To address the problem that traditional fixed-parameter models cannot adapt to battery aging processes, data-driven approaches were deeply integrated with physical model mechanisms, and a state-of-health (SOH)-based dynamic modeling method for lithium-ion batteries was proposed. Convolutional long short-term memory networks were employed to extract multi-scale features from battery time-series data, achieving high-precision SOH prediction. A second-order RC equivalent circuit model was constructed, and a forgetting factor recursive least squares method was adopted for online parameter identification. Random forest regression and Kalman filtering were introduced to realize dynamic updating of model parameters according to aging states. Experimental results show that the proposed method achieves higher prediction accuracy and a more stable error distribution than the traditional fixed-parameter model across all battery aging stages. In the late aging stage, the proposed method improves the prediction performance of the dynamic model by approximately 80%, with a median error of 0.025 V, demonstrating significant advantages over the traditional model.