Three machine learning algorithms, namely LightGBM, XGBoost, and LSTM, were used as benchmark models. Water balance, reservoir water level, and discharge flow constraints were introduced into the loss function, and Kalman filtering was used for real-time correction. KP-LightGBM, KP-LSTM, and KP-XGBoost models were constructed, and a physical mechanism coupled machine learning method for extracting reservoir flood control ensemble dispatching rules was proposed. Based on the combined weighting of best and worst method, entropy weight method, and TOPSIS ranking method, different reservoir scheduling schemes are screened. The application results in the Shifengxi River Basin show that three scheduling scenarios have been selected, including balanced flood control, upstream-oriented flood control, and downstream-oriented flood control. Compared with LightGBM and XGBoost models, the LSTM model has the highest simulation accuracy, and the Nash efficiency coefficient