Application of artificial neural network ensembles in probabilistic hydrological forecasting

Ensemble techniques are used in regression/classification tasks with considerable success. Due to the flexible geometry of artificial neural networks (ANNs), they have been recognized as suitable models for ensemble techniques. The application of an ensemble technique is divided into two steps. The first step is to create individual ensemble members, and the second step is the appropriate combination of outputs of the ensemble members to produce the most appropriate output. This paper deals with the techniques of both generation and combination of ANN ensembles. A new performance function is proposed for generating neural network ensembles. Also a probabilistic method based on the K-nearest neighbor regression is proposed to combine individual networks and to improve the accuracy and precision of hydrological forecasts. The proposed method is applied on the peak discharge forecasting of the floods of Red River in Canada as well as the seasonal streamflow forecasting of Zayandeh-rud River in Iran. The study analyses the advantages of the proposed methods in comparison with the conventional empirical methods such as conventional artificial neural networks, and K-nearest neighbor regression. The utility of the proposed method for forecasting hydrological variables with a conditional probability distribution is demonstrated. The results of this study show that the application of the ensemble ANNs through the proposed method can improve the probabilistic forecast skill for hydrological events.