Predictive analysis

When you first run PEST, it runs in parameter estimation mode. This selects values for the input parameters which cause the outputs to most closely follow the measured values, according to the least-squares rule. The quality of fit is given by PHI " the lower this value is, the better the fit.

The purpose of predictive analysis is to determine the range over which a model output can vary while still being consistent with the measured values supplied. To get maximum and minimum predictions, PEST allows the PHI value to be larger than the minimum value found during parameter estimation.

To do a predictive analysis on a Simile model, check the checkbox in the predictive analysis frame on the Actions tab, and select a model value for which to make a prediction. This can be any non-parameter value in the model, not including those in variable-membership submodels. It can be a value which you are already using for comparison with measured data " for instance, if you are using data measured up to a certain point in time to calibrate the model but then want to predict how the same value will behave at later times on the basis of this calibration, or even use prediction to fill in a gap in your measured data.

You can also make the following selections:

• Whether to predict maximum or minimum values

• What multiple of the original PHI to allow for the predictions

• At what time(s) the prediction will apply.

There are some values on the Settings tab which affect PEST's behaviour during predictive analysis; these are explained in the PEST manual. Simile supplies default values for these, based on the results obtained when running PEST in parameter estimation mode.

If the value being predicted is an array, or if the frequency and duration you set for prediction times are such that a prediction will be made at more than one time, then a complete run of PEST must be done for each prediction, because PEST itself can only predict one value at a time. Since each PEST run includes multiple model runs, predicting large numbers of values can take a very long time. The following steps might speed things up:

• Set display interval to 0, so display tools do not take up time

• Select 'use current values as estimates' on the Inputs tab after the parameter estimation run, so each prediction run starts with good parameter estimates.

During a prediction run, the progress bar shows how many of the prediction runs have been completed. Predicted values will be inserted into the results field, and can be displayed in tabular form and saved as a .csv file using the View button.

It is worth noting that each prediction is arrived at independently. So, if using predictive analysis to create 'error bands' around a graph of model outputs, it might be the case that although each individual value might vary as far as the predicted limits, it would be impossible for the graph line to actually follow the minimum or maximum limits, since multiple predicted values on the limit might be incompatible with each other. Remember, however clever PEST might appear, it's still only a computer program.

In: Contents >> Running models >> PEST