Key points are not available for this paper at this time.
Neural networks have become widely employed in streamflow forecasting due to their ability to capture complex hydrological processes and provide accurate predictions. In this study, we propose a framework for monthly runoff prediction using antecedent monthly runoff, water level, and precipitation. This framework integrates the discrete wavelet transform (DWT) for denoising, variational modal decomposition (VMD) for sub-sequence extraction, and gated recurrent unit (GRU) networks for modeling individual sub-sequences. Our findings demonstrate that the DWT–VMD–GRU model, utilizing runoff and rainfall time series as inputs, outperforms other models such as GRU, long short-term memory (LSTM), DWT–GRU, and DWT–LSTM, consistently exhibiting superior performance across various evaluation metrics. During the testing phase, the DWT–VMD–GRU model yielded RMSE, MAE, MAPE, NSE, and KGE values of 245.5 m3/s, 200.5 m3/s, 0.033, 0.997, and 0.978, respectively. Additionally, optimal sliding window durations for different input combinations typically range from 1 to 3 months, with the DWT–VMD–GRU model (using runoff and rainfall) achieving optimal performance with a one-month sliding window. The model’s superior accuracy enhances water resource management, flood control, and reservoir operation, supporting better-informed decisions and efficient resource allocation.
Yang et al. (Tue,) studied this question.