Dear Groundwater ProfessionalIt appears that the groundwater modeling community is divided on performing transient model calibration using the automated calibration tools. I would like to share the responses which I received to my last posting. (I have also attached my last posting as the GROUNDWATER mailing list members didn’t receive it due to the temporary shutdown of their list-server.)
Assigning weights for observations were discussed in the USGS report 98-4005 (Methods and Guidelines for Effective Model Calibration) and in the PEST manual. The USGS report provided a statistical basis to reflect measurement errors and the PEST manual discussed assigning weights to include different types of observations in the model calibrations. The weights discussed in these reports may only be applied based on the locations where the observations were collected and cannot be used to assign weights for observations at particular locations (ie assigning different weights to time-variant observation from a bore).
Transient groundwater modeling requires reasonably accurate input data which mainly includes time-variant recharge and groundwater pumping. It was pointed out that quantifying recharge for input to transient groundwater models is a challenging task and any observations correspond to uncertain recharge should be weighted out from the objective function which is mathematically minimized by PEST/UCODE in order to select optimum model parameters. It was warned that including these observations in the automated calibration in any other way by assigning different weights, is simply playing with numbers and may lead to hydrogeologically unrealistic model parameters.
Some of the modelers were very critical about mathematically comparing the observed data with the model simulated data, as the observed data was collected at a point in time and the model simulated data at the end of stress periods, assuming the stresses within the stress periods are uniformly distributed.
It was also cautioned that formation of illogical objective functions to include in the automated calibration combined with the errors in hydrological conceptualization may provide misleading unrealistic parameters, as the direct deterministic inverse models have limitation on resolving uncertainties resulting from conceptualization errors.
It was mentioned that the inverse models (PEST/UCODE) ideally require independent observations and the observed water levels from a particular bore are somewhat correlated to each other. The errors in estimating model parameters by including dependent observed with the current inverse models have not been investigated yet. As such, it was suggested that to compare resultant “Q” (inflow-outflow) in each cell per stress-period rather than water levels. Apparently, there is a software code available to provide the resultant “Q”(net-source) for each cell per stress-period with observed groundwater levels. During the automated model calibration, the model simulated net-source (Q) terms are compared with the estimated net-source terms (recharge – groundwater pumping), while any doubtful estimation of net-source terms are completely weighted out in the objective function. I wonder whether any of this type of “detect” model is available for finite-difference grid. This is an interesting concept I would like to hear more on this.
So, what is the verdict. My intuition tells me that if we formulate the object function correctly we may able to use these inverse models for transient model calibration with continuos assistance from modelers. What are the steps involved in formulating the objective function…? How do you apply the inverse models effectively in transient model calibration with real data….? Is the automated transient model calibration an illusion for the real-world models….?
Anyway, Jury (ASTM Subcommittee D18.21) is still out on this, I am eagerly waiting for their verdict. I would like to extend my thanks to everyone who responded to my postings. If anyone (especially software developers) have any other thoughts please share their thoughts with the rest of the world.
José Vaughan
E3 Consultancy
Earth, Environmental and Economics Consultants
we work for sustainable future
>From: "Jose Vaughan" <jose_vaughan@hotmail.com>
>To: GWM-L@gw.cciw.ca
>CC: epoeter@mines.edu
>Subject: Ground Water Model Calibration
>Date: Thu, 10 Feb 2000 11:18:10 GMT
>
>Dear modelers
>
>I would like to seek information on assigning weights in PEST/UCODE for
>calibrating a ground water model. These automated calibration tools are
>increasingly used to calibrate transient ground water models. These tools
>provide ground water modelers with the ability to match time-variant
>observed data with the model simulated data. In my experience with
>irrigation area management models, the distribution of calibrated model
>parameters (hydraulic conductivity, specific yield and storage coefficient)
>provided by these tools, appear to be hydrogeologically unrealistic. These
>tools provide unrealistic values for various reasons due to errors in
>conceptualization, input data etc. I also feel that providing incorrect
>weight for each observation in the automated calibration (PEST / UCODE) may
>lead to these unrealistic values.
>
>I am currently developing a simple and practical methodology to assist
>modelers in assigning the weight for each observation. This methodology
>will only assist modelers in making a decision on weighting observation and
>the modelers have to look at each observation individually (possible
>measurement errors & uncertainty in input data have also to be considered)
>in assigning the weights. The following are my thoughts on this. I wonder
>whether you know anything on assigning weights, or have you got any thoughts
>on this. Your contribution would be very much appreciated and if you wish,
>your response would be kept confidential.
>
>The field observed data mainly ground water levels in observation bores,
>are collected at a point in time and sometimes during/after flooding or
>pumping from near by bore. The ground water model simulates water levels at
>the end of stress periods (usually months) and the stresses (recharge and
>ground water extraction) on the ground water systems are uniformly
>distributed in the stress periods.
>
>Transient simulation starts with a steady-state simulation to confirm the
>initial conditions (heads) support the ground water equation. The model
>hydraulic conductivity parameters are calibrated at this time. Due to
>expansion in irrigation areas, the stresses on the model area are changed
>and the model parameters have to be re-calibrated in order to include the
>changes in stresses.
>
>I feel that it would be acceptable to calibrate hydraulic conductivity
>parameters with PEST/UCODE and the weight on each observation should reflect
>the TREND IN GROUND WATER LEVELS. This may be achieved by performing
>statistical analysis on the deviation observed water levels from 3-point
>average water levels .
>
>Since, the storage parameters reflect the fluctuation of water levels, a
>DIFFERENT set of weights should be assigned to each observation by
>considering frequency and time of observations, when calibrating the model
>storage parameters with PEST/UCODE.
>
>Further, the model hydraulic conductivity parameters might also be
>calibrated by comparing model simulated values for 3-point average input
>(recharge & ground water extraction) with 3-point average water levels.
>
>Please accept my ignorance. If my thoughts are stupid please let me know.
>
>Thanks
>
>José Vaughan
>
>E3 Consultancy
>Earth, Environmental and Economics Consultants
>we work for sustainable future
Email this topic to a friend
Printer-friendly version of this topic
