Find in Library

Flooding is the world’s most catastrophic natural event in terms of losses. The ability to forecast flood events is crucial for controlling the risk of flooding to society and the environment. Artificial neural networks (ANN) have been adopted in recent studies to provide fast flood inundation forecasts. In this paper, an existing ANN trained based on synthetic events was optimized in two directions: extending the training dataset with the use of hybrid dataset, and selection of the best training function based on six possible functions, namely conjugate gradient backpropagation with Fletcher–Reeves updates (CGF) with Polak–Ribiére updates (CGP) and Powell–Beale restarts (CGB), one-step secant back-propagation (OSS), resilient backpropagation (RP), and scaled conjugate gra-dient backpropagation (SCG). Four real flood events were used to validate the performance of the improved ANN over the existing one. The new training dataset reduced the model’s rooted mean square error (<i>RMSE</i>) by 10% for the testing dataset and 16% for the real events. The selection of the resilient backpropagation algorithm contributed to 15% lower <i>RMSE</i> for the testing dataset and up to 35% for the real events when compared with the other five training functions.