Lithium-ion battery capacity prediction is a critical task in battery health management, yet existing methods still face three core challenges: inaccurate modeling of nonlinear degradation trends, insufficient utilization of multi-dimensional sensing features, and prediction deviations caused by capacity regeneration phenomena. To address these issues, this study proposes TFDM-CR, an innovative approach that combines time-frequency features with a diffusion model to model and predict capacity regeneration in lithium-ion batteries. First, the long-term capacity degradation trend is modeled through the forward noise-adding and reverse denoising processes of the diffusion model. Second, key features are selected using the Dynamic Time Warping algorithm, and a joint representation is constructed by combining time-frequency domain analysis. Finally, a dynamic compensation mechanism for capacity regeneration is designed, which corrects prediction deviations through local extreme detection and feature matching. Experimental results demonstrate the outstanding performance of the proposed method on both NASA and CALCE datasets. For 32-step predictions on the NASA dataset, the MSE is 0.00029 and MAPE is 0.00721; on the CALCE dataset, the MSE is 0.00015 and MAPE is 0.02785. As the prediction horizon extends to 48/64 steps, all metrics maintain stable growth trends, validating the model’s effectiveness in capturing battery degradation patterns. This research provides a high-precision solution for lithium-ion battery capacity prediction, particularly suited for real-world scenarios with complex degradation patterns and sudden capacity fluctuations. • Diffusion model predicts battery capacity via denoising degradation for precise forecasting. • DTW selects time-frequency features, boosting prediction via complementary feature fusion. • Adaptive regeneration compensation corrects distortion via pattern-matching, boosting reliability.
Shi et al. (Sun,) studied this question.