|
MODEL INDEPENDENCE
PEST can be used with any model without the needing to do any programming.
To calibrate a model:
- Inform PEST which numbers to adjust on model input files
- Identify those numbers on model output files for which there are corresponding field or laboratory measurements
- Instruct PEST how to run the model (assumed to be command-line-driven)
PEST then takes control of the model, running it as many times as it needs to while adjusting parameter values until the discrepancies between model outputs and corresponding field or laboratory measurements are as small as possible in the weighted least squares sense.
It then lists optimal parameter values, an estimate of the uncertainty associated with optimal parameter values, best-fit model outcomes, model-to-measurement residuals, and a suite of statistics related to the optimal parameter and residual sets.
MODEL CALIBRATION & DATA INTERPOLATION
With PEST you can turn any model into a powerful data interpretation package. PEST allows a modeler to truly understand the capacity that a dataset possesses for the estimation of parameters governing the workings of a system, and how supplementary data are most efficiently gathered in order to increase that capacity.
The possibilities for creativity and elegance in data interpretation and model calibration are truly enormous with PEST. The “model” can be a batch file holding one or many executables. Thus you can:
- Calibrate a model using data gathered over non-contiguous time intervals
- Undertake simultaneous calibration of:
- A steady-state and transient model
- A flow and transport model
- Multiple recharge models together with a flow model
- A flow model combined with regularization functionality
There are no limitations with the number of model input files which PEST can write or the number of model output files which PEST can read.
A common mistake in many modeling exercises is to undertake “sensitivity analysis” after a model has been calibrated in order to estimate the uncertainties in model predictions. There are two problems with this approach.
- When a parameter is varied in order to test the effects of this variation on predictive output, the model may become uncalibrated. Thus the prediction cannot be considered a true model prediction.
-
The variation of individual parameters by a small amount in order to assess predictive uncertainty may seriously underestimate the extent to which parameters could actually vary and still keep the model calibrated.
The trick is to vary not just one, but possibly many correlated parameters together, in such a way that the variation of these parameters has virtually no effect on model outcomes under calibration conditions.
It is the variation of this combination of parameters (rather than each parameter individually) which must be undertaken to perform true predictive analysis.
The modeler can ask PEST to calculate the highest or lowest value of a key model outcome, while at the same time ensuring that the parameter values used to make this prediction are such as to keep the model calibrated. The repercussions for model deployment are profound. Now the user can test best and worst case scenarios with ease, and design a fail-safe remediation system and/or optimize the efficiency of a monitoring network.
Predictive Analysis allows you to quantify the uncertainties typically associated with modeling by directly calculating the definitive uncertainty limits on key model predictions.
PEST2000 comes with sophisticated parallel processing capabilities, enabling it to distribute and manage model runs across a network to significantly reduce optimization times. Model calibration or predictive analysis can now be undertaken using more parameters and larger models than has hitherto been possible.
WinPEST’s high impact and informative graphics allow you to understand the calibration and predictive analyses processes like never before. Through a series of evolving run-time displays you can tell at a glance where the process is going, and whether or not your intervention may be required.
When PEST has finished running, WinPEST presents a further array of colourful and educational plots through which you can examine parameter uncertainty and nonuniqueness, analyze calibration residuals and much more.
Take a look at just a few of the plots that WinPEST creates:
View - Parameter values (Line Graph)
View - Composite parameter sensitivities (Line Graph)
View - Objective function value (Line Graph)
View - Parameter correlation coefficient matrix
View - Individual parameter sensitivities from the Jacobian matrix (Bar Chart)
View - Calculated vs. observed values (Scatter Graph)
View - Calibration residuals (Bar Chart)
View - Calibration residuals (Histogram)
View - Normalized Eigenvectors of the Covariance Matrix
WinPEST is interchangeable with PEST2000 (and previous versions of PEST). So you can import your existing PEST datasets and give them a whole new lease of life as they explode into color.
Sometimes the model calibration or predictive analysis process encounters numerical difficulties. If these are hampering a PEST run, WinPEST’s informative displays not only make this plain, but provide information through which troublesome parameters can be identified.
You can then halt PEST execution, hold the offending parameters at their current values, and re-calculate improvements to the other parameters without having to re-compute the Jacobian matrix. Using this unique methodology you can, more often than not, get a stalled calibration process “back on the rails” with minimum wastage of computer time - often a big issue with large and complex models.
|
Shopping Cart (
Accelerate the way you build your FEFLOW and MODFLOW model 
Advanced contaminant transport & groundwater modeling software. 
Two-dimensional contouring and surface plotting software. 
Print this page
