Abstract: The mechanism by which the transfer error of the mechanism module in the Physics-Informed Machine Learning (PIML) flood forecasting model affects the forecast results is not yet clear. To investigate this issue, a dynamic parameter optimization method based on similar precipitation screening (S method) was proposed, and four coupled scenarios (XAJ, S-XAJ, XAJ-DL, S-XAJ-DL) were constructed using the Xin'anjiang Model (XAJ) and the Convolutional Neural Network Long Short Term Memory Self Attention Joint Model (CNN-LSTM Attention). We conducted flood forecasting experiments in the upper reaches of the Huai River Basin using these four coupling scenarios, systematically studied the error transfer mechanism of the mechanism model in PIML, and proposed the applicable scenarios of the coupling model. The results showed that: ① By dynamically optimizing the mechanism model parameters through similar precipitation screening, the Nash efficiency coefficient of the XAJ Model validation set increased from 0.76 to 0.80 (relative gain of 5.3%); ②When the error of the mechanism model is high (XAJ, validation set NSE value of 0.76), direct coupling will amplify the system error (XAJ-DL, validation set NSE value of 0.69), and the overestimation of flow rate in low water scenarios can reach 23%; ③The dynamic optimization mechanism model parameters can significantly reduce the model system error (S-XAJ, validation set NSE value of 0.80), block the transmission of abnormal error modes, and achieve the best comprehensive performance of the coupled S-XAJ-DL model (validation set NSE value of 0.82). This study takes the commonly used XAJ Model and representative deep learning model in flood forecasting as case studies to explore the optimal coupling scenario of mechanism data dual drive in PIML flood forecasting model, providing a theoretical basis for suppressing the flood forecasting error of the coupling model.