A Data-Driven Framework for Charging Voltage–Capacity Curve Prediction Using Post-Discharge Voltage Relaxation

Accurate monitoring of battery capacity degradation underpins reliable battery operation. Existing methods either estimate the current-cycle maximum capacity via full charge–discharge curves or reconstruct the current full charging curve using local charging segments. Both approaches only characterize current cycle degradation: the former suffers from unavailable full curves in practice and insufficient state description, while the latter cannot predict the aging trend of subsequent cycles. To address the above limitations, this work adopts easily collected post-discharge relaxation voltage as the core input to predict the full voltage–capacity (V-Q) curve of the next charging cycle. An LSTM-CNN hybrid feature embedding module is integrated into a Transformer-based LCformer model to combine temporal dependency modeling and local feature extraction. Furthermore, a novel customized loss function incorporating voltage-interval weighting, a monotonicity constraint, and a smoothness penalty is proposed to improve the plausibility and accuracy of the predicted curves. The proposed method achieves the prediction of a complete charging curve containing rich battery state information for the next cycle, and can provide prospective data support for relevant research based on full charging curves, such as battery state-of-health estimation and lithium plating analysis. Validation on datasets from different laboratories demonstrates that the proposed method achieves a root-mean-square error of less than 0.03 Ah in V-Q curve prediction.