A Novel Runoff Prediction Model Based on Support Vector Machine and Gate Recurrent unit with Secondary Mode Decomposition

Predicting runoff, one of the fundamental operations in hydrology, is crucial for directing the complete exploitation and use of local water resources. However, influenced by factors such as human activities and climate change, runoff displays typical nonlinear, non-stationary dynamic characteristics, which means it is challenging to achieve accurate runoff prediction in the research on water resources. In this research, we developed a hybrid model named CEEMDAN-FE-VMD-SVM-GRU for runoff prediction, which was built on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), fuzzy entropy (FE), variational mode decomposition (VMD), support vector machine (SVM), and gate recurrent unit (GRU). First, CEEMDAN was used to decompose the original daily runoff dataset into several intrinsic mode functions (IMF), followed by the introduction of FE to compute the complexity of each IMF component. The obtained FE calculation results greater than 0.4 were set as high-frequency sequences, and those lower than 0.4 as low-frequency sequences. Then, VMD was applied to perform the secondary decomposition of the high-frequency sequences, and SVM and GRU were trained to predict the primary and secondary decomposition parts, respectively. The results were finally obtained through linear combination. In this study, the daily runoff of the Minjiang River by this model was compared with those of other eight models. The findings demonstrate that the proposed model worked better than other models in a variety of evaluation metrics. In addition, this model showed better applicability in uncertainty interval estimation and flood prediction. Hence, this model proposed in this study has potential to be a preferred data-driven tool in hydrological prediction.