A hybrid MLP-Attention physics-informed neural network for robust SOH estimation of lithium-ion batteries

Accurate estimation of the State of Health (SOH) of lithium-ion batteries is crucial for ensuring operational safety, extending battery lifespan, and optimizing energy management. While data-driven models have shown high performance, they heavily depend on large volumes of high-quality labeled data, which are often costly and difficult to obtain in real-world scenarios. On the other hand, purely physics-based models tend to lack generalization due to simplifications in the modeling process and uncertainties in parameter estimation. In this paper, we propose a novel hybrid Physics-Informed Neural Network (PINN) framework that integrates a Multi-Layer Perceptron (MLP) with a scaled dot-product attention mechanism. The proposed model consists of two synergistic components: a data-driven MLP-Attention module for direct SOH prediction and a physics-informed residual module that incorporates battery degradation dynamics modeled by partial differential equations (PDEs). By leveraging both empirical data and domain-specific physical knowledge, the proposed framework enhances prediction accuracy and generalization, particularly in data-scarce conditions. Experimental results on the XJTU battery degradation dataset show that the proposed approach reduces the Root Mean Square Error (RMSE) by 26% and the standard deviation by 72% compared to the baseline PINN. Additionally, it outperforms conventional models such as CNNs and MLPs, achieving an 86% reduction in RMSE. These results demonstrate the effectiveness of the hybrid PINN in achieving reliable SOH estimation under limited data conditions.