Author(s): Cesar Ambrogi Ferreira Do Lago; Marcio Hofheinz Giacomoni; Eduardo Mario Mendiondo
Linked Author(s):
Keywords: Modelling and visualisation tools; Rapid flood models; Neural networks
Abstract: Hydrodynamic flood models have been widely used for floodplain predictions. However, the hydrodynamic models have substantial computational burdens. One-dimensional (1D) models are unable to simulate lateral flood propagation [1]. The two dimensional (2D) can predict lateral propagation more accurately and is recommended for flood analysis [2]. Yet, the 2D simulation time is significantly higher. These drawbacks can hinder the application of hydrodynamic models in large scale domains, in optimization problems and in real-time control of flood protection infrastructure. One alternative to hydrodynamic models are the low-complexity rapid flood models. These models apply simplified hydraulic concepts to decrease simulation time [3]. Neural Networks (NN) can be used to learn complex rules to improve flood predictions and while maintaining computational efficiency. However, NN generalization is still a challenge as it requires a large number of training samples. This study proposes a novel approach for predicting water surface elevation (WSE) using NN aiming for generalization to different locations. The methodology uses localized characteristics, instead of the whole target domain, to increase the number of training samples and generalization capabilities.
Year: 2021