Comparing ensembles of decision trees and neural networks for one-day-ahead streamflow prediction

Ensemble learning methods have received remarkable attention in the recent years and led to considerable advancement in the performance of the regression and classification problems. Bagging and boosting are among the most popular ensemble learning techniques proposed to reduce the prediction error of learning machines. In this study, bagging and gradient boosting algorithms are incorporated into the model creation process for daily streamflow prediction. This paper compares two tree-based ensembles (bagged regression trees BRT & gradient boosted regression trees GBRT) and two artificial neural networks ensembles (bagged artificial neural networks BANN & gradient boosted artificial neural networks GBANN). Proposed ensembles are benchmarked to a conventional ANN (multilayer perceptron MLP). Coefficient of determination, mean absolute error and the root mean squared error measures are used for prediction performance evaluation. The results obtained in this study indicate that ensemble learning models yield better prediction accuracy than a conventional ANN model. Moreover, ANN ensembles are superior to tree-based ensembles.