PEST - confidence intervals and data units.

I have been using PEST to get a handle on the controling factors of my model.  It returns values of hydraulic conductivity about 100,000 times higher that I was expecting and so I thought it was returning K expressed as m/day - the same as they are written in the element file - and not in value x 10-4m/s as they are shown in the GUI editor.  I turned these back into 10-4 units by dividing by 86400, put them into the model and got completely different results.  I then checked the confidence limit calculations in the PEST rec output file and I got a reasonable but not exact agreement (taking into account the log transformation used - although I am having trouble working out how PEST calculates the degrees of freedom for the Confidence Interval calculation).  After a day of two of scratching my head I noticed that the values in the PEST output were the same as values in the Feflow GUI editor with the important difference that the factor of 10-4 was missing.  It seems that PEST transfers the mantissa to and from Feflow but does the output calculations based upon the mantissa alone.  This means that the covariance matirx in my problem is 10E8 times too big and the confidence intervals are way out.

I have been having problems getting the PEST model to produce 'sensible' answers and tried to 'force' the model into the right magnitude by including flows but no matter how I start the model - i.e. with positive or negative values for flows (either transfer or flux) the model floods.  Given the above, it made me wonder if it is a 'problem' with the way data are transferred between PEST and Feflow.  I have partly resolved it by using a fixed head boundary and fixing the hydraulic conductivity of the layer next to the boundary to control the cross flow.

Has anyone noted similar problems with unit transfers or resolved whether to include negative signs in PEST for flows into or out of the model?

There was a post earier on about whether people thought PEST was useful.  I wouldn't go any further that accepting that it is useful, and I certainly wouldn't accept its findings for all but the simplest problems.  In my case, it gets the RMS (phi) error down but the parameter values it produces in doing so are pretty much nonsense for the aquifer in question.  I have seen models with almost perfect filts to the observation data but the hydraulic conductivity/transmissivity arrays look like patchwork quilts and I can't believe they resemble the world they are trying to model.  Minimizing the RMS error seems to have become a Holy Grail at the expense of understanding the underlying controls on the hydrogeology and I think PEST etc. are partly to blame for this.  I think that the Principle of Parsimony should be applied more liberally and the one of the minimizing the RMS error a bit less so.

William

I can't help you on the first matter, but as far as I remember I had no problems with unit transfers.

On the second matter, I agree with you in that PEST is just a tool, it can't build the model by itself.
As a tool, it is indeed helpful in highlighting the errors in the model: if it minimizes the RMS, and you're sure that it's the global minimum (not a local one), and you get unreasonable parameter values, than:

1. your head, flux or concentration measures are wrong
AND/OR
2. you're focusing only on some parameters, keeping fixed (i.e. considering well known) parameters that should be changed together with the first ones
AND/OR
3. your conceptual model is wrong, you're not simulating what really happens.

Unfortunately, PEST can't tell where's the problem, understanding the system's structure is all in our hands.

Good luck!

Giovanni Formentin