Abstract: To improve the accuracy in medium- and long-term prediction of reservoir runoff, this paper develops a combined prediction model integrating variational mode decomposition, phase space reconstruction, and a deep gated network. For a series of historical runoff data, this model first generates its multiple mode components using variational mode decomposition, and uses phase space reconstruction to reconstruct the components into a high-dimensional feature space to generate inputs for deep learning. Then, detailed characteristics of the series are obtained and predicted using a deep gated network, and the reconstruction is completed by superimposing the predicted values of the mode components. In a case study of the Baishan reservoir, this combined prediction model is compared with typical single prediction models and other combined prediction models. The results show that our method can effectively decompose non-stationary runoff series, learn the intrinsic hydrological patterns, and offer the smallest prediction error, the highest goodness of fit, and more reasonable error distributions over the whole test set, thus helping the planning and management of reservoir water resources.

Key words: variational mode decomposition, phase space reconstruction, deep gated network, medium and long-term runoff prediction, evaluation index