A learnable graph spatiotemporal attention physics-informed network for lithium-ion battery state-of-health estimation

Abstract Accurate estimation of the state of health (SOH) is critical for the safe operation and effective management of lithium-ion batteries. However, data-driven methods face challenges in modeling the complex correlations among health factors (HFs) and often lack physical interpretability. In this paper, we propose a learnable graph (LG) spatiotemporal attention physics-informed network for accurate and reliable SOH estimation. Specifically, HFs are first extracted from the raw charging and discharging data of the batteries. A LG construction module is then employed to adaptively learn the topological structure among features and the importance of different time steps. Subsequently, the spatiotemporal attention network designed with a dual-path input architecture maps the inputs to SOH. One path deeply fuses spatiotemporal features across multiple time steps through the novel spatiotemporal feature extraction module, while the other path retains precise information from the current time step via a multi-layer perceptron, thereby achieving complementarity between long-range dependencies and immediate features. Finally, temporal attention weights are employed to compute the feature-specific partial derivatives across multiple time steps, and a degradation dynamics learning network is utilized to integrate the fitting capability of the data-driven approach with the physical constraints of battery degradation, enabling precise SOH estimation. Extensive experiments on the MIT and HUST datasets validate the effectiveness of our method and demonstrate its considerable potential for practical SOH estimation.