Data-driven state of charge estimation for lithium-ion battery packs based on Gaussian process regression. (15th August 2020)

• Record Type: Journal Article
• Authors: Deng, Zhongwei; Hu, Xiaosong; Lin, Xianke; Che, Yunhong; Xu, Le; Guo, Wenchao
• Is Part Of: Energy. Volume 205(2020)
• Publication Date: 2020-08-15
• Subjects: Battery pack -- State of charge -- Data-driven -- Feature selection -- Gaussian process regression -- Autoregressive model; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2020.118000
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: Since a battery pack consists of hundreds of cells in series and parallel, inconsistencies between cells make it difficult to create an explicit model to simulate its behaviors effectively. Therefore, the widely used and sophisticated model-based methods (such as Kalman filters) are difficult to apply to SOC (state of charge) estimation of battery packs. In this paper, a data-driven method based on Gaussian process regression (GPR) is proposed to provide a feasible solution. Its superiority includes the ability to approximate nonlinearity accurately, nonparametric modeling, and probabilistic predictions. First, a feature extraction strategy, including data preprocessing, correlation analysis, and principal component analysis, is employed to obtain a compacted input set with a high correlation with SOC. Second, the squared exponential kernel function is used, and the automatic relevance determination is applied to optimize the weights of features. Third, besides the regular GPR model, an autoregressive GPR model is also constructed to further improve estimation accuracy and confidence. The experimental results verify that the autoregressive model has better SOC estimation performance than the regular model, and its estimation error under different dynamic cycles, temperatures, aging conditions, and even extreme conditions is lower than 3.9%, and the confidence interval is also much narrower. Highlights: A feature extraction method is employed to obtain a crucial and compacted data set. Gaussian process regression is used to predict the state of charge of battery pack. Automatic relevance determination is used to optimize the weights of features. An autoregressive model is created to improve estimation accuracy and confidence. The method is verified in different dynamic cycles, temperatures, and aging states.