A monthly runoff prediction model based on ICEEMD-L-SHADE-SRU

Medium- and long-term runoff time series are characterized by strong nonlinearity and non-stationarity, which makes it difficult to predict accurately in the model. Taking the runoff data from three hydrological stations in the lower reaches of the Yellow River as an example, an improved complementary ensemble empirical mode decomposition (ICEEMD) is proposed. The runoff prediction model is established by combining the success-history adaptive differential evolution algorithm for linear population size reduction (L-SHADE) and simple recurrent unit (SRU). ICEEMD reconstructs the components of CEEMD with similar frequency and amplitude to obtain the frequency terms, which is used as the training data for the SRU. L-SHADE is employed to complete the parameter optimization of the SRU. The results showed that ICEEMD-L-SHADE-SRU achieved the best performance in runoff prediction, showing distinct improvements when compared to tested models in terms of both NSE (0.91–0.93) and QR (72%–74%). Considering the influence of extreme weather, the atmospheric circulation factors that have a greater impact on runoff are screened out and fused into the frequency term for prediction, and the model performs well. The innovation of this study lies in the distinct functionality of each component: ICEEMD effectively decomposes complex runoff data, L-SHADE optimizes model parameters, and SRU provides accurate predictions, collectively enhancing the reliability of runoff prediction. It can be used for medium- and long-term monthly runoff prediction.