SOH Prediction Methods Of Lithium-ion Battery Under Random Discharge Conditions

Lithium-ion batteries have been widely used in the Battery Energy Storage System（BESS）due to their high energy density,low self-discharge rate,and relatively long life,etc.Accurate prediction of the State Of Health（SOH）of the battery is essential to ensure the safe operation of the BESS.However,the standard charging and discharging modes of lithium batteries do not reflect the operating conditions of lithium batteries in practical applications.Therefore,it is very important to study the SOH prediction of batteries under random discharge conditions.By analyzing the experimental data of randomly discharged batteries recorded by NASA,the thesis conducts research on SOH prediction based on the condition monitoring data of lithium-ion batteries.The main contents are as follows:（1）Aiming at the problem that standard tests are difficult to measure the reference capacity of lithium-ion batteries under constant current charging and random discharging conditions,an extraction method of Health Indicator（HI）is proposed.From the state parameters that can be measured online,HIs used to characterize the SOH battery are extracted.Considering that the battery is fully discharged,several potential HIs containing battery health degradation information are extracted from the data collected by the voltage and current sensors,and then these potential HIs are dimensionally reduced based on the Principal Component Analysis（PCA）to remove redundant information.Secondly,considering that the battery is not fully discharged under random discharge conditions,the internal resistance within a short time and voltage standard deviation within a short time are extracted from some random discharge data to represent the battery SOH.And the high correlation between the two HIs and the battery capacity is verified by the Pearson correlation coefficient.（2）Deep Extreme Learning Machine（DELM）can use extreme learning machine-autoencoder to initialize the input weights and hidden layer biases of each hidden layer.This process does not require reverse fine-tuning,so the network learns fast,good generalization performance.But since the input weights and hidden layer biases of DELM are randomly set,these significantly affect the prediction accuracy of the original DELM.Therefore,the thesis uses the Sparrow Search Algorithm（SSA）to optimize two key parameters of the DELM network,and uses the fusion HI obtained by the PCA algorithm as the network input to predict the battery SOH.Compared with other machine learning methods,it shows that the designed SOH prediction method has high precision and strong robustness.（3）Similar to other optimization algorithms,SSA also has shortcomings such as easy to fall into local optimal solution and low accuracy.To this end,this thesis studies an improved SSA to further improve the prediction performance of deep extreme learning machines.First,two indirect health indicators were extracted from random partial discharge voltage and current,and the appropriate extraction interval was selected as the input of DELM by Pearson correlation analysis.Then,the exploration and development process of the sparrow search algorithm is balanced by combining the elite reverse learning algorithm and the Cauchy-Gaussian mutation strategy to prevent the algorithm from falling into a local optimum.Compared with other combined methods of improved optimization algorithm and DELM network,the experimental results show that the RMSE of ISSA-DELM is 0.134%,and the proposed method has higher prediction accuracy.