Abstract: A teleconnection relationship exists between watershed runoff and large-scale climate indexes. For medium- and long-term runoff prediction, a major difficulty is how to pick out those that are strongly correlated with runoff from various factors such as hydrology, meteorology, atmospheric circulation, and ocean current. This study applies a random forest model based on Bayesian optimization (sequential model-based optimization for general algorithm configuration) to selecting runoff predictors from the set of high-dimensional hydrometeorological and climatic factors according to their importance scores, and constructs a general regression neural network, an extreme learning machine, and a support vector regression runoff prediction models. The method is applied to runoff predictions for the Jinsha River. Compared with those of the factor selection model based on correlation coefficients, our new prediction model using the random forest for factor selection improves the generalization capability. Meanwhile, adding appropriate teleconnection climatic factors to the prediction model inputs can help improve accuracy of monthly runoff prediction and provide physical basis for the model.

Key words: monthly runoff prediction, random forest, Bayesian optimization, teleconnection, Jinsha River