Hybrid OCSSA-VMD and Optimized Deep Learning Networks for Runoff Forecasting

To improve accuracy and address the non-linearity and non-stationarity in monthly runoff forecasting, this paper proposes a method that integrates intelligent optimization techniques with Deep Learning (DL) network. The Osprey-Cauchy-Sparrow Search Algorithm (OCSSA) is employed to fine-tune the parameters of Variational Mode Decomposition (VMD), which is utilized to break down the original runoff data into multiple Intrinsic Mode Functions (IMFs). A deep learning framework, combining Convolutional Neural Networks (CNN), Bidirectional Gated Recurrent Units (BiGRU), and Attention Mechanisms (AM), was developed, referred to as CNN-BiGRU-AM. Prior to feeding the IMFs into the network, the hyperparameters of models were optimized by using the Northern Goshawk Optimization (NGO) algorithm. Empirical analysis using 45 years of monthly runoff data from Hankou and Yichang Station in Hubei Province demonstrates that the proposed model achieves lower forecasting errors and higher accuracy compared to several commonly used models, offering a robust scientific foundation for runoff forecasting and water resource management.