Hydrological models play a crucial role in forecasting future water resource availability and water-related risks. It's essential that they realistically represent and simulate the processes of interest. However, which model structure is most suitable for a given task, catchment and data situation is often difficult to determine. There are only few tangible guidelines for model structure selection, and comparing multiple models simply to choose one to use in further work is a cumbersome process. It is therefore not surprising that the hydrological community has spent considerable effort on improving model parameter estimation, which can be treated as an automatized process, but the selection of a suitable model structure (i.e., the specific set of equations describing catchment function) has received comparatively little attention.
To facilitate easier testing of different model structures, this thesis introduces an approach for Automatic Model Structure Identification (AMSI), which allows for the simultaneous calibration of model structural choices and model parameters. Model structural choices are treated as integer decision variables while model parameters are treated as continuous model variables in this approach. Through combining the modular modelling framework Raven with the mixed-integer optimization algorithm DDS, the testing of different structural hypotheses can thus be automated. AMSI then allows to effectively search a vast number of model structure and parameter choices to identify the most suitable model structures for a specific objective function.
This thesis uses four experiments to test and benchmark AMSI's performance and capabilities. First, a synthetic experiment generates “observations” with known model structures and tests AMSI’s ability to re-identify these same structures. Second, AMSI is used in a real-world application on twelve diverse MOPEX catchments to test the feasibility of the approach. Third, a comprehensive benchmark study explores how reliably AMSI searches the available model space by comparing AMSI’s outcomes to a brute force approach that calibrates all feasible model structures in the available model search space. Fourth, the model space AMSI searches was compared to a much wider model hypothesis space, as defined by 45 diverse and commonly used model structures taken from the MARRMoT-Toolbox. This evaluation of AMSI’s performance is based on mathematical accuracy (tested via statistical metric performance) and hydrological adequacy (tested via the performance on several hydrological signatures) to assess the advantages and limitations of the method.
The re-identification experiments showed that process choices that show little impact on the hydrograph are difficult to re-identify due to near equivalent diagnostic measures. The real-world experiment showed that AMSI is capable of identifying feasible and avoiding infeasible model structures for the twelve tested MOPEX catchments. The performance of the identified models was compared to that of eight other models configured for the MOPEX catchments. AMSI's performance is in the top half of the performance range found by these eight, partially more complex, models, and is therefore considered satisfactory. However, the high variance in the identified model structures with comparable objective function values reflects substantial model equifinality. This was also seen in the benchmark study. While AMSI reliably identifies the most accurate model structures in a given model hypothesis space, the equifinality in model choice as measured through an aggregated metric such as KGE is considerable. In some catchments up to 30\% of the tested model choices obtain comparable KGE scores. These models, however, show significantly different behaviour in their internal storages, showing that a wide range of simulated hydrologic conditions can lead to comparable efficiency scores and therefore a wide ensemble of different model structures may appear suitable. Using AMSI with aggregated statistical metrics therefore provides only limited insights into which models are most suitable for the given catchment. Further investigations showed that the large number of identified mathematically accurate models (as measured through good metric performance) could hardly ever also be considered hydrologically adequate models (as measured through good signature performance). In nine out of twelve catchments none of the accurate models was also considered to be adequate, while only between one (0.1\%) and 49 (0.7\%) of all tested model structures met the defined adequacy requirements in the other three catchments. This glaring disconnect between mathematical accuracy and hydrological adequacy applies to all model selection approaches tested in the benchmark experiments. Neither AMSI, nor the brute force search, nor the MARRMoT models are able to provide accurate as well as adequate model structures when calibrated for the aggregated statistical metric KGE. Therefore, no distinct advantages of commonly used, expert-developed conceptual model structures could be identified over the data-derived AMSI models, as long as model performance is assessed only with aggregated efficiency scores.
This has relevant implications for all modelling studies, as despite many papers suggesting to do otherwise, assessing model performance only through mathematical accuracy (i.e., with scores such as NSE or KGE) has remained the standard practice. The great empirical evidence of the inherent constraints of aggregated metrics such as KGE provided in this thesis may help to convey the message that relying only on these scores simply cannot guarantee hydrologically adequate model structures due to the equifinality in the combined model and parameter selection problem. The results also indicate that the AMSI method is able to identify model structures that are just as mathematically accurate and hydrologically inadequate as previously developed methods for model selection yield, but at a reduced work load to the modeller. Multi-variate datasets and better model performance metrics are often mentioned as ways to reduce equifinality. If such improved methods are implemented in the calibration procedure, AMSI's ability to discriminate between more granular process equations will increase. AMSI could then be a promising way forward to reduce the subjectivity in model selection, and to explore the connections between suitable model structures and catchment characteristics.
| Identifer | oai:union.ndltd.org:DRESDEN/oai:qucosa:de:qucosa:92573 |
| Date | 01 August 2024 |
| Creators | Spieler, Diana |
| Contributors | Schütze, Niels, Merz, Ralf, Melsen, Lieke, Technische Universität Dresden |
| Source Sets | Hochschulschriftenserver (HSSS) der SLUB Dresden |
| Language | English |
| Detected Language | English |
| Type | info:eu-repo/semantics/publishedVersion, doc-type:doctoralThesis, info:eu-repo/semantics/doctoralThesis, doc-type:Text |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | 10.1029/2019wr027009, 10.1029/2023WR036199 |
Page generated in 0.0021 seconds