Quadratic decomposition and WOA-optimized CNN-ResNet-BiGRU-Attention: A novel hybrid framework for hydrological runoff forecasting

Accurate and reliable monthly runoff prediction is a crucial component of scientific water resources management. Due to the impacts of human activities and climate change, runoff series exhibit non-stationarity, which typically deteriorates the performance of most existing runoff prediction models. To address this challenge, this paper proposes a novel hydrological runoff forecasting model based on double decomposition and deep learning. The model first decomposes the runoff series using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). Then, the K-Means algorithm classifies the subsequences according to the sample entropy criterion and frequency characteristics. The Singular Spectrum Analysis-Variational Mode Decomposition (SSA-VMD) model is introduced to conduct a secondary decomposition of the high-frequency components. Subsequently, a CNN-ResNet-BiGRU model integrated with the Attention mechanism is optimized by the Whale Optimization Algorithm (WOA) to simulate and predict the runoff series. The results of the monthly runoff series at the Longmen Hydrological Station in the Yellow River Basin demonstrate that the Nash-Sutcliffe Efficiency (NSE) coefficient of the proposed model reaches 0.9743. Compared with the single VMD decomposition model, the Mean Squared Error (MSE) of the double decomposition model is reduced by 80.86%, and the NSE is increased by 18.27%. Moreover, simulation predictions at three hydrological stations in different regions of the middle and lower reaches of the Yellow River Basin further verify the model's potential applicability, with all NSE values exceeding 0.95. This study provides an effective approach for monthly runoff prediction in non-stationary environments, which is of great significance for flood control and scientific water resources management.