Abstract
In addition to requesting water quality models that contain a robust and reliable calibration, clients and stakeholders are increasingly demanding some measure of model uncertainty. The estimation of parameter values for water quality models like WaterCAST can be a difficult task for modellers, especially for large and complex models. The uncertainty surrounding the resulting parameter values, and the predictions from the model that use these values, is hard to quantify and communicate to a broad audience. The methodologies investigated by this study aim to more adequately satisfy the clients driving the production of these models. A model-independent parameter estimation program, PEST, has been used to calibrate the basic rainfall-runoff and constituent generation models contained in a complex WaterCAST project representing the Fitzroy River Catchment in central Queensland, Australia. The calibrated parameter values have been calculated through a supervised approach which incorporates both Tikhonov and subspace techniques of mathematical regularisation, maintaining a high degree of modeller satisfaction and confidence. The calibration process considers all of the complex interactions possible within the subcatchment-node-link network, meaning that optimised parameter values are immediately suitable for application to the model. One of the PEST tools employed in this process provides assistance in applying singular value decomposition. This assistance significantly reduced the number of model runs needed to achieve a satisfactory calibration. The same assistance was then used to attempt the unsupervised calibration of 100 random parameter sets in a Monte Carlo style approach. A high level of model run efficiency of the recalibration process is achieved through using null space projection of random parameter fields to replace solution space components with those estimated through the previous calibration exercise. A total of 50 of these random parameter sets were able to have the statistical objective function minimised by PEST to within 0.5% of that achieved with the supervised calibration. These 50 calibrated parameter sets allow a quantifiable analysis of the uncertainty surrounding both the range of parameter values that provide a satisfactory calibration, and the range of model predictions that each of these parameter sets leads to. The same techniques were also applied to sediment generation processes in a sub-region of the Fitzroy catchment. By calibrating model parameters within the WaterCAST environment, rather than in a secondary environment like the Rainfall Runoff Library, the calibration was able to consider all internal interactions. This has resulted in a statistical fit of predictions to observations that is a clear improvement (average coefficient of efficiency for daily flow at 20 locations increased from 0.50 to 0.79, average percent flow volume difference at 20 locations reduced from 91% to 15%). This calibration also satisfies the modellers need for an acceptable visual fit. Estimates of sediment generation from the WaterCAST model are accompanied by uncertainty estimates that most encapsulate the variability seen in a sporadic and variable observation data set.
Original language | English |
---|---|
Pages | 3158-3164 |
Number of pages | 7 |
Publication status | Published - 2009 |
Externally published | Yes |
Event | 18th World IMACS Congress and International Congress on Modelling and Simulation: Interfacing Modelling and Simulation with Mathematical and Computational Sciences, MODSIM 2009 - Cairns, Australia Duration: 13 Jul 2009 → 17 Jul 2009 |
Conference
Conference | 18th World IMACS Congress and International Congress on Modelling and Simulation: Interfacing Modelling and Simulation with Mathematical and Computational Sciences, MODSIM 2009 |
---|---|
Country/Territory | Australia |
City | Cairns |
Period | 13/07/09 → 17/07/09 |
Keywords
- Null space
- Parameter estimation
- Regularisation
- Uncertainty