Research On State Of Health Estimation And Remaining Useful Life Prediction Of Lithium-ion Battery

Lithium ion battery is the most widely used secondary battery for energy supply and storage,and its application fields include industrial equipment,new energy automobile,electronic products and so on.In various applications of lithium-ion batteries,the battery management system is a very important part.State of Health(SOH)estimation and Remaining Useful Life(RUL)prediction of lithium-ion battery are key functions of battery management systems,and it is of great significance to study them.The effect of data-driven methods is proper,but it is difficult to extract the direct health indicators of battery on-line.Therefore,the indirect on-line state of health estimation and remaining useful life prediction methods are studied in this paper.In order to extract indirect health features,two feature extraction methods are used.The health indicator of constant current charge time is extracted and its validity is proved by Spearman rank correlation coefficient.Gramian Angular Field(GAF)is used to transform time series into two-dimensional data to represent the implicit information of time series effectively.Aiming at solving the problem that artificial feature extraction needs experience,Convolutional Neural Networks(CNN)is used to automatically extract features.A feature extraction method based on GAF-CNN is established,and the input of it is charging voltage fragment.Compared with the traditional neural network,Simple Deterministically Constructed Cycle Reservoirs with Regular Jumps(CRJ)has the advantages of simple training,high precision and echo characteristics,and it is used for SOH estimation.In this paper,the parameters of CRJ are optimized by Arithmetic Optimization Algorithm(AOA),which solves the problem of difficult selection of CRJ parameters.Many time series prediction methods have the problems of no time series dependence between output values and low accuracy.The prediction output of Sequence to Sequence(Seq2Seq)model has the advantages of timing and high accuracy of short-term prediction,thus it can be applied to the short-term prediction of RUL.First,Seq2 Seq was used to predict the late-cycle constant current charging time sequence,and the relationship model between constant current charging time and capacity was established by CRJ.Then,the predicted late-cycle health indicator sequence was input into the established relationship model for capacity prediction,thus RUL is obtained.Among them,the relational model is the SOH estimation model,which is the basis of life prediction,but its input is the predicted constant current charging time.In this paper,two kinds of extracted health features are input into AOA-CRJ model for SOH estimation.Compared with other methods,the results show that AOA-CRJ has the best estimation accuracy,stability and generalization ability.When compared with the CNN and charging voltage fragment based methods,the results show that the feature extracted by GAF-CNN is effective.The RUL prediction with different cycle length is studied,which can be divided into 80 and 50 cycles.The influence of the parameter combination of Kernel Extreme Learning Machine(KELM),the leaking-rate of CRJ and the randomness of neural networks on the prediction is also studied.In the RUL prediction for B6 battery after 50 cycles,1.4Ah was taken as the failure threshold,and the error of CRJ was only 1.The results show that the health indicator predicted by Seq2 Seq is effective,CRJ has the best comprehensive performance and does not have the trouble of randomness of many neural networks.KELM and CRJ have the best parameters.