Predict seasonal low flows in the upper Yangtze River using random forests model

Predictor selection and model construction are two key issues in long-term streamflow forecasting. This study introduces a random forests model for selecting predictor set from measured streamflows and 74 hydro-climatic indices of the period from January to October provided by China National Climate Center, and predicts seasonal low flow in the upper Yangtze for the period from November to next May. The results show that: 1) as predicting lead-time increases, streamflow auto-correlation becomes weaker and hydro-climatic teleconnection gets stronger than the auto-correlation; 2) the predicted streamflows at the Pingshan and Cuntan gauges show a strong linear relationship with the observations and their relative errors are less than 20%; 3) analysis of the prediction uncertainty indicates that the random forests model can also be applied to probabilistic prediction of streamflow.