Multi-state joint prediction algorithm for lithium battery packs based on data-driven and physical models

In the Industry 4.0 era, precise and efficient estimation of battery State of Health (SOH) and State of Charge (SOC) is essential for battery management systems. However, traditional SOC estimation methods often overlook the influence of battery health and internal physical characteristics, leading to inaccuracies in SOC estimation. To address this challenge, an innovative approach is proposed for the concurrent estimation of lithium battery pack states by seamlessly integrating data-driven and physical models. Initially, an automated aging matching model is established to capture inter-vehicle correlations and predict the target vehicle's capacity. SOC is then refined through a combination of the predicted capacity, the ampere-hour integration technique, and correction factors, ensuring high accuracy. Additionally, the physical model of lithium batteries is incorporated into the Channel-Spatial Attention-Gated Recurrent Unit (CSA-GRU) architecture, resulting in the considering physical information neural network (CPINN), specifically designed for SOC estimation. This approach effectively addresses the accuracy and interpretability issues commonly faced by data-driven methods in SOC estimation. Finally, the prediction accuracy, robustness, and real-time performance of the model are validated using electric bus battery data. Experimental results demonstrate that, compared to other models, the CPINN model offers superior accuracy, stability, and enriched information, along with enhanced interpretability.