Research On The State-of-health Estimation Of Lithium-ion Power Batteries Based On Neural Network Methods

In order to slow down the massive consumption of traditional fossil energy and reduce the pollution from the exhaust emissions of fuel vehicles,the country vigorously promotes the new energy vehicle industry.As the main power source of electric vehicles,the sta-ble and reliable operation of the power battery is the most essential to ensure the life and property safety of the driver.The battery management system is responsible for monitoring the operational indicators of the power battery among which the state of health（SOH）es-timation is one of its core functions.Accurate SOH estimation can improve the utilization efficiency and duration ability of the power battery,extend its useful life,and help users achieve the best balance between system safety and economic benefits.In addition,accu-rate SOH estimation also helps the system to make better energy management decisions.With the advantages of high energy density,low self-discharge rate and long useful life,lithium-ion batteries have become the mainstream on-vehicle power battery.In artificial intelligence times,the neural network（NN）method provides a brand new solution idea to the SOH estimation problem.Instead of analyzing the complex aging mechanism inside the battery,online SOH estimation can be accomplished only by learning the historical ag-ing data.In this paper,two SOH estimation methods based on NN models for lithium-ion batteries are proposed,using the data of ternary lithium-ion batteries whose cathode mate-rial is nickel-cobalt-manganese collected by the research group in the laboratory,the data of NASA and the data of CALCE as the research objects.The main research work is as follows:（1）Taking ternary lithium-ion power battery as the research and test object,the internal structure,working principle and aging causes are introduced.Then the battery aging test platform is designed and built to obtain battery aging data by conducting cyclic charge/dis-charge tests on brand-new lithium-ion power batteries at a constant charge/discharge rate,which provides data support for the implementation of the SOH prediction method in Chap-ter 4.（2）The SOH estimation method of lithium-ion batteries based on the long short-term memory（LSTM）NN combined with attention mechanism（AM）is proposed to address the problems of predicted values drifting and inaccurate prediction in the late stage of prediction by traditional NN methods.First,the lithium-ion battery capacity data are preprocessed by moving average filter for denoising.Then,based on the battery capacity data with different datasets and discharge rates,the SOH prediction is completed by giving different weights to the LSTM hidden layer by AM to enhance the important information.Finally,the model can be migrated to brand-new lithium-ion battery data for whole life testing and is compared with the currently popular NN prediction methods.The experimental results show that the AM-LSTM-based SOH prediction method has the features of accurate whole life prediction,simple network structure and high model robustness,and is expected to be applied to the actual operation of electric vehicles.（3）To solve the problem of the inefficiency of manually adjusting NN hyperparameters and the poor accuracy of directly predicting SOH,a SOH prediction method for lithium-ion batteries is proposed using the empirical mode composition（EMD）combined with the the bidirectional gated recurrent unit（Bi GRU）NN which is optimized by slime mould algo-rithm（SMA）.First,a battery SOH sequence is decomposed into some intrinsic mode func-tion sequences and a final residual term by EMD.Then,SMA is used to search the optimal combination of hyperparameters for the Bi GRU NN model.Finally,combined models are constructed to predict the SOH of lithium-ion batteries.According to the experimental re-sults,the combined model exhibits strong robustness and generality,and shows excellent prediction performance on both the Li Co O₂battery dataset from the Center for Advanced Life Cycle Engineering（CALCE）at the university of Maryland and the ternary lithium-ion battery dataset from our own research group.This indicates that the combined model is of great practical value.In summary,two kinds of SOH estimation methods based on NNs are proposed for lithium-ion power batteries in this paper,which lays a solid foundation to guarantee the safe and stable operation of lithium-ion batteries in applications such as electric vehicles and energy storage fields.