Estimation Of State Of Health Of Lithium-ion Batterier Based On Graph Neural Network

Currently,environmental pollution and energy crisis are becoming more and more serious,and lithium-ion batteries are attracting much attention as a clean energy source.State of Health(SOH)is an important parameter reflecting the aging degree and health status of lithium-ion batteries,and accurate estimation of SOH is essential to grasp the battery health status.However,lithium-ion battery performance degradation exhibits complex nonlinear variations,and measurable parameters such as voltage,current and temperature do not directly reflect the SOH of the battery.In order to estimate SOH accurately,this thesis extracts health indicators(HIs)characterizing SOH of lithium-ion batteries based on the analysis of the variation characteristics of charge-discharge cycle data,and establishes the relationship between HIs and SOH using deep learning methods,and investigates the method of estimating SOH of lithium-ion batteries based on graph neural networks(GNN).Firstly,the characteristics of charge/discharge curves are studied,the shapelet-based HIs are proposed,and the shapelet-based SOH estimation method for lithium-ion batteries is designed.Then,the correlations among the HIs are deeply explored,and the multi-health indicator-based SOH estimation method for lithium-ion batteries is designed.Finally,according to the characteristics of different types of HIs,multi-models are introduced to mine the deep information of different types of HIs,and the multi-model feature fusion-based SOH estimation method for lithium-ion batteries is designed.The main innovation points and research of the thesis are as follows.(1)In order to reduce the sensitivity to irrelevant data in the discharge curve,a shapeletbased SOH estimation method for Li-ion batteries is proposed,which is divided into an offline processing stage and an online estimation stage.First,in the offline processing stage,the variation characteristics of the discharge voltage curve are analyzed,and the voltage data are divided into multiple sub-series as shapelet candidates using a sliding window.Then,the best candidate object is selected as the shapelet by correlation analysis.Finally,in the online estimation stage,the long short term memory(LSTM)neural network achieves online SOH estimation by calculating the subinterval of the shortest Euclidean distance between the shapelet and the discharge voltage curve as input.The maximum relative error of SOH estimation results for all tested cells is kept within 5%,which has a good stability.(2)In order to further improve the SOH estimation accuracy,a multi-health indicator-based SOH estimation method for lithium-ion batteries is proposed by mining the temporal and spatial features of HIs.An undirected graph based on HIs is constructed as input,LSTM mines the temporal features of HIs with cyclic cycles,and Graph Sample and Aggregate(Graph SAGE)mines the interconnected spatial features of HIs.In the SOH estimation results of all tested cells,the maximum relative error is kept within 5%,and the mean absolute percentage error is reduced to 0.49% when compared with the SOH estimation results based on shapelet,which proves that the proposed method achieves better SOH estimation.(3)In order to enrich the feature diversity of HIs and fully complement the information,a multi-model feature fusion-based SOH estimation method for Li-ion batteries is proposed.First,the HIs are extracted from the multi-source signals(voltage,current and temperature)of the charging and discharging processes,and the HIs are classified into positive,negative and weak correlations using Pearson correlation coefficients.Then,three feature vectors are obtained by mining different types of HIs using multiple models.Finally,the feature space of the three feature vectors is fused to achieve SOH estimation.Among the SOH estimation results for all the tested cells,the method achieves better SOH estimation accuracy compared to the single model and different subsets of HIs.