A hybrid SOH prediction estimation framework for Lithium-ion Batteries integrating feature engineering and optimized Gaussian process regression

• Published in: Physica Scripta
• Main Authors: Fan, Xingming, Huang, Wenxi, Wang, Chao, Xu, Yanke, Zhang, Xin
• Other Authors: Guangxi Key Technologies R&D Program, National Natural Science Foundation of China, Natural Science Foundation of Guangxi Province
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: IOP Publishing 2025
• Subjects: Vae; Beluga; Beluga whale; Beluga*
• Online Access: https://doi.org/10.1088/1402-4896/ae1ad8; https://iopscience.iop.org/article/10.1088/1402-4896/ae1ad8; https://iopscience.iop.org/article/10.1088/1402-4896/ae1ad8/pdf

Abstract Accurate assessment of the State of Health (SOH) for Lithium-ion Batteries (LIBs) is critical for ensuring the operational safety of battery management systems. To overcome the accuracy limitations of existing SOH prediction methods, this study proposes an integrated framework that synergizes feature engineering, an Improved Beluga Whale Optimization (IBWO) algorithm, and Gaussian Process Regression (GPR). The methodology involves several key stages. First, Health Features (HFs) are extracted from raw charge/discharge data using a Variational Autoencoder (VAE). Optimal HFs are then selected through Pearson correlation analysis and augmented using a Generative Adversarial Network (GAN). Subsequently, the Variational Mode Decomposition (VMD) algorithm is applied to adaptively decompose the HFs. The decomposed components are predicted and reconstructed using GPR models, effectively suppressing noise within the HFs. Following this, GPR models establish a nonlinear mapping between the reconstructed HFs and the actual SOH to enable prediction. Meanwhile, to address the challenge of GPR hyperparameter selection, an IBWO algorithm is proposed. The IBWO incorporates Circle chaotic mapping initialization, quasi-opposition-based learning, and a dynamic weight adjustment strategy to improve optimization performance. Experimental validation on the NASA and MIT public battery dataset demonstrates the framework’s superior performance, with prediction error indicators consistently below 1%, confirming high accuracy and robust generalization capability.