Transferable Electric Vehicle Battery Capacity Estimation From Real-World Charging Data Using Spectral Learning

Reliable estimation of electric vehicle battery capacity is critical for range prediction, maintenance planning, and second-life reuse. While vehicle manufacturers employ proprietary algorithms within the battery management system, these methods are closed, model-specific, and inaccessible to fleet operators, charging networks, and regulators. This gap creates a clear need for transferable, data-driven approaches that use universally available signals, such as charging traces to deliver transparent, fleet-scale capacity estimates. We present a spectral learning framework based on Fourier neural operators that map per-cycle multivariate charging data to capacity estimates. The model incorporates both per-cycle predictions and sequence-level constraints to ensure smoothness and temporal consistency. Evaluated on a real-world dataset comprising 83 EVs and over 15 800 charging cycles, the method achieves a relative root mean square error (RMSE) of 1.9%, reducing error by nearly 40% compared to strong baselines. Importantly, we show that proposed approach generalizes across vehicle manufacturers via light fine-tuning with state-of-the-art accuracy recovered on independent subsets. These results demonstrate that frequency-domain operator learning provides a compact, transferable, and interpretable tool for reliable battery health estimation beyond proprietary systems, enabling fleet-scale monitoring and supporting the wider adoption of electric mobility.