Multisource Data Fusion and Stacked Ensemble Modeling for Accurate Lithium-Ion Battery Capacity Estimation

Area of Science:

• Energy Storage Systems
• Machine Learning for Battery Health Management
• Predictive Maintenance in Electric Vehicles

Background:

• Accurate lithium-ion battery lifespan prediction is crucial for operational reliability and cost reduction in electric vehicles and smart grids.
• Existing methods struggle with the heterogeneity and variability of multisource battery data.
• Need for advanced frameworks to enhance battery health management and safety.

Purpose of the Study:

• To introduce a novel hybrid learning framework for accurate lithium-ion battery lifespan prediction.
• To address limitations of existing methods by integrating multisource data fusion and stacked ensemble modeling.
• To improve the scalability and interpretability of battery health management systems.

Main Methods:

• Developed a hybrid learning framework combining multisource data fusion with a stacked ensemble (SE) model.
• Utilized heterogeneous datasets from NASA, CALCE, and TRC, incorporating nickel cobalt aluminum (NCA) chemistries.
• Employed a variance-aware weighting mechanism and an SE model integrating ridge regression, LSTM, and XGBoost.

Main Results:

• Achieved superior predictive performance with MAE of 0.0058, RMSE of 0.0092, and R² of 0.9839.
• Demonstrated significant improvements over baseline models: 46.2% increase in R² and 83.2% reduction in RMSE.
• SHAP analysis identified differential discharge capacity (Qdlin) and temperature of measurement (Temp_m) as key aging indicators.

Conclusions:

• The proposed hybrid learning framework offers a scalable and interpretable solution for battery health management.
• The framework enhances operational reliability and reduces maintenance costs in diverse energy storage systems.
• This approach supports optimized maintenance and safety for lithium-ion batteries in critical applications.