Research on runoff interval prediction method based on deep learning ensemble modeling with hydrological factors

Abstract

Precise prediction of runoff is not only conducive to the prevention of floods and droughts but also to the rational use of water resources. Due to the frequency of weather extremes and the complexity of runoff variability, achieving accurate runoff predictions is challenging. This research develops a deep-learning ensemble model for interval prediction based on meteorological and hydrological factors. The model can be divided into four stages: feature extraction, decomposition, point prediction, and interval prediction. First, Pearson's correlation coefficient filters out key driving variables affecting runoff. Next, the original data are decomposed by variational modal decomposition (VMD) to intrinsic modal function (IMF); Then, each IMF is decomposed by complementary ensemble empirical modal decomposition (CEEMD) to capture more data details. Following, the runoff point prediction portion is realized by the attention mechanism fusion gated recurrent unit (AM-GRU). In this study, data from Dunhuang and Panjiazhuang stations, located in the upper and lower reaches of the Shule River in China, were used to validate and analyze the VMD-CEEMD-ISSA-AM-GRU (VCIAG) model. The results show that (1) the VCIAG model has the best fitting effect which the NSE values of Dunhuang and Panjiazhuang stations are 0.97 and 0.96, respectively. (2) In the multi-period prediction in advance, the highest prediction accuracy is achieved when the prediction period is 1 day and the accuracy of the prediction decreases gradually as the prediction period becomes longer. (3) In flood early warning, the VCIAG performs well at both stations, which suggests that the proposed model can take precautionary measures in advance before the floods come. (4) In terms of interval prediction, the VCIAG model has the narrowest prediction interval width and the highest prediction accuracy, which enhances the application value of the model.