Battery state estimation via hybrid P2D modeling and adversarial deep learning in electric vehicles

Abstract

Accurate multi-state estimation of lithium-ion batteries (LIBs) is essential for electric vehicle (EV) battery management systems. Existing electrochemical models face challenges in parameter calibration, while purely data-driven methods lack physical interpretability. To address these limitations, this study proposes an integrated framework combining a pseudo-two-dimensional (P2D) electrochemical model with a generative adversarial network-long short-term memory (GAN-LSTM) architecture. A hybrid simulated annealing-particle swarm optimization (SA-PSO) algorithm was developed for non-invasive parameter calibration of the Tesla Model S battery P2D model, achieving a mean absolute error (MAE) of 0.027 V in terminal voltage prediction during 1C constant-current discharge. The calibrated model, integrated with vehicle dynamics simulations, generated physics-based multivariate time-series data across diverse operational scenarios. These data were utilized to train the GAN-LSTM framework, which synergizes LSTM’s temporal modeling with GAN’s adversarial training for robust state estimation. Experimental results demonstrate the framework’s high accuracy, with determination coefficients (R²) of 0.9965 for state of charge (SOC) and 0.9843 for state of health (SOH). This work establishes a novel methodology that bridges electrochemical mechanisms with data-driven modeling, providing a physics-informed solution for multi-state battery estimation without relying on artificial feature engineering or unvalidated assumptions. The proposed framework offers practical value for next-generation battery management systems in real-world EV applications.