- Home
- Assessments
- Bioregional Assessment Program
- Gloucester subregion
- 2.6.1 Surface water numerical modelling for the Gloucester subregion
- 2.6.1.5 Uncertainty
- 2.6.1.5.1 Quantitative uncertainty analysis
The aim of the quantitative uncertainty analysis is to provide a probabilistic estimate of the change in the hydrological response variables due to coal resource development at the receptors. A large number of parameter combinations are evaluated and, in line with the Approximate Bayesian Computation outlined in companion submethodology M09 (as listed in Table 1) for propagating uncertainty through models (Peeters et al., 2016), only those parameter combinations that result in acceptable model behaviour are accepted in the parameter ensemble used to make predictions.
Acceptable model behaviour is defined for each hydrological response variable based on the capability of the model to reproduce historical, observed time series of the hydrological response variable. For each hydrological variable, a goodness of fit between model simulated and observed annual hydrological response variable is defined and an acceptance threshold defined.
The ensemble of predictions are the changes in hydrological response variable simulated with the parameter combinations for which the goodness of fit exceeds the acceptance threshold. The resulting ensembles are presented and discussed in Section 2.6.1.6.
Design of experiment
The parameters included in the uncertainty analysis are the same as those used in the calibration, with the exception that in the uncertainty analysis parameter ne_scale is included.
Table 7 lists the parameters used in the uncertainty analysis and the range uniformly sampled in the design of experiment. The AWRA-L parameters in Table 7 are explained in the AWRA-L v4.5 documentation (Viney et al., 2015).
Table 7 Summary of AWRA-L parameters for uncertainty analysis
AWRA-L = Australian Water Resource Assessment Landscape, na = data not applicable
Through a space filling Latin Hypercube sampling (Santer et al., 2003), 10,000 parameter combinations are generated from the AWRA-L parameters, with the ranges and transform in Table 7. These ranges and transforms are chosen by the modelling team based on previous experience in regional and continental calibration of AWRA-L (Vaze et al., 2013). These mostly correspond to the upper and lower limits of each parameter during calibration.
The parameter combinations are generated together with the parameter combinations for the regional analytic element groundwater model and the alluvial groundwater model (see companion product 2.6.2 for the Gloucester subregion (Peeters et al., 2018)). This linking of parameter combinations allows the results to consistently propagate from one model to another, as outlined in the model sequence section (Section 2.6.1.1).
Each of the 10,000 parameter sets is used to drive AWRA-L to generate streamflow time series at each 0.05 x 0.05 degree (~5 x 5 km) grid cell (Jones et al., 2009).
Observations
Four catchments used for the calibration contribute flow to the river systems in the Gloucester subregion. For these catchments the historical observations of streamflow are summarised into eight of the nine hydrological response variables for all years with a full observational record (ZFD is not used because it is not a meaningful HRV in rivers with perennial flow (Section 2.6.1.4)). The equivalent historical simulated hydrological response variable values are computed from the 10,000 design of experiment runs. The goodness of fit between these observed and simulated historical hydrological response variable values is used to constrain the 10,000 parameter combinations and select the best 10% replicates (i.e. 1000 replicates) that are used for predictions in Section 2.6.1.6.
Predictions
For each of the 30 receptor catchment nodes the post-processing of design of experiment results in 10,000 time series with a length of 90 years of hydrological response variable values for baseline, , and coal resource development conditions,
.
These two time series are summarised through the maximum raw change (amax), the maximum percent change (pmax) and the year of maximum change (tmax). The percentage change is defined as:
|
(3) |
As the predictions include the effect of surface water – groundwater interaction through the coupling with the groundwater models, it is possible that the groundwater parameters affect the surface water predictions.
Sensitivity analysis
Figure 11 shows the sensitivity indices of the absolute change in the 1st percentile of flow in all receptor catchment nodes to all parameter values for both surface water and groundwater models. These are computed with the density based algorithm described in Plischke et al. (2013) from the results of the design of experiment. It is very clear from this plot that there are only a handful of AWRA-L parameters that control the change in the 1st percentile of flow. These are consistent across catchments. None of the parameters of the groundwater models have a sizeable impact, mainly because of the limited size of the change in baseflow due to coal resource development, compared to the total streamflow.
Figure 11 Sensitivity indices and parameter values for the surface water and groundwater models
The figure shows sensitivity indices of the absolute change in the 1st percentile of flow in all receptor catchment nodes (x axis) to all parameter values for both surface water and groundwater models (y axis). High values indicate high sensitivity of the prediction to a parameter.
See Section 2.6.2.6 in companion product 2.6.2 for the Gloucester subregion (Peeters et al., 2018) for explanations of the parameter names.
Selection of behavioural parameter combinations
The acceptance threshold for each hydrological response variable is set to the 90th percentile of the average goodness of fit between observed and simulated historical hydrological response variable values obtained from four gauges. This means that out of the 10,000 model replicates, the 1000 best (or 10% best) are selected for each hydrological response variable.
The selection of the 10% threshold is based on two considerations: (i) guaranteeing enough prediction samples to ensure numerical robustness and (ii) their performance approaching to that obtained from the high-streamflow and low-streamflow model calibrations. Furthermore, it is expected that the full 10,000 replicates contain many with infeasible parameter combinations and that these are likely to be filtered out by sampling only the best 10% of replicates. Nevertheless, selecting the 10% best replicates is determined arbitrarily, and its strength and weakness are further discussed in section 2.6.1.5.2.
