Physics-Informed Temporal Convolutional Networks for State-of-Health Battery Estimation of Real-world Electric Vehicles

Accurate State-of-Health (SOH) estimation for lithium-ion batteries in electric vehicles (EVs) under real-world driving conditions is challenged by fleet heterogeneity and the scarcity of labeled degradation data. Purely data-driven models are prone to shortcut learning, failing to generalize to batteries with nominal capacities unseen during training, i.e., the Out-of- Distribution (OOD) scenario. This paper proposes a two-stage hybrid framework based on Temporal Convolutional Networks (TCN). In the first Stage A, the model learns degradation dynamics from a simulator using capacity-agnostic features (Crate, normalized voltage, temperature, and SOC). In the second Stage B, Domain Adaptation on heterogeneous real-world data is performed via Partial Freezing and a Physics-Informed loss function (PINN loss) that penalizes non-physical temporal jumps in the SOH estimate. Experiments on heterogeneous real-world EV battery datasets show that the proposed approach mitigates capacity-related overfitting, enabling stable real-time onboard estimation and yielding temporally smoother SOH estimates than purely data-driven baselines, and that it generalizes to a battery pack of unseen nominal capacity in on-road tests.