A state-wide assessment of optimal groundwater hydrograph time series models

Abstract

A number of groundwater hydrograph time series models have been proposed over recent years but, to our knowledge, there has been no systematic review of their performance and thus no means of selecting a model to suit the prevailing conditions. This paper presents an evaluation of the new groundwater hydrograph time series models presented in Peterson and Western (2011) against existing models on 620 bore hydrographs distributed throughout Victoria. Bores that monitor water level under natural conditions and having at least 20 years of data were used. The aim of this study is to rigorously demonstrate the strength of the Peterson and Western (2011) models (hence referred to as soil moisture store-transfer function noise model, or SMS-TFN) and ascertain which forms of the various soil moisture components within the model perform best and under what conditions. To assess the relative performance, the widely used HARTT model (Ferdowsian et al. 2001, Ferdowsian et al. 2002) and the standard transfer function noise model (von Asmuth et al., 2002) were also investigated. This investigation into the groundwater head time series modelling was assessed by evaluating the performance of eleven model variants of three classes of models (SMS-TFN, TFN, HARTT) using the Coefficient of Efficiency (CoE) and the Akaike Information Criterion (AIC) as the performance measures for the calibration and evaluation periods. The results showed that the SMS-TFN model (Peterson and Western, 2011), significantly improves the predictive model performance compared to the performance of the traditional TFN model. The SMS-TFN model with ground water recharge as the forcing component shows better model calibration and predictive performances than models with infiltration as the forcing component. These model variants produced the best median calibration period CoE of 0.655 (where 1.0 is a perfect fit) and the best evaluation period unbiased CoE of 0.270 (see Figure 1). The predictive performance of the HARTT model was shown to be highly variable and inconsistent across the bore hydrographs tested. If a sustainable linear time trend exits in the bore hydrograph, the model produced good results as indicated by performance measures during both calibration and validation periods. However, in the absence of such trends, the model performed poorly. This illustrates the potential risk in assuming a non-climatic time trend which may or may not exist in the bore hydrographs. More importantly, the SMS-TFN model with ground water recharge as the forcing component was shown to be the most robust model which can explain most of the bore hydrographs from climate data alone.