Aiming at the problem that it is difficult for a single model to take into account the interpretability of physical mechanisms and the adaptability to complex environments in streamflow prediction, this study proposes a hybrid prediction model XAJ-TCN-GRU&XGBoost (XTGX), which combines hydrological mechanisms and deep learning. Firstly, the maximum mutual information coefficient (MIC) is used to screen key variables from hydrological and remote sensing satellite meteorological data. Then, the Xin'anjiang model based on hydrological mechanism and the TCN-GRU model capturing temporal features were used to generate streamflow predictions, the former analysing the physical processes and the latter mining the deep laws of the data. Finally, XGBoost is used to construct a nonlinear weight allocation mechanism, which dynamically integrates the complementary advantages of the two types of models to improve the prediction accuracy. The study shows that XTGX outperforms the single benchmark model in streamflow prediction and flood forecasting of Luoqing River and Chishui River, and the Nash efficiency coefficients (NSEs) are improved to 0.981 and 0.957, respectively. The results of the study are of great significance for optimizing water resources management and mitigating flood disasters in small and medium-sized rivers.