Abstract: Making full use of modern technological means to improve the accuracy of runoff forecasting plays an important guiding role in basin flood and drought disaster defense and joint scheduling of reservoir groups. However, existing deep learning models have problems such as lack of model transparency and poor physical interpretability. To address the above problems, in this study, a conceptual hydrological model EXP-Hydro is embedded into the P-RNN layer of recurrent neural network, and a deep learning hybrid model Hybrid-DL coupled with physical mechanism is modeled. The hybrid model adopts a differential framework to realize the deep bidirectional fusion of conceptual model and neural network, which is able to train the parameters of conceptual model and neural network at the same time. And an application study is carried out in the upper reaches of Qingjiang River as an example. The results show that compared with RNN, EXP-Hydro, BP and SVM models, the Nash efficiency coefficient（NSE） of the Hybrid-DL model increases by 6.08%, 21.01%, 37.09% and 73.92%, the rootmean-square error（RMSE） decreases by 10.82%、33.73%、54.70% and 95.57%, the KGE efficiency coefficient increases by 4.78%、12.68%、26.79% and 55.74%, and the peak error TPE decreases by 4.96%、13.12%、252.84% and 297.81%. The Hybrid-DL model has good robustness and adaptability, and can provide a reliable theoretical tool for runoff forecasting in the upper reaches of Qingjiang River and even in other basins.