加窗 LSTM 网络的动力电池 SOC 预测优化算法
2021-09-13李泓沛刘桂雄
李泓沛 刘桂雄



摘要:动力电池荷电状态(stateof charge, SOC)的实时预估对提高电动汽车电池利用率、优化电动汽车性能意义重大。该文提出一种可学习电池诸多干扰因素的加窗 LSTM 网络动力电池 SOC 预测算法,首先构建加窗 LSTM 网络的动力电池 SOC 预测算法,由动力电池电流、电压、温度结合历史数据实时预测动力电池 SOC;其次,分析并提出加窗 LSTM 网络优化超参数为 LSTM 网络的窗口大小、cell 数量、全连接层激活函数,试验证明,优化 LSTM 网络能实时预测动力电池 SOC,在不同充放电、路谱工况下的动力电池 SOC 预测最大误差均不超过2.20%。
关键词:动力电池;荷电状态;长短期记忆;超参数优化
中图分类号: TH89文献标志码: A文章编号:1674–5124(2021)12–0087–05
SOC prediction optimization algorithm for power battery based on LSTM network with window function
LI Hongpei,LIU Guixiong
(School of Mechanical and Automotive Engineering, South ChinaUniversity of Technology, Guangzhou 510640, China)
Abstract: Real-time estimation of the state of sharge (SOC) of the power battery is of great significance for improving battery utilization and optimizing the performance of electric vehicles. Firstly, a long short term memory (LSTM) network with window function based SOC prediction algorithm is proposed, to predict the SOC of the power battery in real time by combining the power battery current, voltage and temperature with the historicaldata. Analysisof the LSTM network with windowfunctionindicates that the windowsize, number of cells, and full connection layer activation function mode are hyperparameters that are more relevant toinput/outputfor the LSTM network with windowfunction,and itsoptimizationalgorithmis proposed. Experiments showed that the optimized LSTM network with window function could predict the SOC of the power cell in real time, and the maximum error of the SOC prediction was less than 2.20% under different charging, discharging, and driving conditions.
Keywords: power lithium battery; state of charge; long short term memory; hyper-parameter optimization
0引言
動力电池荷电状态(state of charge, SOC)是指动力电池剩余容量与最大可用容量之比[1],它是关系到电动汽车电池利用率、优化电动汽车性能的重要参数,其难以直接测量,受充放电速率、电池状态诸多因素影响,SOC 准确实时预估电池领域研究热点。目前动力电池荷电状态的实时估计方法,主要包括安时积分法[2]、开路电压法[3]、等效电路模型[4-10]、深度学习模型[11-14]等。安时积分法易受到电流传感器噪声影响,积分计算产生累计误差[2];开路电压法需将电池较长时间静置达到稳定,不利于实现实时测量[3];时域等效电路模型需用滤波方法[4]对电池内部参数建模、辨识,采用自适应卡尔曼滤波[5-7]、粒子滤波[8-9]与机器学习结合算法[10]实现 SOC 预测,但在电流变化较大动态工况下易产生发散[11]。深度学习模型应用非……
