Neural Network-based SOC And SOH Estimation Method For Lithium-ion Batterie

Lithium-ion battery has the advantages of long life,low self-discharge rate and high energy density,which is a reliable energy storage device for electric vehicles.To further ensure the safety and stability of electric vehicles in the actual running process,a perfect battery management system(BMS)is needed.State of charge(SOC)and state of health(SOH)of batteries are important state parameters in BMS,but their accurate values cannot be obtained directly through measurement.Inaccurate SOC and SOH will affect the effective control of BMS on the vehicle battery pack.Therefore,how to design a reasonable method to obtain relatively accurate SOC and SOH indirectly from direct measurements is an urgent problem to be further studied.Based on neural networks,this paper sets up models suitable for SOC and SOH estimation of battery from the perspectives of basic network topology selection,network structure re-innovation,efficient optimization of network dynamic and static parameters,and solves the problem of SOC and SOH estimation.Specific research contents are as follows.Firstly,based on the Neware battery tester,dynamic stress test(DST),urban dynamometer driving schedule(UDDS)and Los Angeles 92(LA92)data were collected to verify the SOC estimation model.Then,the Oxford battery degradation data was analyzed to extract effective information for model verification of SOH estimation.Secondly,in order to achieve high precision SOC estimation,a BAS-RLSRecurrent-ELM model based on the beetle antennae search(BAS)algorithm and recursive least squares(RLS)algorithm to optimize recursive extreme learning machines is studied.Recurrent-ELM with time-delay lines can map the nonlinear relationship between the multi-moment terminal voltage,current,SOC at the old time and the SOC at the present time.And it can capture more dynamic information in input than ELM.BAS algorithm can automatically search the number of hidden layer nodes in Recurrent-ELM,and RLS algorithm can effectively train the weight of output layer.Under the condition that the training set and the test set are the same and the actual SOC is greater than 30%,the model has high precision.Thirdly,in order to achieve SOC estimation in more scenarios and increase the universality of the model,A conjugate gradient(CG)algorithm is utilized to optimize Multi-reservoir echo state networks(ESN)to build the CG-MESN model.In the MESN structure,the establishment of nonlinear mapping only requires the current moment terminal voltage,current and the current moment SOC,and the reservoir pools can be set up to remember the information of the past moment.CG algorithm is used to optimize the weight of the output layer for MESN,avoiding large-scale matrix inversion.In addition,an appropriate amount of Gaussian noise with an average value of 0 was added to the input data of the training set for data enhancement to improve the generalization ability of the model.Under the condition that the training set and the test set are only similar and the actual SOC value is greater than 10%,the model has good estimation performance.Fourthly,in order to achieve effective SOH estimation,a conjugate gradients(CG)algorithm is utilized to optimize broad learning system(BLS)to build a CG-BLS model.Based on the Oxford battery aging data set,three voltage rise times in the early charging period were extracted as shallow aging characteristics.A BLS structure capable of random feature mapping was used to establish the nonlinear and linear hybrid relationship between input and output.CG algorithm is used to optimize the weight of output layer and also avoid large-scale matrix inversion.In addition,an appropriate amount of Gaussian noise with a mean of 0 is added to the input of the training set to improve the generalization ability of the model.Simulation results show that the model has good SOH estimation performance,and data enhancement can improve the model’s estimation performance when the training set is insufficient.