Abstract: Real-time flood control scheduling needs to consider multiple constraints and comprehensive objectives, which has a high degree of complexity. Using the Three Gorges reservoir as a study case, this paper describes a new real-time reservoir flood control dispatching model based on deep learning. We simulate the dispatching process of the reservoir to generate training sample data; then using the sample data, we generate high-dimensional tensor input data, and extract high-dimensional data features through network parameter training to learn how to fit real-time reservoir flood control dispatch modes. This dispatching model, based on a real-time scheduling model equipped with a deep convolutional neural network, extracts the characteristic information of the gate opening in the training process. It uses reinforcement learning algorithms for optimizing the model parameters iteratively, and updating sample data successively through online learning to implement optimal scheduling decisions. In the case study of the reservoir’s real-time flood level control model and real-time discharge flow control model, the relative errors of simulated discharges are around 1.4% and 1.0%, respectively, against the site observed data. And the calculations show that our deep learning model has a good convergence, and it is applicable to the real-time flood control dispatching of reservoirs.

Key words: Reservoir flood control operation, simulation operation, deep neural network, reinforcement learning, decision