- Title
- A robust and efficient optimization algorithm for hydrological models
- Creator
- Qin, Youwei
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2017
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- Hydrological models are widely used in research and engineering applications. An ongoing challenge for hydrologic modelling is that some or all model parameters need to be calibrated to observed data. In past decades, much research has been directed to developing robust and computationally efficient optimization algorithms for hydrological modelling. However, there remains a strong demand for optimization algorithms that are both robust and efficient. Manual calibration, which relies on the hydrologist’s experience, can be labor-intensive and time consuming. The Shuffled Complex Evolution (SCE) method search, widely regarded as the most robust search method, has been shown to incur significant computational cost, particularly in highly dimensioned problems where its robustness can also be deteriorated. Gradient-based Newton-type methods have been reported to perform poorly in hydrological model calibration because of the small and large-scale nonsmoothness and multi-modality of the objective function surface. In response to this challenge, this thesis develops a robust Gauss-Newton method to preserve the inherent computational efficiency of the Gauss-Newton method, while substantially improving its robustness using the following three novel heuristics: 1) adoption of an adaptive sampling scale strategy to enable the search to “see” the large-scale features of the objective function surface; 2) use of a best sampling point strategy to take advantage of the opportunities of jumping to different regions of attraction when using a large sampling scale; and 3) a relaxation of the singular value decomposition truncation threshold to assist the search to escape from a local optimum. These three heuristics were applied in conjunction with a random multistart strategy. The performance of the robust Gauss-Newton (RGN) method was validated with extensive numerical experiments using four computationally “fast” conceptual hydrological models, HYMOD, SIXPAR, SIMHYD and FUSE-536, applied to three contrasting Australian catchments. RGN was compared against SCE and benchmark Gauss-Newton and quasi-Newton methods. The results showed that the RGN and SCE methods had comparable robustness for the HYMOD, SIXPAR, and SIMHYD models; while for the FUSE-536 model, RGN was more robust than SCE. Of practical significance was the finding that RGN very significantly outperformed SCE in terms of the computational cost. The comparative performance of RGN was also evaluated using two computationally “slow” models, the river basin model IQQM and the distributed watershed model SWAT2009. It was found that RGN was significantly more efficient but comparably robust as SCE for calibration of the IQQM model with 6 and 14 parameters. In the case of 38 parameter SWAT2009 calibration, RGN very significantly outperformed SCE both in robustness and computational efficiency. Moreover, the trajectory traces also showed that RGN exhibited faster and better convergence behavior than PEST implementation of the Gauss-Newton method. The numerical experiments clearly suggest that in representative hydrological model calibration ranging from simple conceptual to complex distributed models, the robust Gauss-Newton method exhibits robustness comparable to the SCE search, but with a significantly superior efficiency. This finding is of particular significance for the calibration of computationally costly distributed and highly parameterized hydrological models.
- Subject
- hydrological model; model caliration; RGN; SCE; PEST; river basin model
- Identifier
- http://hdl.handle.net/1959.13/1350350
- Identifier
- uon:30536
- Rights
- Copyright 2017 Youwei Qin
- Language
- eng
- Hits: 1321
- Visitors: 1573
- Downloads: 293
Thumbnail | File | Description | Size | Format | |||
---|---|---|---|---|---|---|---|
View Details Download | ATTACHMENT01 | Thesis | 5 MB | Adobe Acrobat PDF | View Details Download | ||
View Details Download | ATTACHMENT02 | Abstract | 337 KB | Adobe Acrobat PDF | View Details Download |