Simplification error analysis for groundwater predictions with reduced order models

Abstract

Groundwater resource management often requires detailed numerical models that make calibration and predictive uncertainty analysis computationally challenging. Reduced order models (ROMs) alleviate the computational burden but can potentially lead to predictive bias and underestimation of uncertainty. A paired model approach has previously been proposed to estimate the predictive uncertainty of models compared to highly complex, synthetic realities. This approach is modified to analyze and compare the simplification error for groundwater predictions of a real-world numerical MODFLOW model of the Wairau Plains Aquifer in the Marlborough Region of New Zealand. Two different ROM types were applied in this study to predict groundwater heads, spring discharge and river–groundwater exchange fluxes: (1) a drastically simplified MODFLOW model, and (2) artificial neural networks (ANNs). The different ROMs exhibit very different patterns of bias and magnitude of model simplification error. The method accurately captures the simplification error for most predictions by the MODFLOW model, but underestimates the error for predictions highly dependent on the variability of the complex model. The simplified MODFLOW model shows significant parameter surrogacy and non-optimality of simplification, thus questioning the adherence to expert-knowledge based parameter limits. For predictions where historic data sets are available, ANNs provide superior predictive power. However, they cannot be applied to predictions of data types and locations not contained in the calibration data set. For those predictions, simplified numerical models can be applied with varying degree of accuracy.