Sensitivity analysis, uncertainty quantification and parameter estimation for a numerical tide and storm surge model
File(s)
Author(s)
Warder, Simon Charles
Type
Thesis
Abstract
Storm surges pose a significant hazard to coastal communities worldwide. Accurate and reliable storm surge numerical models, combined with an understanding of model uncertainties, are therefore vital. In this work, an adjoint-capable numerical coastal ocean model, Thetis, is extended and used to perform a sensitivity analysis for a hindcast case study in the North Sea, revealing the spatial patterns of the sensitivity of modelled surges to three model inputs. These sensitivities are used to gain physical insight, and to perform uncertainty quantification for each of the model inputs. The results indicate that, while the greatest contribution to uncertainty is made by meteorological inputs, uncertain bottom friction is nevertheless significant.
This motivates a comparison of parameter estimation methods for a spatially varying bottom friction coefficient. This is performed for a tide-only case study consisting of the Bristol Channel and Severn Estuary. Here, a gradient-based optimisation method via the adjoint model is compared with Bayesian inference via a Markov Chain Monte Carlo algorithm, utilising a Gaussian process emulator as a surrogate for the full numerical model. Three friction parameters, based on the distribution of sediment types within the model domain, are estimated, with the results from each calibration method consistent within the estimated parameter uncertainties. Furthermore, the estimated parameters are found to reduce model-observation misfit for a second numerical model, TELEMAC-2D, suggesting that the calibration process has identified physically meaningful parameters.
Since such a model calibration relies on the availability of observation data, this work considers a framework for the identification of observation locations which will perform best in a parameter estimation exercise. In addition to an application within the Bristol Channel case study, this work takes steps towards guiding new observations in the Maldives, where present datasets are severely limited and the development of well-calibrated numerical models would be highly valuable.
This motivates a comparison of parameter estimation methods for a spatially varying bottom friction coefficient. This is performed for a tide-only case study consisting of the Bristol Channel and Severn Estuary. Here, a gradient-based optimisation method via the adjoint model is compared with Bayesian inference via a Markov Chain Monte Carlo algorithm, utilising a Gaussian process emulator as a surrogate for the full numerical model. Three friction parameters, based on the distribution of sediment types within the model domain, are estimated, with the results from each calibration method consistent within the estimated parameter uncertainties. Furthermore, the estimated parameters are found to reduce model-observation misfit for a second numerical model, TELEMAC-2D, suggesting that the calibration process has identified physically meaningful parameters.
Since such a model calibration relies on the availability of observation data, this work considers a framework for the identification of observation locations which will perform best in a parameter estimation exercise. In addition to an application within the Bristol Channel case study, this work takes steps towards guiding new observations in the Maldives, where present datasets are severely limited and the development of well-calibrated numerical models would be highly valuable.
Version
Open Access
Date Issued
2020-04
Date Awarded
2020-08
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Piggott, Matthew
Cotter, Colin
Ham, David
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/L016230/1
Publisher Department
Earth Science & Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)