 |
| Titel |
Interactive model evaluation tool based on IPython notebook |
| VerfasserIn |
Sophie Balemans, Stijn Van Hoey, Ingmar Nopens, Piet Seuntjes |
| Konferenz |
EGU General Assembly 2015
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 17 (2015) |
| Datensatznummer |
250109152
|
| Publikation (Nr.) |
 EGU/EGU2015-9033.pdf |
|
|
|
|
|
| Zusammenfassung |
| In hydrological modelling, some kind of parameter optimization is mostly performed. This
can be the selection of a single best parameter set, a split in behavioural and non-behavioural
parameter sets based on a selected threshold or a posterior parameter distribution derived
with a formal Bayesian approach. The selection of the criterion to measure the
goodness of fit (likelihood or any objective function) is an essential step in all of these
methodologies and will affect the final selected parameter subset. Moreover, the
discriminative power of the objective function is also dependent from the time period
used.
In practice, the optimization process is an iterative procedure. As such, in the course of
the modelling process, an increasing amount of simulations is performed. However, the
information carried by these simulation outputs is not always fully exploited. In this
respect, we developed and present an interactive environment that enables the user to
intuitively evaluate the model performance. The aim is to explore the parameter space
graphically and to visualize the impact of the selected objective function on model
behaviour.
First, a set of model simulation results is loaded along with the corresponding parameter
sets and a data set of the same variable as the model outcome (mostly discharge). The ranges
of the loaded parameter sets define the parameter space. A selection of the two parameters
visualised can be made by the user. Furthermore, an objective function and a time
period of interest need to be selected. Based on this information, a two-dimensional
parameter response surface is created, which actually just shows a scatter plot of the
parameter combinations and assigns a color scale corresponding with the goodness of fit
of each parameter combination. Finally, a slider is available to change the color
mapping of the points. Actually, the slider provides a threshold to exclude non
behaviour parameter sets and the color scale is only attributed to the remaining
parameter sets. As such, by interactively changing the settings and interpreting the
graph, the user gains insight in the model structural behaviour. Moreover, a more
deliberate choice of objective function and periods of high information content can be
identified.
The environment is written in an IPython notebook and uses the available interactive
functions provided by the IPython community. As such, the power of the IPython notebook as
a development environment for scientific computing is illustrated (Shen, 2014). |
|
|
|
|
|
|