Abstract: River basin flood forecasting is of great significance for flood control and emergency response as well as water resource management. To explore the application of mechanism-deep learning coupling in river basin flood forecasting, this study takes the upper reaches of the Dongtiao River Basin as the research object, and constructs the Xin'anjiang model (XAJ), two deep learning models (LSTM, CNN-LSTM), and six mechanism-deep learning coupling models (PIML1/2/3-LSTM/CNN-LSTM). The application effects of these 9 models in flood forecasting are systematically discussed. The results show: (1) The overall stability of the XAJ model is relatively high, with an average NSE of 0.87 in the validation set, but there is a systematic underestimation in the high-flow section, and the absolute values of the RE and REP indicators are all greater than 15%; (2) The deep learning models (LSTM, CNN-LSTM) perform better than the XAJ model in the high-flow section, but have a slightly lower NSE, which is 5.7% and 10.3% lower than that of the XAJ model respectively; (3) Among the coupling models, only the PIML1 type model that introduces the output flow (QXAJ) of the XAJ has the best accuracy, among which, the PIML1-LSTM model's RE and REP indicators are 71.9% and 33.2% higher than those of the XAJ model respectively; while the PIML2 type model that introduces soil moisture content (QSXW) has a significant performance decline, indicating that some intermediate variables in the physical model may inhibit the learning effect of machine learning; (4) The coupling models sacrifice a small amount of overall fitting degree to achieve a significant improvement in key flood characteristic indicators, which has important practical value for flood control decisions. The research suggests that the PIML1 type architecture should be prioritized, while maintaining the physical interpretability to optimize the forecast accuracy. This study reveals the influence rules of variable selection in mechanism models on the performance of coupling models, providing a theoretical basis for optimizing hydrological physical-data-driven coupling modeling.