Optimisation of machine learning sediment forecasting models for a typical watershed in the middle reaches of the Yellow River

ABSTRACT

The middle reaches of the Yellow River exhibit severe soil erosion and complex water–sediment dynamics, requiring accurate sediment concentration forecasting for flood mitigation and ecological management. This study evaluates the applicability of Long Short-Term Memory (LSTM), Artificial Neural Networks (ANN), and Transformer models for short-term sediment concentration forecasting in the Jing River and Beiluo River basins. Precipitation, discharge, and sediment concentration data from 1979 to 2024 were used to construct models for 168 representative flood events with 1h, 3h, and 6h lead times. An adaptive empirical quantile method was employed to assess model performance across different sediment concentration levels. Results demonstrate that forecasting accuracy decreases with increasing lead time. LSTM achieved the best overall performance, with NSE values of 0.975 and 0.759 for 1- and 6-h forecasts at Qingyang Station, respectively. ANN showed intermediate performance, while Transformer exhibited weaker capability in capturing sediment peaks. The analysis revealed systematic prediction biases, characterized by underestimation of high sediment concentrations and overestimation of low concentrations, highlighting the challenges of purely data-driven models in maintaining physical consistency during extreme sediment transport processes. These findings provide valuable guidance for selecting deep learning models in high sediment-load regions.