Calibration and Reliability in Groundwater Modelling: Credibility of Modelling

(Proceedings of ModelCARE 2007 Conference, held in Denmark, September 2007). IAHS Publ. 320, 2008, 267-271.

Model predictive error: how it arises and how it can be accommodated

JOHN DOHERTY

Watermark Numerical Computing, Brisbane, Australia

johndoherty@ozemail.com.au

Abstract This paper presents a brief overview of the means by which the potential error associated with predictions made by a calibrated model can be estimated. This is important both in its own right, and as a means of optimizing the acquisition of future data to reduce this potential error. It is shown that there are two major contributors to the predictive error of a calibrated model, referred to herein as the “null space” and “solution space” contributions. The first is an outcome of the fact that system properties can only be estimated in broad detail through the calibration process; to the extent that a prediction is sensitive to finer detail than this, its potential for error is not reduced through model calibration. The second is an outcome of the fact that even estimates of broad-scale system properties carry some error, this resulting from the fact that they are estimated on the basis of data contaminated by (measurement and structural) noise. Methodologies for computation of these two terms are presented; reference is made to publicly available software through which these methodologies are implemented.

Key words regularisation; uncertainty; parameters; calibration