Study on runoff prediction method under the influence of various runoff mechanisms

Runoff prediction is an important step for optimal operation of water resources, early warning of flood disasters and maintenance of ecological balance, and its accuracy improvement is of great significance for comprehensive watershed management. The traditional hydrological model is not applicable in areas with few data because it needs a lot of data support in the study area. Taking the Shitoukoumen Reservoir Basin as a typical area, this study compared three machine learning models including random forest (RF), support vector regression (SVR) and gradient lifting tree (GBDT), and two deep learning models including recurrent neural network (RNN) and long-term and short-term memory network (LSTM) under different runoff components in the same basin, and evaluated the performance differences of the models in terms of prediction accuracy through klingupta efficiency coefficient (KGE), determination coefficient (R2 ) and other indicators. It is found that the GBDT model is the best in the prediction accuracy, while the LSTM model is the worst of the five models in the prediction accuracy. The multi model comparison framework proposed in this study provides a theoretical support for the selection of runoff prediction models under different runoff components in the same basin, especially for the runoff prediction of similar small and medium-sized basins.