8 Calibration, sensitivity analysis and uncertainty analysis

Due to the bioregional assessment (BA) requirement that groundwater modelling should take as many forms of uncertainty as possible into account, a ‘conventional’ (i.e. deterministic) calibration process will not be followed in the bioregional assessments. A global sensitivity analysis will be conducted on each model prediction using as many parameters as possible within the model. The sensitivity analysis will determine which parameters each model prediction is most sensitive to. The uncertainty analysis will be conducted using plausible ranges of values for each of the sensitive parameters using (i) a Monte Carlo procedure if there are no constraining data available or (ii) a Markov Chain Monte Carlo procedure when there are data to constrain the prediction. As the computational cost of a thorough uncertainty analysis using groundwater models is generally prohibitive, a limited number of model runs (i.e. in the order of thousands) will be conducted to train a statistical model emulator. For each prediction of interest, a Gaussian Process emulator will be built which can be run more efficiently (and enable model runs in the order of tens of thousands) to quantify the probability distribution function of the required output. The details of the uncertainty analysis can be found in the companion submethodology M09 (as listed in Table 1) for propagating uncertainty through models (Peeters et al., 2016).

Constraining data to be used for model predictions will ideally include hydraulic heads in various aquifers as well as fluxes such as baseflow and volumes of co-produced water. The location of these data points will need to be evaluated thoroughly to ensure they are responding to regional stressors rather than local effects (which are not captured by the regional model).

The sensitivity and uncertainty analyses undertaken for BAs require that groundwater models are built with this use in mind. This will require robust models that are capable of converging for a broad range of parameter values. This will likely require model grid simplifications to aid convergence and reduce run times. These requirements have been defined before a groundwater model can be passed to the risk team:

  1. Coal resource development pathway (CRDP)
    1. the final CRDP is implemented in the model.
  2. Model nodes
    1. a preliminary list of model nodes is identified (90% final)
    2. model output for these locations is generated through the observation functionality of MODFLOW or via ZONEBUDGET (not via post processing the heads or budget file in a graphical user interface (GUI)).
  3. Parameterisation
    1. an exhaustive list of parameters is compiled. For each parameter it describes:
      1. name
      2. units
      3. description (in case of parameter zones, reference needs to be made to maps and cross-sections)
      4. preferred value
      5. minimum plausible value
      6. maximum plausible value (the plausible range of hydraulic properties is expected to vary over at least two orders of magnitude).
    2. The value of each parameter can be changed via a script in an automated way, either via the native parameter functionality of MODFLOW or via a custom script.
    3. The parameterisation and model run can be executed as a single batch-file from command line, independent of the GUI used for development.
  4. Convergence
    1. The model converges for the steady state, baseline transient and CRDP transient for the preferred parameter values. The model convergence criteria for these runs are not to be changed in the subsequent stress testing runs.
    2. The model also converges for the extreme parameter combinations (e.g. minimum plausible recharge with maximum plausible hydraulic conductivity with minimum specific storage).
    3. Non-converging parameter combinations can be acceptable if a sound hydrogeological reason is provided for the non-convergence.
    4. In case of acceptable non-convergence parameter combinations, the most extreme parameter combination of that type for which the model converges needs to be established.
  5. Head and flux observations
    1. An objective function is formulated, combining and weighting all historical observations, both heads and fluxes.
    2. The objective function is part of the model output, either via the native parameterisation and observation functionality of MODFLOW or customised scripting.
Last updated:
17 October 2018