Abstract: Fast and accurate prediction of the cross-section velocity field of a channel is of great significance for the design and maintenance of the open channel and the improvement of irrigation efficiency.However, it is more difficult to predict the velocity field in real time for channels with fast and large water level changes.Taking the main canal of People's Canal in Dujiangyan Irrigation area as an example, a new method for real-time velocity field prediction was proposed.Firstly, the Computational Fluid Dynamics(CFD)was used to simulate the cross section flow filed of the target open channel.Then, a machine learning model SaDE-ELM was established, and a fully connected three-layer and three-input and one-output neural network was constructed by SaDE-ELM model.The parameters of the hidden layer nodes were calculated by the differential evolution algorithm that adaptively selects the evolutionary strategy in the strategy library according to the characteristics of the flow field data, and the Moore-Penrose generalized inverse was used to calculate the weight of the output layer of the network.Finally, SaDE-ELM model was trained by CFD cross-section simulation data.After the training is completed, as long as the water level of the open channel and the position coordinates of any point in the section are input, the velocity of the point can be output, and the velocity field of the whole section of the open channel can be obtained.The application practice shows that the prediction results of the model conform to the general law of flow velocity distribution in open channels with high accuracy, which can be used as a reference for similar projects.