| Lithium-ion batteries are widely used in power batteries of pure electric vehicles due to their high energy density and strong performance.As lithium-ion batteries are constantly recycled,their performance will gradually degrade.In order to ensure the safe and reliable operation of electric vehicles,its energy source,namely lithium-ion battery,should be monitored in real time.The state of health(SOH)and state of charge(SOC)of lithium-ion battery are two important parameters of the battery.It is of great significance to accurately obtain the health status and state of charge values of lithium-ion battery to improve the efficiency of the battery,prevent the battery from overcharging or overdischarging,ensure the safe operation of the vehicle,and reduce the cost of the battery.However,SOH and SOC are state quantities,which are difficult to be measured directly.In view of the shortcomings of current estimation methods,this paper carries out the following research:(1)State of health estimation of lithium-ion battery based on temporal convolutional networks.At present,the commonly used deep learning methods for lithium-ion battery health state estimation,such as recurrent neural network,have the problem of gradient disappearance when estimating long time series.This paper proposes a time series estimation network,namely time domain convolutional network.This network supports parallel computing,which can improve the efficiency of the algorithm.For the local capacity regeneration problem of lithium-ion battery in the cycle process,causal convolution and dilated convolution were used.In order to improve the accuracy of the estimation model,the spatio-temporal attention mechanism is introduced,and the original data-related aging features are extracted as the network input.The residual connection and random dropping technology were used to improve the training speed of the model and avoid the overfitting problem of the network.Finally,the effectiveness of the proposed method is verified by experiments on NASA data sets.(2)State of charge estimation of lithium-ion battery based on transfer learning.Traditional machine learning methods do not consider the influence of inconsistent data distribution when estimating the state of charge of lithium-ion batteries,and the trial-and-error training method cannot guarantee the optimal solution of model parameters.In addition,when the new battery lacks label data,the estimation accuracy of traditional methods is low.Therefore,this paper proposes a double long short-term memory network,which realizes feature transfer from source domain network to target network through transfer learning,and introduces backtracking line optimization algorithm to control a reasonable amount of knowledge transfer to improve the estimation accuracy of the network.Finally,it is verified in three different battery data sets and battery cells under dynamic stress conditions.Compared with traditional machine learning methods,the proposed method has higher estimation accuracy for the state of charge estimation of lithium-ion batteries. |
