Shandong 266111,China;School of Traffic&Transportation Engineering,Central South University,Changsha,Hunan 410075,China)

机构地区：[1]中车青岛四方机车车辆股份有限公司,山东青岛266111 [2]中南大学交通运输工程学院,湖南长沙410075

出　　处：《控制与信息技术》2023年第5期83-90,共8页CONTROL AND INFORMATION TECHNOLOGY

基　　金：湖南省科技人才托举工程项目(2022TJ-H14)。

摘　　要：碱性镍镉蓄电池是动车组辅助装置的核心能源,对其荷电状态(state of charge,SOC)进行精确估计,对于延长电池使用寿命及提高能量利用率具有显著意义。考虑现有的SOC估计方法在处理小样本电池循环数据时的局限性,文章提出了一种融合注意力机制的卷积-门控循环单元(CNN-GRU)电池SOC估算模型,并通过亚通达LPH140A型动车组镍镉蓄电池进行实验验证。模型通过卷积神经网络(CNN)提取电池循环数据中长序列的短时特征依赖关系,然后采用融合注意力机制的门控循环单元(GRU)对提取的特征数据进行长空间距离依赖关系的捕捉,从而达到更准确的电池SOC估算精度。同时为适应小样本电池循环数据SOC精准估算,文章将连续回归模型转化为分类问题,将电池SOC区间离散化,将最终预测结果转化为电池SOC区间离散值。实验结果表明,文章所提算法的预测结果与CNN-GRU算法的相比在均方根误差、平均绝对误差以及平均相对误差这3个关键指标上分别提升了18.90%、17.92%和19.78%。可见该模型在预测准确性和稳定性方面具有出色性能。Nickel-cadmium alkaline batteries are the core energy source for auxiliary devices of electric multiple units.Hence,an accurate estimation of their state of charge(SOC)is significantly important for prolonging battery life and improving energy efficiency.Given the limitations of existing SOC estimation methods when dealing with small-sample battery cycling data,this paper proposes an attention mechanism integrated convolutional neural network-gated recurrent unit(CNN-GRU)model for battery SOC estimation,and experimental validation is conducted on the LPH140A model nickel-cadmium batteries used in electric multiple units.The model employs a convolutional neural network(CNN)to extract short-term feature dependencies from long sequences within the battery cycling data.Then,an attention mechanism-integrated gated recurrent unit(GRU)is adopted to capture long spatial distance dependencies of the extracted feature data,resulting in more precise battery SOC estimation.To precisely estimate the SOC of small-sample battery cycling data,this paper transforms the continuous regression model into a classification problem,discretizes the battery SOC ranges,and converts the final prediction result into discrete SOC values.The experimental results show that compared with the CNN-GRU algorithm,the proposed approach improves three key metrics—root mean square error,mean absolute error,and mean relative error by 18.90%,17.92%and 19.78%,respectively,demonstrating impressive prediction accuracy and stability.

关 键 词：镍镉蓄电池 注意力机制 卷积-门控循环单元 SOC估计 

分 类 号：TM912.2[电气工程—电力电子与电力传动]