Environmental Engineering Reference
In-Depth Information
can play a crucial role for the overall system dynamics. When these parameters
aggregate many external and internal influences, the resulting high context speci-
ficity will limit the global applicability of the model and its outcome. If an
underlying mechanism can be assumed and measurement data are available, it
is frequently possible to determine which parameter value would be best to
minimize the differences between modelled and observed data. In the sensitivity
analysis it is possible to resolve which parameters are most critical for further
adjustments.
23.3.2 Calibration
Calibration (also referred to as “parameter identification”) is a procedure in which
model parameters are changed to minimize the difference of the model output and a
given set of measurement values. For this purpose, a fully developed, executable
ecological model must be available. In addition, a target dataset is required that
demarcates the output which the model should generate in case of an ideal fit. It is
important to ensure that this dataset is independent of the measurements which were
used for model specification, because otherwise it would strongly limit any conclu-
sion on the reliability of the identification process.
Systematically, parameter values are varied, applied to the model and the
outcome is compared, to see whether the fit was improved. The direction and extent
of the change is then used to calculate a new set of parameter values which is again
tested in an iterative procedure in which the quality of the fit can be expressed
quantitatively (Janssen and Heuberger 1995).
As with other optimization processes, it is not guaranteed that the iterative
procedure always finds the best overall fit when stopping at an extreme value.
Therefore, it is usually advisable to start parameter identification as close as
possible to the assumed values, and repeat the procedure with a number of slightly
different starting points to assess whether the results remain comparable. Often we
can find a strong dependency between model complexity and data requirements.
Lack of data and the use of over-parameterized models may also limit the success of
model calibration (see, e.g. Marsili-Libelli and Checchi 2005).
Although most of the established optimization techniques have been originally
developed and applied for differential equation-based models as these regularly
operate with aggregated parameters, there are also adapted calibration processes
being developed for individual-based models IBM (e.g. Pereira et al. 2008), which
operate in analogy, i.e. change quantities in model specification and compare the
results with regard to an optimality criterion.
The approach of machine learning (see Chap. 19) conceptually expands the
approach beyond a variation and adaptation of parameter values. In addition also
functional expressions (e.g. different forms of nonlinearity) are changed and then
tested, as to whether this would lead to an improved approximation of the target
values.
