University of Twente Student Theses


Dike breach flood prediction of an LSTM compared to the HAND.FLOW model for real-time flood forecasting

Besseling, L.S. (2022) Dike breach flood prediction of an LSTM compared to the HAND.FLOW model for real-time flood forecasting.

[img] PDF
Abstract:The most common method to model flood dynamics is using two-dimensional depth-averaged (2DH) hydrodynamic models (Chu et al., 2020). However, these models generally have long computation times of many hours or even days. As a result, they cannot be used for scenario analysis in a real-time flood forecasting system after a warning for an incoming discharge is issued (Teng et al., 2017). The aim of this study is to identify which surrogate model is the most promising model for real-time flood forecasting in case of a dike breach: a new conceptual HAND.FLOW model or a data-driven neural network. The neural network that was developed in this study is a Long Short Term Memory (LSTM) neural network, since it has been found suitable for predicting time series due to its ability to store information and to learn long-term dependencies in data (Le et al., 2019). In this study, data from 73 flood events modelled in a 1D2D-hydrodynamic model developed by Bomers (2021) was used to train and assess the LSTM. The outflow hydrograph of the dike breach functioned as the input, and the water depth in the hinterland was predicted per time step on every grid cell of the study area. The model architecture and hyperparameters such as dropout, number of neurons, activation function and learning rate were optimized using Bayesian optimization for the lowest value of the error function on water depth (Mean Absolute Error, MAE). The original HAND model is a conceptual model that only requires the Digital Elevation Model (DEM) of the study area to be set up. It takes the river water level as its input and floods all cells along the river with a Height Above Nearest Drainage (HAND) value lower than this water level. A dike breach, on the other hand, is a point source of a flood. To model the flood propagation from the breach into the hinterland, the new HAND.FLOW model was created with a number of adaptations: a pathfinding algorithm for finding the steepest downstream path from the dike breach into the hinterland, a distance limit relationship limiting the pathfinding algorithm per time step to simulate flood propagation behaviour, and a volume component allowing the model to take the outflow hydrograph as input instead of the river water level. Both models were tested on 15 flood events that were excluded from the LSTM training procedures. The LSTM performance was very accurate with a MAE of just 0.045 meters on an average water depth of 1.49 meters: an error of just 3% compared to HEC-RAS. The NSE values were close to 0.99 on nearly all grid cells in the study area, and the CSI metric for comparing the inundation areas was on average 0.94. The HAND.FLOW model was less similar to the water depths of HEC-RAS, due to a terrain feature not modelled in HEC-RAS. In a single corrected simulation, the MAE was 0.21 meters (error of 15%), the NSE was around 0.8 for large parts of the study area and the CSI averaged around 0.7. After a change in the hinterland, the HAND.FLOW model also correctly predicted the new flood pattern. All in all, the data gathering for the LSTM requires a lot time (800 hours for Bomers (2021)). It has to be retrained for a change in the hinterland or for another breach location, so it is not flexible. After the training procedure, however, it can predict the flood event near instantly and very accurately. The HAND.FLOW model requires a much shorter set-up time of around 30 minutes, so it is very flexible for changes in the hinterland, simulating other breach locations, or adapting spatial/temporal resolutions. The simulation time was 30 minutes on a detailed resolution of 10x10 meters, and only 1.5 minutes on the 150x150 meter resolution used by HEC-RAS and the LSTM. Therefore, the HAND.FLOW model offers in a relatively short simulation time a reasonable insight in how a dike breach flood will propagate in the hinterland, and could be the suitable and flexible model needed for a real-time flood forecasting system if it is further developed.
Item Type:Essay (Master)
Faculty:ET: Engineering Technology
Programme:Civil Engineering and Management MSc (60026)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page