hide
Free keywords:
-
Abstract:
The two-dimensional hydrodynamic flood models are computationally expensive, especially in large-scale watersheds, which limits their application when several simulation runs are required or in large-scale catchments. More recently, the literature presented a significant increase in the use of Deep Learning as an alternative to hydrodynamic models because of their computational efficiency. However, studies show that Artificial Neural Networks (ANN) applications for flood prediction lack the ability to predict floods outside the training datasets satisfactorily. In this study, we used a conditional generative adversarial network (cGAN) aiming to generalize pluvial flood predictions (cGAN-Flood) in catchments not included in the training dataset. Our approach uses two generators to solve a rain-on-grid problem by first identifying wet cells and then calculating the water depths. cGAN-Flood distributes a target flood volume (vt) in a given catchment, which can be calculated via water balance from any hydrological simulations. Our approach was trained on ten and tested on five urban catchments. The testing areas have distinct characteristics and were compared to HEC-RAS outputs for different rainfall magnitudes and surface roughness. The averages of the root square mean error (RSME) and critical success index (CSI) were 0.12m and 70.1%, respectively. The successful water depths predictions in the testing areas show that cGAN-Flood can generalize. cGAN-Flood, combined with an SWMM model to compute vt, was 250 times faster than HEC-RAS. Due to its computational efficiency and accuracy, cGAN-Flood can be applied when fast simulations are necessary, and it can be a viable for flood forecasts in large-scale watersheds.
