Hybrid deep learning framework integrating CNNs, transformers, and adaptive optimization for accurate suspended sediment concentration prediction.

Reliable prediction of suspended sediment concentration is essential for reservoir operation, sustainable water management, and river ecosystem restoration. However, accurate forecasting remains challenging due to the strong nonlinearity and nonstationarity of sediment transport processes, as well as the complex interactions between hydrological variability and sediment supply dynamics.This study develops an integrated forecasting framework that combines signal decomposition, feature selection, deep learning, and optimization. In the proposed framework, variational mode decomposition is used to decompose suspended sediment concentration time series into multiple subcomponents, enabling the separation of temporal patterns at different scales. An improved sparrow search algorithm calibrates the decomposition parameters and improves component stability. Multigene genetic programming is then applied to select informative subcomponents and reduce redundancy. A hybrid convolution-transformer network is constructed to model both local fluctuation patterns and longer-range dependencies within the selected components. Finally, the improved sparrow search algorithm(ISSA) optimizes the weighting of individual model outputs to form an integrated adaptive predictor. The framework is evaluated using daily streamflow(Q) and suspended sediment concentration observations from 1977 to 1986 at the Tangnaihai Hydrological Station in the upper Yellow River Basin. The results show high predictive accuracy, achieving a Nash-Sutcliffe efficiency of 0.9535 on an independent test period with consistently low error metrics. Overall, the proposed framework provides a robust and generalizable approach for suspended sediment concentration forecasting under complex environmental conditions.