Uncertainty aware hybrid learning framework for fast and safe charging of lithium-ion batteries using multi-fidelity observers.

Accurate and real-time estimation of the state of charge (SoC) is critical for ensuring the performance, safety, and longevity of lithium-ion batteries in electric vehicles (EVs). This research work presents an Uncertainty-Aware Hybrid Learning Framework for accurate and safe State-of-Charge (SoC) prediction in lithium-ion batteries, particularly under dynamic and thermal-varying conditions during fast charging. The framework integrates multi-fidelity electrochemical observer outputs with a Tuned Multilayer Perceptron (MLP) Regressor, enabling real-time estimation of high-fidelity SoC while quantifying uncertainty using a Negative Log-Likelihood (NLL) loss. Case studies were conducted on a modified dataset simulating sensor faults and thermal distortions. The proposed model achieved a peak R² of 0.9921 and MSE of 0.000021 under clean conditions. Under thermal noise, the retrained MLP maintained strong generalization with R² as 0.9657, MSE as 0.00053, and Prediction Interval Coverage Probability (PICP) of 0.975 with a narrow MPIW of 0.200. Comparative evaluation against Random Forest, Gradient Boosting, and Linear Regression confirms the MLP's superior adaptability and robustness. This hybrid framework is particularly suitable for deployment in real-time Battery Management Systems (BMS) and offers a foundation for future thermal-aware predictive control in electric vehicles. The proposed framework also highlights the potential of integrating low-complexity observer models with learning algorithms for real-time applications in EV systems.