A deep neural network based battery state of charge: electric vehicle application

Abstract

Accurate State of Charge (SoC) estimation is critical for the performance, safety, and longevity of Li-ion batteries in electric vehicles (EVs). Traditional model-based approaches, such as equivalent circuit models and Kalman filters, often suffer from computational complexity and sensitivity to parameter variations, while data-driven methods face challenges in generalization due to limited training data or suboptimal algorithm selection. To address these limitations, this study proposes an intelligent SoC estimation process based on a deep neural network, which learns an algebraic expression describing the SoC evolution directly from voltage, current, and temperature measurements. A systematic comparative study evaluates three training algorithms Levenberg–Marquardt, Bayesian Regularization, and Conjugate Gradient under varying data splits to determine the optimal balance between precision and robustness. Results demonstrate that Bayesian Regularization achieves the highest accuracy when trained on 70% of the dataset, with 15% each for validation and testing, reducing the SoC prediction error to below 2%. This outcome not only validates the effectiveness of the proposed data-driven approach but also highlights the importance of algorithm and data split selection in overcoming the generalization challenges of existing methods. The study provides a practical and reliable solution for real-time EV battery management systems.