Physics-inspired window-wise learning framework for capacity degradation prediction using diverse real-world battery packs

Capacity degradation prediction of battery packs poses significant challenges to the safety, reliability and lifecycle management of electric vehicles (EVs). Battery degradation in real-world operation is physically consistent, with the current state governed by cumulative degradation effects embedded in historical usage trajectories. Physics-Informed Neural Networks (PINNs) provide a principled way to encode such physical degradation laws. However, their inherent point-wise learning style restricts them from capturing historical information, limiting their ability to predict capacity degradation. To address this issue, we propose a physics-guided window-wise learning framework based on PINNs to enable trajectory-aware capacity degradation prediction at the battery pack level. Specifically, we introduce an adaptive exponential weighting mechanism to emphasize degradation-relevant historical information in a window-wise representation for subsequent model learning. We then design a lightweight and non-recursive feature–channel attention encoder to model long-term trajectory information in a computational-efficiency manner. To validate the proposed framework, we collect a 1–3-year EV dataset including data from 50 vehicles with mileage ranging from 80,000 to 100,000 km, along with two public real-world EV datasets. These datasets cover diverse vehicles, battery systems, and operating strategies. Experimental results demonstrate the effectiveness of the proposed method, which achieves an average MAE of 1.912 Ah across the three datasets. The code will be made publicly available at https://github.com/Chan1w/PIW2L-EV .