PG-DAT: Physics-Guided Domain-Adaptive Transformer for Battery State-of-Health Forecast

Nowadays, accurate battery state-of-health (SoH) forecasting is limited by data-driven models’ poor extrapolation beyond training ranges and high computational latency, especially when only scarce labelled degradation data are available for new deployments. To address these challenges, we propose the physics-guided domain-adaptive Transformer (PG-DAT) for battery SoH forecasting in smart-city energy systems. PG-DAT integrates electrochemical priors obtained from the open-source Python Battery Mathematical Modelling library (PyBaMM), captures local-to-global aging patterns through a Transformer combined with a bidirectional gated recurrent unit (Transformer-BiGRU) encoder, and learns domain-invariant features via gradient-reversal training. A least-recently-used (LRU) cache reduces simulation overhead to less than 1 ms per query, eliminating the latency bottleneck while preserving physical insight. Experiments on eight Li-ion cells (CS2-35/36/37/38 and CX234/36/37/38) spanning two chemistries, three nominal chargerate (C-rate) conditions, and multiple temperatures show that PG-DAT achieves an average root-mean-square error (RMSE) of 0.0519 and mean absolute error (MAE) of 0.0369—improving on current models by 32%. End-to-end inference remains below 10 ms on a single CPU core, satisfying online control requirements and offering a practical solution for battery analysis and management in smart-city applications.