The occurrence of catastrophic floods will increase the uncertainty of hydrological forecasting at downstream hydrological stations. In order to solve the problems of the unclear propagation law of catastrophic floods in the middle and lower reaches of the Yangtze River and the inadaptability of traditional forecasting methods, this paper uses the M-K trend test method to analyze the annual average flow and annual average water level of the Yichang and Hankou stations. For conventional floods and catastrophic floods, Random Forest (RF), Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and CNN-LSTM neural networks are used to simulate the water level/flow of Hankou station. The simulation results are analyzed by Nash–Sutcliffe Efficiency Coefficient (NSE), Kling–Gupta efficiency coefficient (KGE), Root Mean Square Error (RMSE), and Symmetric Mean Absolute Percentage Error (SMAPE). The results show that the annual average flow and annual average water level of Yichang station show a downward trend and the annual average water level of Hankou station shows an upward trend. By comparing the four indicators of NSE, KGE, RMSE, and SMAPE, the CNN-LSTM coupling model was determined to be the best fitting model, with NSE and KGE greater than 0.995 and RMSE and SMAPE less than 0.200. The proposed coupling model can provide technical support for flood control optimization, scheduling, emergency rescue, and scheduling impact analysis of the Three Gorges Power Station.