University of Twente Student Theses

Login

Exploring the use of surrogate models to reconstruct historic discharges : artificial neural networks for the reconstruction of the 1809 flood event of the Rhine river delta

Fredrix, Yorick (2018) Exploring the use of surrogate models to reconstruct historic discharges : artificial neural networks for the reconstruction of the 1809 flood event of the Rhine river delta.

[img] PDF
3MB
Abstract:Currently over the last 120 years discharges are measured in the Dutch river delta. These values are used to consider the possible discharge of once every 1.000 -10.000 year. This requires extrapolation of these 120 data points. By reconstructing historic floods, additional extreme values can be added to this dataset. Using these cases will reduce the uncertainty in what will occur once every 1.000 - 10.000 year. Such a reconstruction can be made with the help of physical models. A physical model of the geophysical situation around the event can be made. In these models however multiple parameters remain unknown, most noticeable the discharge and roughness values. Whilst having multiple unknown parameters, standard calibration methods fail. Due to the complex calculation nature of these models, the calculation time constrains the use of multivariable optimization. In this regard meta models might have a solution. A meta model is a simpler model that represents the detailed hydrodynamic model. Within meta models there are two options, namely lower fidelity modelling and data modelling. The lower fidelity modelling is still a physical model with less details, like a coarser grid. Data modelling leaves all physical relations behind and tries to find relations between the input and output of the original model. In this thesis only data modelling is considered as it has the potentially largest speed increase. For the reconstruction of the flooding of 1809, a detailed 2D model with the use of the software D-Flow FM is built. This model describes the geophysical parameters accurately and has a range of parameters for the unknown discharges and roughness sections in the summer bed. This range of parameters gives a sphere of fitting for the used surrogated model, namely a NARX. The NARX is trained and verified on the different potential runs of D-Flow FM. With this a highly accurate trained NARX is created that has an R2 between 0.99 and 0.75, for the best FM run and the worst FM run respectively. This means that the NARX can mimic the D-Flow FM model. The NARX is used in combination with an interior point barrier function algorithm to reconstruct the original discharges and roughness values of the summer bed. The resulting original discharges however were unphysical. Usually the model resulted in discharges over 4*104, which is extremely more than any literature. This shows that the method has some flaws in this specific case. The flaws are most likely caused by one of the following three reasons: too many variables, outside the sphere of fitting, or failing optimization algorithm. First, the too many variables are mostly the number of roughness sections compared to the number of measurement stations. There are relatively few measuring stations in the area, compared to the number of roughness sections. This could be solved by reducing the number of roughness sections. Second, being outside the sphere of fitting, means that the NARX cannot represent the physical model for the measured water levels. This can be fixed by changing the type of experiments used in the training phase. Third, the failing of the optimization algorithm, this means that the problem does not fit to the requirements of the optimization algorithm. To solve this another optimization algorithm can be used. Concluding, the accurate representation of the physical models shows that the NARX is capable in representing the physical model. This is possible even though there are changes to the physical properties of the area. However, it showed that in the current approach there are still difficulties to reach a desired water level. This means that further research is needed in the three aspects that are most likely causing this problem.
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/76742
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page