A novel feature optimization and ensemble learning method for state-of-health prediction of mining lithium-ion batteries

• Published in: Energy (Oxford) 2024-07, Vol.299, p.131474, Article 131474
• Main Authors: Feng, Juqiang, Cai, Feng, Zhao, Yang, Zhang, Xing, Zhan, Xinju, Wang, Shunli
• Format: Article
• Language: English
• Subjects: batteries; Canonical correlation analysis; data collection; energy; Ensemble learning method; Health features optimization; Mining lithium-ion batteries; multivariate analysis; prediction; State of health prediction
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2024.131474

In this work, accelerated aging tests at different temperatures are conducted for 228Ah mining lithium-ion batteries, and an SOH prediction method with health features (HFs) optimization and ensemble learning method is proposed. Firstly, six health features are extracted from cyclic charge/discharge data. Simultaneously, to solve the problem of a single feature not fully reflecting the SOH at multiple temperatures, canonical correlation analysis is introduced to construct the feature fusion vector to obtain the comprehensive health feature (C–HF). Secondly, the complementary ensemble empirical mode decomposition method is used to smooth the features and the SOH of the battery to extract the raw data of the battery to be tested in the stable frequency range. Then, four different datasets are used to comprehensively evaluate the performance of C–HF in the ensemble learning method. Compared with other HFs, the optimized feature C–HF has the best SOH prediction in all datasets, with high prediction accuracy and strong robustness. Finally, we compare SVM, LSTM, and LSTM-SVM with the proposed method in this paper for SOH prediction. Whether 70 % or 30 % of training datasets are used, the proposed method's estimation results are closer to the actual in SOH. Figure 1. The framework based on ensemble learning for battery SOH prediction.The graphical abstract clearly illustrates the ensemble learning approach for LSTM and SVM models. Initially, the initial values of the optimization methods are set for both the LSTM and SVM models. Subsequently, the optimal hyperparameters of each model are determined using the sparrow search algorithm (SSA), and the error of the training set on each model is recorded. Finally, the training set's LSTM and SVM model errors are compared, and the model with the minor error is selected as the final model. Suppose the error difference between the LSTM and SVM models falls below a certain threshold. In that case, the models are combined in a weighted manner, with the final prediction equal to the weighted sum of the predictions of both models. [Display omitted] •Proposed optimization of HFs using CAA to obtain C–HF features.•Proposed a novel ensemble learning algorithm for accurately estimating SOH of MLIB.•Aging experiments were conducted on MLIBs across diverse temperature conditions.•The proposed methods have undergone effective validation on experimental datasets.