To explore the fusion methods of physical mechanism models and data-driven models, the implementation pathway of existing physics-guided fusion driving methods was analyzed. The research status of physics-guided fusion driving methods based on theory-guided data science (TGDS) in different fields was classified and summarized, and new classification methods, such as the physics-guided feedback fusion driving method and the physics-guided coding fusion driving method, were proposed. According to the characteristics of the hydrological modeling field, the challenges of the physics-guided fusion driving method in hydrological model construction were elaborated in detail. And the future research directions have been prospectively discussed. It was believed that the physics-guided fusion driving method based on TGDS could enhance the physical consistency of prediction results, reduce the accumulation of prediction errors, and improve the interpretability of the model, providing a feasible path for improving flood forecasting-oriented hydrological models. It could not only improve the low prediction accuracy caused by the generalization process of the mechanism model but also effectively improve the interpretability problem caused by the excessive dependence of data-driven models on samples. However, this method faces the problems of limited computing power, inaccurate extraction of multi-source data features, and inflexible parameter adjustment in its application. Therefore, in future research, differentiable modelling (DM) can be combined, or large models combined with domain knowledge graphs can be used to further explore the modeling of flood time series prediction, so as to better meet the needs of flood forecasting in complex environments.