基于改进Bi-LSTM网络下的多时变状态锂电池剩余寿命预测方法

Remaining useful life prediction method of lithium-ion battery in multiple time-varying states based on an improved Bi-LSTM network

针对现有模型仅考虑一种内部状态对锂电池性能退化影响的问题,同步建立3个模型分别预测3种时变状态随锂电池性能退化的变化轨迹,并以内阻与温度预测为基础实现锂电池容量的实时更新;针对传统神经网络中的sigmoid与ReLU激活函数存在梯度消失与神经元坏死问题,在双向长短时记忆(bi-directional long short term memory,Bi-LSTM)网络与全连接网络中引入一种新的Mish激活函数,使模型以平稳的梯度流提取更高质量的特征用于剩余使用寿命(RUL)预测的建模分析。最后利用蒙特卡洛(Monte Carlo,MC)与Dropout技术对锂电池RUL的预测结果不确定性进行分析。在美国Kristen教授课题组所公开的锂电池数据集上进行对比试验的结果表明,所提改进Bi-LSTM模型预测的均方误差(mean squared error,MSE)、平均绝对误差(mean absolute error,MAE)与R^(2)可达9.16×10^(-5)、0.00795、99.794%。随着获取数据量的增加,模型对锂电池RUL预测的精度越高,RUL平均预测误差可达2.3个循环,验证了所提模型能有效地实现锂电池循环RUL的实时更新。

In response to the problem that the existing models only consider the influence of one internal state on the performance degradation of lithium battery,three models are established simultaneously to predict the trajectory of three time-varying states with the performance degradation of lithium battery,and to realize the real-time updating of lithium battery capacity based on the prediction of internal resistance and temperature.In response to the problems of gradient disappearance and neuron necrosis in the sigmoid and ReLU activation functions of traditional neural networks,a new MISH activation function is introduced in bi-directional long short term memory(Bi-LSTM)network and fully connected network,which make the model extract higher quality features with smooth gradient flow for RUL prediction.Finally,the Monte Carlo(MC)and Dropout techniques is used to analyze the uncertainty of the prediction results of lithium battery RUL.The results of the comparison test on the lithium battery dataset disclosed by Prof.Kristen's group in the United States show that the mean squared error(MSE),mean absolute error(MAE)and R^(2)predicted by the improved Bi-LSTM model can reach 9.16×10^(-5),0.00795,and 99.794%.With the increase of the amount of data obtained,the higher the accuracy of the model's prediction of lithium battery RUL,the average prediction error of RUL can reach 2.3 cycles,which verifies that the proposed model can effectively achieve real-time update of RUL of lithium batteries.