In marine dynamic positioning scenarios, accurate ship motion prediction is vital for warnings and motion compensation. Addressing the challenges posed by rapid and highly nonlinear ship motion changes in such scenarios, this study propose a hybrid model combining physical and data-driven approaches. The hybrid model employs a physical module to make initial predictions motion and corrects errors through a correction module. Building on this, a control force prediction module is introduced to update the parameters of the physical module, thereby improving prediction accuracy. The study utilizes dynamic positioning model test data for the training and testing of the hybrid model, and conducts an in-depth assessment of model performance with respect to prediction horizon and the number of training samples. Results show that for surge, sway, and yaw, the hybrid model improves 15-second prediction accuracy by 15%~57% over Long ShortTerm Memory (LSTM), with the most significant improvements observed at motion peaks and troughs. Additionally, the hybrid model exhibits stronger prediction stability when dealing with unevenly distributed datasets. In summary, the hybrid model offers higher accuracy and stability, holding promise for providing more accurate and stable warning information and motion compensation support for dynamic positioning operations in marine environments.