Ensemble Flood Forecasting By Artificial Neural Network

In recent years, the frequency and intensity of extreme rainfall events have significantly increased due to climate change. Therefore, there is a need for constructing reliable flood prediction models. The applications of artificial neural networks, such as flood prediction/warning have been well developed and recognized. The main content of this study is to integrate the concept of ensemble concept into artificial neural networks. The ensemble neural networks are built for flood predictions by generating the ensemble members through initialization, resampling and network structure methods and combining the model outputs through the arithmetic average, Bayesian model average (BMA) and stacking average in order to take the model uncertainties into consideration. Then, the study investigates various ensemble strategies on two study sites where the watershed size and river system are different. In this case, it can help to directly realize if the ensemble flood predictions are sensitive to hydrologic and physiographic factors, and the applicability and availability of the model can be easily evaluated.The main achievements and innovations of this study are as follows:(1) The ensemble neural network models greatly improved the accuracy of flood prediction as compared to single neural network models in both watersheds. Moreover, the results obtained from different ensemble strategies are similar, indicating that the ensemble flood predictions are not sensitive to the hydrologic and physiographic factors. (2) Among various ensemble strategies, the combination of initialization and arithmetic average had simpler structure and less computational time, and its improvements in Longquan river basin and Jinhua river basin are about 16%-32% and 10%-28%, respectively, as compared to the single model. Whereas, the accuracy obtained from the combination of boosting and BMA increased 22%-35% and 12%-30% in Longquan and Jinhua river basins, respectively, and the overall performance of this combination is better than other strategies. (3) The study constructed reliable flood prediction models in both Longquan and Jinhua river basins and the intervals of ensemble flood predictions may reflect the possible maximum flood which can be the reference for flood prevention.