Groundwater management is often supported by numerical groundwater modelling. In the Murray-Darling Basin, groundwater models support an accounting framework that manages how the States and Commonwealth manage River Murray salinity. Historically these models have been manually calibrated, using a process informed by expert knowledge where judgement calls can be difficult to document comprehensively. In contrast, automated calibration methods should enable greater transparency.
To improve transparency, a trial of the automated calibration Parameter Estimation Tool (PEST) was undertaken for the Border to Lock 3 groundwater model, which is maintained by South Australia’s Department for Environment and Water. The model estimates the salt moving from groundwater into the River Murray.
During the trial, numerous PEST options were tested for their impact on the calibration. Both pilot point and zone approaches were tested. The calibration was constrained using combinations of: Tikhonov regularization, prior information, SVD, SVD assist, and lower and upper bounds. Different numbers of pilot points were used. Steady-state and transient runs were considered. Calibrated parameters included aquifer properties, irrigation recharge rates, and boundary condition parameters. The objective function evaluated the match between observed and modelled potentiometric heads, weighted by data quality and relevance. Other data sources were used as a check on model outputs but were not included in the objective function.
The trial highlights a number of considerations when using PEST for more complex management models. Calibration results can be very sensitive to PEST options, sometimes in unanticipated ways. Model run times become extremely important, as PEST calibration may take weeks or months to generate. Iterations of PEST runs are also needed as initial results inform further constraints or changes in both the model and PEST setup. This process can be challenging to undertake within the timeframes associated with model development. Policy and management constraints also need to be explicitly brought into the calibration process, which can prove challenging. Most importantly, an automatic procedure must be tempered by actual physical knowledge of the regional area.