University of Twente Student Theses
Surrogate models: a solution for real-time inundation forecasting? Surrogate modelling for three case studies in the Netherlands
Janssen, Laura (2023) Surrogate models: a solution for real-time inundation forecasting? Surrogate modelling for three case studies in the Netherlands.
PDF
11MB |
Abstract: | The Netherlands is facing an increased risk of pluvial flooding, mainly due to the expected rise in the frequency and intensity of extreme rainfall events caused by climate change. Recent events have highlighted the importance of accurate flood forecasting to minimize damage. Hydrological inundation models are crucial in flood mitigation as they support water managers in an operational setting by predicting the size and timing of a flooding event, such that they can make informed decisions and implement appropriate measures (e.g. adjusting weir levels, alerting relevant parties). To ensure the operational usability of hydrological inundation models, they must meet certain criteria. Detailed hydrological inundation models provide accurate results but have long computational times, making them impractical for the operational context. Surrogate models have a shorter computational time by approximating the detailed hydrological inundation model. Surrogate models may offer a solution, but their quality and added value in an operational setting compared to existing hydrological models are yet to be determined. Waarschuwing voor Wateroverlast (W2O) is an example of a conceptual bucket model that is currently used in the operational setting. It combines a probabilistic rainfall forecast with geographical characteristics (e.g. land use, soil type) and the available soil storage to calculate the probability of flooding. This conceptual hydrological inundation model covers the entire area of the Netherlands and has a computational time of only one minute. It could therefore be, next to surrogate models, a useful alternative to predict inundation in an operational setting. In this study, the added value of surrogate models in an operational setting compared to the W2O model is researched. Three cases studies are selected to evaluate the hydrological inundation models’ compliance with the end-users’ requirements. These case studies include: the municipality of Amersfoort who would like to have an accurate forecast of the locations where inundation is expected within the city, the municipality of Tilburg who would like to have an accurate forecast of urban surface inundation, and Hoogheemraadschap van Rijnland who would like to know probability of flooding when the interaction between open waters (surface waters) and inundation on surface level is included. Semi-structured interviews with the end-users took place to specify the modelling goals, perspectives for action, relevant output variables, and the required accuracy. For all three case studies, the end-users are interested in the size and location of the flooding on a 2D map to alert relevant authorities (e.g. contractors). The waterboard is, next the location and size of the flooding, also interested in the water levels in the study area’s waterways, enabling proactive response to flood forecasts by changing weir levels and installing temporary pumping stations when necessary. For each case study, a surrogate model is created based on available data, models, and the end-user’s preference. For the municipality of Amersfoort, a Machine Learning (ML) model is used that predicts flood volume timeseries for manholes in the sewage system based on rainfall timeseries as input. This ML model meets most of the end-user’s requirements. However, the output variable (flood volume per manhole), is relatively difficult to interpret and is therefore not suitable for non-experts. The ML model created for the municipality of Tilburg predicts the maximum inundation depth on surface level using the rainfall timeseries as input. This ML model meets all requirements by the end-user, except of the required accuracy of the model. The critical success index, a performance indicator that measures the model’s ability to correctly predict flooding or not, is only 48% and therefore insufficient. Other studies in literature have shown that ML models are able to accurately reproduce the output of hydrological inundation models. This type of surrogate model has thus a high potential, but further research is necessary. For Hoogheemraadschap van Rijnland, a surrogate model is created based on a detailed hydrological inundation model by applying simplifications to the model schematisation. This model meets all end-user’s requirements but is at its limits regarding the maximum computational time (30 minutes). Also the quality of the W2O model is assessed using the requirements from the end-users. For both municipalities, the W2O model meets most of the requirements. Further research is necessary to determine if the W2O model also meets the accuracy requirement. For the waterboard, the W2O model suffices in the information provision on the 2D grid, but since the W2O model does not calculate the water level in waterways, it only partially meets the requirements for the required output variable. Overall, it can be concluded that, for municipalities, the W2O model provides sufficient information in an operational setting. It is advised to further research the accuracy of the W2O model to ensure that the W2O model also meets this requirement. The ML of the second case study has high potential in being an additional value to the W2O model. It is able to rapidly predict inundation depths on a high spatial resolution, but the accuracy of this ML model needs to be improved. Since generating the training data for the ML model is a computationally and memory expensive process, it is not advised to apply this type of surrogate model on larger model domains. Instead, ML models could be used as an addition to the W2O model, where the W2O model is used to get a general impression of the predicted inundation and the ML model is used additionally for the most vulnerable areas. For waterboards, the W2O model suffices in the information provision on the 2D grid, but additional models are needed to provide also the information regarding the water levels in waterways. The surrogate model created for the third case study could be used for this. Due to the long computational time, it is not advised to apply this surrogate to a larger model domain. Instead, this surrogate model could be used for the most vulnerable areas. Using the W2O model to get a general impression of the predicted inundation and combining this with a surrogate model for the most vulnerable areas, results in all information needed to make informed decisions on flood mitigation. |
Item Type: | Essay (Master) |
Faculty: | ET: Engineering Technology |
Subject: | 56 civil engineering |
Programme: | Civil Engineering and Management MSc (60026) |
Link to this item: | https://purl.utwente.nl/essays/96434 |
Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Repository Staff Only: item control page