GWO-driven adaptive runoff forecasting via hybrid deep learning networks of VMD-RPV and GRU-KAN

ABSTRACT Accurate runoff prediction is an important part of optimizing water resource allocation, flood control, drought relief, and ecological protection. In recent years, artificial intelligence technologies have made significant progress in hydrological forecasting, but challenges still remain. To address these issues, this paper proposes a runoff prediction model based on modal decomposition and deep learning - Variational Mode Decomposition (VMD)-Runoff Process Vectorization (RPV)-gated recurrent unit (GRU)-Kolmogorov-Arnold Network (KAN). This model employs the VMD) method optimized by the Grey Wolf Optimization (GWO) algorithm, using minimum envelope entropy as the fitness function to optimize the decomposition parameters and extract the time-frequency features of runoff data. It incorporates the RPV method to enhance the model's ability to capture flood and extreme hydrological events. Finally, the GRU-KAN network is used for modeling, where KAN replaces the fully connected layers of GRU to improve feature extraction capability. Experiments conducted on the dataset from the Pengshan Hydrological Station in the Minjiang River Basin show that the proposed VMD-RPV-GRU-KAN model outperforms baseline models, achieving a root mean squared error of only 1.8436 m3/s and a mean absolute error of only 1.2244 m3/s. The research offers an efficient method for future runoff prediction and extreme hydrological event forecasting.