The 12th International Conference on Scour and Erosion (ICSE 2025)

Mountain rivers, such as the Jinsha River, play a crucial role in China’s large-scale hydropower development. When upstream dams trap mainstream sediment, sedimentation in downstream reservoirs becomes dominated by tributary inputs, introducing significant uncertainty and complicating sedimentation prediction. Reservoir operations further induce backwater effects, leading coarse particles to deposit upstream while finer sediments are transported into the main reservoir area, where they readily flocculate and settle to accelerate sediment accumulation. To address this issue, this study incorporated tributary sediment inputs and flocculation processes into a hydro-sediment numerical model. A machine learning-based framework was further developed for efficient and accurate sedimentation prediction in large mountain reservoirs. The proposed method reduced the total sedimentation error from 53.42% to 3.44%, and the maximum group-wise error from 90.88% to 13.46%, confirming the critical roles of tributary inputs and fine sediment flocculation. The machine learning models effectively captured the dynamics of sedimentation rates (R2 > 0.80, KGE > 0.85), with the Random Forest model showing the best overall performance and interpretability, outperforming models of XGBoost, Support Vector Regression, and Artificial Neural Network. This study provides an effective tool for accurate and rapid sedimentation assessment, and offer valuable insights for modeling sediment dynamics in mountainous basins and large reservoir systems.

Tributary sediment; Sedimentation modeling; Flocculation process; Machine learning; Mountain reservoirs; Jinsha River