Abstract: The accuracy and forecasting range of flood forecasting are crucial for effective reservoir flood warning and dispatching.The application of artificial intelligence models in flood forecasting can significantly enhance its accuracy.A K-means clustering analysis method was employed to scientifically partition the Pankou Reservoir basin, and then an Informer deep learning model was used for flood forecasting, and it was compared with the traditional LSTM model.Finally, based on the Informer model, four forecasting schemes were designed to analyze the impact of upstream reservoirs on the flood forecasting accuracy of the Pankou Reservoir.The results indicate that: ① the forecasting performance of the Informer model surpasses that of the LSTM model; ② for the optimized Informer model, the overall Nash coefficient for both the training and testing sets is 0.892, with a total flood volume error of 6.64% and a flood peak error of 7.69%.Both the average flood volume error and flood peak error meet Class A standards; ③ the actual test Nash coefficient values for the Informer model in 2023 and 2024 are 0.878 and 0.827, respectively, with both flood volume and flood peak error pass rates reaching 100%, meeting Class A requirements.The intelligent flood forecasting based on the Informer deep learning model not only enhances the prediction accuracy of flood volume and flood peak but also possesses strong practical application potential, providing a decision-making basis for reservoir flood warning and disaster prevention and mitigation.