Coupled temporal-frequency learning framework for lithium-ion battery capacity prediction

Abstract

Lithium‑ion battery capacity is critical for guaranteeing the system safety and stable operation. Since the true battery capacity cannot be directly measured by sensors, it can be estimated from available data. Nevertheless, battery aging represents a highly non‑stationary process, which consists of long‑term fading trends together with local fluctuations originating from capacity regeneration, working‑condition shifts and sensor‑introduced measurement disturbances. Therefore, single‑temporal models are insufficient to fully characterize battery degradation dynamics. To mitigate this limitation, this paper proposes a time‑frequency collaborative soft‑sensing framework for capacity estimation using health indicators extracted from charge‑discharge curves. A Multi‑scale Convolutional Attention Module (MCAM) is utilized to capture local degradation variations and long‑range aging dependencies from sensor sequences. Meanwhile, a Frequency‑Aware Mixture‑of‑Experts module (FA‑MoE) leverages learnable Gaussian spectral filtering and expert aggregation to build adaptive frequency‑domain representations for suppressing non‑stationary interferences. Adaptive fusion of temporal and spectral features enhances the characterization of complicated degradation trajectories under diverse aging behaviours. Validations on public NASA, CALCE, and TJU datasets reveal that the developed framework yields the lowest average RMSE, MAE and MAPE compared with competing approaches, demonstrating favourable indirect‑measurement stability under various battery aging conditions.