基于深度学习的数据中心VRLA电池健康状态估计  

SOH estimation of VRLA battery in data center based on deep learning

作　　者：陈志鹏 左信[1] 宋东力 CHEN Zhipeng;ZUO Xin;SONG Dongli(Department of Automation,College of Information Science and Engineering,China University of Petroleum(Beijing),Beijing 102249,China;Changan Communication Technology Co.,Ltd.,Beijing 100029,China)

机构地区：[1]中国石油大学(北京)信息科学与工程学院自动化系,北京102249 [2]长安通信科技有限责任公司,北京100029

出　　处：《电源技术》2023年第2期210-214,共5页Chinese Journal of Power Sources

摘　　要：健康状态(state of health,SOH)是数据中心阀控式铅酸电池(value regulated lead acid,VRLA)容量及安全管理的关键指标,而常用的SOH测量方法因检测过程放电时间长、深度大无法满足运营需求,因此对数据中心VRLA电池的健康状态估计研究是非常必要的。针对SOH数据驱动建模存在的估计精度低的问题,提出一种基于时空注意力(spatio-temporal attention,STA)和长短期记忆(long short term memory,LSTM)网络的STA-LSTM深度学习模型。该模型用时空注意力机制在输入数据的特征和时间步上分配注意力权重从而生成新的输入,使用LSTM网络对新的输入进行编码以及实现SOH估计输出。基于电池放电深度50%的数据建模应用结果表明,STA-LSTM模型取得最优估计精度,注意力机制的引入提升黑箱模型的收敛速度、估计精度及物理可解释性。State of health(SOH)is a key index for the capacity and safety management of valve regulated lead acid(VRLA)batteries in the data center.The commonly used SOH measurement methods can not meet the operation requirements due to long discharging time and large depth in the detection process.Therefore,it is very necessary to study SOH estimation of VRLA batteries.Aiming at the problem of low estimation accuracy in data-driven modeling,a deep learning model STA-LSTM based on spatio-temporal attention(STA)and long short term memory(LSTM)networks is proposed.The model uses spatio-temporal attention mechanism to allocate attention weights on the characteristics and time steps of input data,so as to generate new inputs;then uses LSTM network to encode new inputs and realize SOH estimation output.The modeling application results based on data of the battery 50% discharge depth show that STA-LSTM model achieves the optimal estimation accuracy,and the attention mechanism improves the convergence speed,estimation accuracy and physical interpretability of the black box model.

关 键 词：时空注意力 长短时记忆 SOH估计 VRLA电池 

分 类 号：TM912[电气工程—电力电子与电力传动]