University of Twente Student Theses
Rapid generation of probabilistic inundation forecasts by utilizing cloud computing and deep learning
Hop, F.J. (2023) Rapid generation of probabilistic inundation forecasts by utilizing cloud computing and deep learning.
|
PDF
30MB |
| Abstract: | Heavy rainfall events are occurring more frequently due to climate change, and can lead to inundation which poses risks for society. Inundation forecasts are crucial to alleviate these risks. The forecasts are used as a warning system and can help making decisions on the measures to take to reduce the inundation hazards. The most important variable for predicting inundation is the rainfall, however the uncertainty in rainfall forecasts is usually high. To include the uncertainty of the rainfall forecasts in the inundation forecasts, probabilistic inundation forecasts are made. These forecasts consider an ensemble of rainfall forecasts, and predict the inundation for each ensemble member. Based on the number of ensemble members where inundation occurs the probabilities of inundating can be calculated, resulting in a probabilistic inundation forecast. Conventional hydraulic models are too slow to generate probabilistic inundation forecasts, with total com- putation times generally exceeding one hour. This is because inundation for each of the ensemble members of the rainfall forecast has to be simulated. This study applies and analyses two methods of making probabilistic inundation forecasts fast enough to be used operationally. The study area is polder de Tol, located north-west of Utrecht in the Netherlands. The size of the area is about 12.5 km2, mostly consisting of rural landscape. For this area, a 1D2D hydraulic model is setup and calibrated such that inundation depths during pluvial flooding can be predicted. The first method to rapidly generate probabilistic inundation forecasts utilises cloud computing such that many simulations can be executed simultaneously. This significantly reduces the computational time required for making probabilistic inundation forecasts, since simulations for multiple ensemble members can be executed in parallel. In this study, cloud computing has proven to be a feasible solution, reducing the time required for probabilistic inundation forecasts to the computation time of a single hydraulic simulation. The costs of utilising cloud computing are dependent on the model complexity and the number simulations that are required. For the model used in this study, the costs of generating a probabilistic inundation forecast consisting of 50 ensemble members at a 10 meter resolution is 0.40 euros. The methodology used for this study can also be applied to other study areas, and is scalable when more simultaneous simulations are required. The second method is utilising deep learning by training a neural network to make inundation forecasts. The neural network trained for this study can predict inundation depths with a 10 meter resolution at 12 time steps. The network is trained on 1600 hydraulic simulations, and can accurately predict inundation depth progression over time. The accuracy of the neural network’s inundation forecasts is assessed by comparing the neural network predictions to hydraulic model forecasts for 200 different rainfall events. For each of these 200 events the mean absolute error is calculated, which is between 5.7 ∗ 10−5 and 0.07 cm, with an average of 0.01 cm. The network also performs very well in generating probabilistic inundation forecasts: For 99.6% of the predictions the neural network inundation probability is within 2% of the probability predicted by the hydraulic model. The neural network can generate a probabilistic inundation forecast within seconds. The neural network architecture used for this study is also expected to perform well in other study areas as long as the network is trained on data for that study area. Both methods of providing probabilistic inundation forecasts within an operational time frame have proven successful. Which method is preferred depends on the application. The neural network can make forecasts much faster, and at a negligible cost. However to do so a large number of training simulations have to be performed beforehand, and this has to be re-done when changes in the hydraulic model are required. Utilising cloud computing is more expensive and slower, but the hydraulic model used to make the inundation predictions can be changed whenever required. There are several suggestions for future research to improve the accuracy and applicability of the neural network. These suggestions include: studying the impact of additional input variables on the network, evaluating the network’s performance in areas with diverse topography and urban environments, and creating a generalised neural network that can be applied to any area without specific training. Also, it is recommended to evaluate different network architectures and to study how the amount of training data affects the accuracy of predictions. Future research opportunities to improve the accuracy and applicability of the neural network have been identified. These include: investigating the effects of additional input variables on the network, assessing the network’s performance in diverse topographical and urban environments, developing a generalizable neural network that can be applied without training for a specific area, evaluating various network architectures, and examining the impact of the quantity of training data on prediction accuracy. |
| Item Type: | Essay (Master) |
| Clients: | Hydrologic |
| Faculty: | ET: Engineering Technology |
| Programme: | Civil Engineering and Management MSc (60026) |
| Link to this item: | https://purl.utwente.nl/essays/94319 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page