Data-Driven Performance Prediction of Lead-Carbon Batteries: Integrating Experimental Validation and Reduced-Order Model-Guided Neural Networks

Abstract

Accurate and efficient prediction of battery degradation is essential for optimizing energy storage system design and control. This study introduces a hybrid modeling framework that combines reduced-order modeling (ROM) insights with experimentally validated deep neural networks (DNNs) to predict degradation in lead-carbon (PbC) batteries. Using voltage-capacity profiles from 258 experimental charge/discharge cycles, we extract four physically meaningful input features-cycle number, capacity, charge voltage, and discharge voltage-to train a ROM-guided DNN surrogate. The model predicts two key health indicators: capacity retention (CapRet) and end-of-discharge voltage (EoDV). It generalizes well across five scenario types, including extrapolated conditions up to 700 cycles and varying voltage/capacity inputs. Predictions remain smooth and physically consistent, with validation yielding R-2 > 0.99 and low MSE. In terms of computational performance, the DNN achieves sub-second inference (similar to 0.02 s), offering over five orders of magnitude speedup compared to full COMSOL simulations (similar to 25 h), and similar to 1000x faster than ROM (similar to 22 s). This enables rapid scenario testing and real-time diagnostics. The proposed framework provides a scalable and interpretable solution for battery performance forecasting, well-suited for deployment in digital twins, battery management systems, and advanced energy storage design workflows.