Abstract: The intensity of geomagnetic disturbances during geomagnetic storms is characterized by SYM-H, and accurate prediction of SYM-H is crucial for space weather warnings. The composite model is the latest short-term SYM-H prediction model integrating physical principles and neural network technology. It is extended to predict SYM-H throughout the entire magnetic storm process via an iterative strategy. The model establishes a SYM-H time evolution equation based on the total energy balance equation of the ring current, and optimizes key model parameters using neural networks, balancing interpretability and the ability to handle complex scenarios. Trained with solar wind and SYM-H observation data, the study compares the prediction performance of SYM-H during magnetic storms under different iterative time steps. The optimal performance is achieved at an iterative step of 60 minutes, with a root mean square error (RMSE) of 16.3 nT and a determination coefficient R² of 0.767. Prediction errors are mainly concentrated in the sudden storm commencement (SSC), main phase, and early recovery phase, and errors for intense magnetic storms are significantly higher than those for weak ones. This study verifies the feasibility of the composite model in full magnetic storm prediction, providing an alternative for space weather forecasting.