Author(s): Aristotelis Mantoglou

Linked Author(s):

Keywords: Inverse groundwater modeling; stochastic groundwater modeling; heterogeneous aquifers; random fields; groundwater simulations; parameter estimation

Abstract: This paper investigates the prediction ability of models of various complexities and the optimal model structure depending on the real system complexity and the quality of measurements. A simulation approach is used where the underlying transmissivity map is assumed a member of a random field and multiple realizations of transmissivity maps are generated using stochastic simulation. For each simulation run, several models with different structures and complexities, based on zonation parameterization, are calibrated and the model transmissivities are estimated using measurements of head data. Then, using the calibrated models, the heads and the flow terms are evaluated in the numerical cells of the aquifer. The corresponding prediction errors in the heads and flow terms are calculated and the optimal models having the lowest prediction errors are determined. The method is applied in a case study of a two-dimensional aquifer with steady groundwater flow with pumpage. For small correlation lengths of transmissivity variability, the numerical results indicate that a simple singlezone model yields the best predictions and outperforms complex models with more parameters. For larger correlation lengths however, models of medium complexity have best prediction characteristics and should be preferred. It is further shown quantitatively that as the quality of measurements decreases the number of parameters and the complexity of the optimal model decreases. The reduced flexibility of a model with fewer parameters protects it from being adversely affected from the errors contained in the measurements.

DOI: https://doi.org/10.1080/00221680509500156

Year: 2005