Uncertainty Aware Hybrid Learning Framework for Fast and Safe Charging of Lithium-Ion Batteries Using Multi-Fidelity Observers

Abstract

The paper presents a framework for accurately estimating state of charge in lithium-ion batteries during fast charging. The approach integrates multi-fidelity electrochemical observer outputs with a Tuned Multilayer Perceptron (MLP) Regressor to enable real-time predictions while quantifying uncertainty. Testing on modified datasets showed peak performance with an R² of 0.9921 under optimal conditions and maintained strong results (R² of 0.9657) under thermal stress. The authors conclude this hybrid method is particularly suitable for deployment in real-time Battery Management Systems (BMS) in electric vehicles.

Key Contributions

• Novel hybrid architecture combining multi-fidelity electrochemical observers with a tuned MLP regressor for SOC estimation
• Explicit uncertainty quantification embedded in the prediction pipeline — enabling safe charging decisions under model uncertainty
• Demonstrated real-time feasibility suitable for onboard BMS deployment in EVs
• Robust performance under thermal stress conditions, maintaining R² of 0.9657 even under elevated temperature scenarios
• Multi-fidelity observer design that fuses low-fidelity (simplified equivalent circuit) and high-fidelity (electrochemical) model outputs

Methodology

The framework fuses outputs from multi-fidelity electrochemical observers (spanning simplified equivalent circuit models to full physics-based electrochemical models) as inputs to a tuned Multilayer Perceptron Regressor. The MLP is hyperparameter-optimized for SOC estimation accuracy. Uncertainty quantification is incorporated to flag prediction confidence, enabling conservative charging control when uncertainty is high. Evaluation conducted on modified battery datasets under normal and thermal stress conditions.

Results

• R² of 0.9921 under optimal operating conditions
• R² of 0.9657 under thermal stress conditions
• Framework demonstrated real-time feasibility for BMS deployment
• Uncertainty quantification enables adaptive charging rate control, improving safety margins during fast charging
• Hybrid multi-fidelity approach outperforms single-fidelity model baselines across test conditions

Limitations

• Evaluated on modified datasets; broader validation on diverse real-world battery chemistries and aging states not fully demonstrated
• Computational overhead of multi-fidelity observer pipeline on embedded BMS hardware not quantified in detail
• Long-term performance degradation of the MLP regressor as battery ages not addressed
• Generalizability to batteries outside the training dataset chemistry requires further investigation

Source: Uncertainty Aware Hybrid Learning Framework for Fast and Safe Charging of Lithium-Ion Batteries Using Multi-Fidelity Observers by Parimala et al., Scientific Reports