Deriving reservoir operation rule based on Bayesian deep learning method considering multiple uncertainties

Abstract

Reservoir operation rules play a key role in real-time reservoir operation; the main factors affecting operation decisions are the current reservoir status and the future inflows. However, the future reservoir inflows are stochastic and always contain uncertainty. To study the influence of inflow uncertainty on reservoir operation rules, this paper proposes a Bayesian Deep learning method that considers both model parameter uncertainty and inflow uncertainty. In the model, the Monte Carlo integration is used to convert the complex integrals of inflow probability into a summation form. Variational inference is employed to obtain the posterior distribution of model parameters. The proposed method is applied to a real-world application at Three Gorges Project on the Yangtze River. The uncertainty estimation results show that the influence of inflow uncertainty on reservoir operation rule is greater than model parameter uncertainty, and the decision of reservoir operation is more sensitive to the reservoir inflow during dry season than other seasons. The experimental results demonstrate that the proposed Bayesian deep learning performs better than the comparison method in term of hydropower generation and the root mean square errors. Moreover, the proposed method is more robust than the comparison method when considering the inflow uncertainty.