Deep Learning Hydrodynamic Forecasting for Flooded Region Assessment in Near-Real-Time (DL Hydro-FRAN)
Authors:
Francisco Haces-Garcia,
Natalya Maslennikova,
Craig L Glennie,
Hanadi S Rifai,
Vedhus Hoskere,
Nima Ekhtari
Abstract:
Hydrodynamic flood modeling improves hydrologic and hydraulic prediction of storm events. However, the computationally intensive numerical solutions required for high-resolution hydrodynamics have historically prevented their implementation in near-real-time flood forecasting. This study examines whether several Deep Neural Network (DNN) architectures are suitable for optimizing hydrodynamic flood…
▽ More
Hydrodynamic flood modeling improves hydrologic and hydraulic prediction of storm events. However, the computationally intensive numerical solutions required for high-resolution hydrodynamics have historically prevented their implementation in near-real-time flood forecasting. This study examines whether several Deep Neural Network (DNN) architectures are suitable for optimizing hydrodynamic flood models. Several pluvial flooding events were simulated in a low-relief high-resolution urban environment using a 2D HEC-RAS hydrodynamic model. These simulations were assembled into a training set for the DNNs, which were then used to forecast flooding depths and velocities. The DNNs' forecasts were compared to the hydrodynamic flood models, and showed good agreement, with a median RMSE of around 2 mm for cell flooding depths in the study area. The DNNs also improved forecast computation time significantly, with the DNNs providing forecasts between 34.2 and 72.4 times faster than conventional hydrodynamic models. The study area showed little change between HEC-RAS' Full Momentum Equations and Diffusion Equations, however, important numerical stability considerations were discovered that impact equation selection and DNN architecture configuration. Overall, the results from this study show that DNNs can greatly optimize hydrodynamic flood modeling, and enable near-real-time hydrodynamic flood forecasting.
△ Less
Submitted 5 July, 2023; v1 submitted 19 May, 2023;
originally announced May 2023.
