A novel graph-based framework for state of health prediction of lithium-ion battery. (February 2023)

• Record Type: Journal Article
• Authors: Yao, Xing-Yan; Chen, Guolin; Pecht, Michael; Chen, Bin
• Is Part Of: Journal of energy storage. Volume 58(2023)
• Publication Date: 2023-02
• Subjects: LIBs lithium-ion batteries -- SOH state of health -- HIs health indicators -- CL-GraphSAGE the proposed method in this paper -- GNN graph neural network -- GraphSAGE graph sample aggregate -- CNN convolutional neural network -- RNN recurrent neural network -- LSTM long short-term memory neural network -- CC the current curve of the constant current -- CV the current curve of the constant voltage -- G=VE graph G, E and V are the edge and node set -- RMSE root mean square error -- MAPE the average absolute percentage error -- R2 coefficient of determination; Lithium-ion battery (LIB) -- State of health (SOH) -- GraphSAGE -- Graph neural network (GNN); Energy storage -- Periodicals; Energy storage -- Research -- Periodicals; 621.3126
• DOI: 10.1016/j.est.2022.106437
• Languages: English
• ISSNs: 2352-152X

Abstract

Abstract: State of health (SOH) plays a vital role in lithium-ion batteries (LIBs) safety, reliability and lifetime. Health indicators (HIs) are a powerful approach to predict battery SOH. The existing methods for battery SOH prediction according to HIs only consider the temporal features of HIs. The spatial features of interdependence between HIs enrich predicational information especially for the limited data. This paper proposes a novel framework CL-GraphSAGE to predict battery SOH based on graph neural network (GNN), which takes into both temporal and spatial features of HIs. Firstly, the Pearson's correlation coefficients between HIs and SOH are obtained to extract highly correlated HIs to build a graph. Subsequently, the temporal features are extracted by convolutional neural network (CNN) and long short-term memory neural network (LSTM). Finally, the spatial features are obtained by the graph sample aggregate (GraphSAGE) to propagate messages on a pre-defined graph structure, which uncovers the deep information among HIs. The effectiveness of the proposed approach in predicting battery SOH is verified by MIT, NASA datasets and the experimental datasets, and compared with CNN, LSTM and graph convolutional network and graph attention network. The results show that the root mean square error of the proposed approach CL-GraphSAGE can achieve 0.2 %, and the different datasets verify its feasibility. Highlights: A novel framework base on graph neural network (GNN) for battery SOH prediction is proposed. Both the temporal features and the spatial features of health indicators are employed. HIs highly correlated with SOH are extracted to form a spatial-temporal graph. GNN is the first time to use for battery prediction problem. The accuracy and feasibility of CL-GraphSAGE for SOH prediction are validated in various working conditions.