Cloud-Based Artificial Intelligence Framework for Battery Management System
Year: 2026
Type: conference-paper
Abstract: Accurate state-of-charge (SoC) estimation remains a critical challenge in lithium-ion battery management systems because of measurement noise, model uncertainty, and operating condition variability. Coulomb Counting (CC) provides reliable short-term SoC dynamics. However, it accumulates drift over time. Additionally, model-based estimators such as Extended Kalman Filters (EKF) depend on accurate parameter identification. Although purely data-driven models are flexible, they often suffer from limited generalization when trained on restricted operating conditions. This paper presents a physics-guided hybrid SoC estimation framework in which a deep learning model is first pretrained using data generated from a Doyle-Fuller-Newman (DFN) electrochemical battery model and subsequently fine-tuned on real-world battery datasets. The pretrained model learns electrochemically consistent behavior across a wide range of current profiles, temperatures, and initial SoC conditions. In real-time operation, the fine-tuned network acts as a correction anchor for Coulomb Counting, periodically compensating accumulated drift without relying on open-circuit voltage measurements or online equivalent circuit parameter identification. The proposed approach is validated through extensive simulations and real data fine-tuning experiments. Results demonstrate a reduction in SoC root-mean-square error of approximately 57% compared to CC+OCV correction and approximately 46% compared to EKF-based estimation across 2322 dynamic test profiles. These findings indicate that physics-based pretraining combined with data-driven adaptation provides a robust and scalable solution for accurate SoC estimation under diverse operating conditions.
Institution: German International University
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202296239