Monthly Runoff forecasting using A Climate‑driven Model Based on Two-stage Decomposition and Optimized Support Vector Regression

Accurate and reliable monthly runoff forecasting is crucial for water resource management, but the increasing non-stationarity of runoff series poses new challenges to the development of forecasting models. To overcome these problems, this study proposes a novel climate-driven hybrid model based on two-stage decomposition and.

optimization, called ICEEMDAN-SVMD-ESPSO-SVR. Firstly, the original runoff is decomposed into a set of components by improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), and high-frequency components identified through entropy analysis are further decomposed into multiple subcomponents by successive variational mode decomposition (SVMD). Secondly, the input variables are selected from the candidate set consisting of runoff, subcomponents and key climate factors by using correlation coefficient and mutual information, respectively. Thirdly, support vector regression (SVR) is employed to forecast each sub-component, and particle swarm optimization algorithm using eagle strategy (ESPSO) is used to select each model’s parameters. Finally, the forecasted values of all sub-components are aggregated as the final result. Monthly runoff data from three different hydrological stations in China's Poyang Lake Basin are employed to evaluate the performance of the proposed hybrid model and other comparable models. Results indicate that the proposed hybrid model outperforms single models and other combination models based on single-stage decomposition, and the correlation coefficient method is more suitable for input variable selection in climatedriven models, thus confirming the effectiveness of two-stage decomposition, the ESPSO algorithm, and input variable selection in enhancing modeling accuracy. Therefore, the proposed hybrid model is a feasible and promising new method for monthly runoff forecasting.