Recharge Model Performance
- Reviewed predictive performance of 4 differently complex recharge models that are currently used or can potentially be used in conjunction with groundwater allocation planning in Western Australia
- Provides regulators with proof-of-concept for how to use a suite of recharge methods in future groundwater models.
- Guidance for model parameters and improved understanding of bias in predictive performance for future models.
Overview
In south-west Western Australia, groundwater allocation planning decisions for aquifers close to or above full allocation is usually assisted by numerical models to estimate groundwater levels. These models need to account for changing climate, landcover, abstraction and artificial recharge. Recharge is a key flux in these models, particularly under a changing climate. To date, recharge in the Perth Region of WA is simulated using the Vertical Flux Model Manager (VFM), which contains complex physical processes, but is slow to simulate and difficult to parameterise. Recharge models with less complex physical process simulation can avoid these issues and are used in other areas on the Swan Coastal Plain, but these may incur predictive bias due to over-simplification. Replacing the VFM with simpler, less data intensive models that are easier to parameterise and less computationally expensive requires that they are reliable and capable of assessing and reducing uncertainty of decision-salient predictions that inform groundwater management.
The acceptability of a simpler recharge model depends on whether simplification-induced errors (a) exceed the uncertainties a simpler model is capable of quantifying and reducing through history matching, and (b) are significant in comparison to the total prediction uncertainty. It also depends (c) on the impact of prior parameter definition and observation weights on these trade-offs. Such uncertainty-bias trade-offs introduced by model simplification were assessed using the simple-complex analysis. The complex model’s parameterisation of soil and land-use properties was kept spatially variable at a cell level (VFM cell-by-cell). The selected four simpler models are characterised by increasingly simpler recharge process representation:
- “VFM-lumped” is a spatially upscaled version of the VFM across different soil and land use types;
- “MODFLOW-OWHM” is based on an analytical model and 1-D-simpfilied Richard’s equation;
- “LUMPREM” is a lumped-parameter model that uses van Genuchten soil-water constitutive functions;
- “EMPRICAL model from lookup tables” is trained on a VFM model.
Locations of three typical cross sections across the Swan Coastal Plain
Outcomes of Simple-vs-Complex Analysis
The predictive performance against the complex model was measured by the simple model’s predictive bias and variance and ability to include the ‘truth’ within the quantified uncertainty range (defined as ‘feasibility’). We also explored the impact of prior data conflicts and the accommodation of predictive bias by increased prior variance. High-level guidance for the Regulator was provided on the utility of simpler models, their relative impact on prediction error, systematic bias through parameter lumping or recharge process simplification, or their potential improvements to reduce uncertainty and bias:
- Under historical stresses, uncertainty and bias were greater for all models when simulating net recharge as opposed to heads. For the latter, no significant bias was evident using any of the simpler models, but the VFM-Lumped did perform best.
- Prior data conflicts of several models call for a better definition of the prior probability distribution to mitigate the impact that conceptual differences, spatial lumping, or recharge-process simplification have on bias and uncertainty.
- Pragmatic considerations for the model selection are ease of use, stability, and run time, which were best for the EMPIRICAL model, worst for the VFM models, and medium for LUMPREM and OWHM.
The project provides regulators with proof-of-concept for using a suite of recharge methods in future groundwater models. It provides guidance for setting prior parameters for landcover/soil combinations generally found on the coastal plain and an understanding of bias in predictive performance for some of the land cover/soil combinations and some methods to avoid the potential for predictive bias.
Cross-Section of the combined groundwater and recharge model with numbered Observations and Prediction cell locations
History-matching against complex model outcomes at monitoring bore heads and net recharge at lysimeters using PEST-IES