![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Data driven model generation based on computational intelligence |
VerfasserIn |
Peter Gemmar, Oliver Gronz, Christophe Faust, Markus Casper |
Konferenz |
EGU General Assembly 2010
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 12 (2010) |
Datensatznummer |
250041093
|
|
|
|
Zusammenfassung |
The simulation of discharges at a local gauge or the modeling of large scale river catchments
are effectively involved in estimation and decision tasks of hydrological research
and practical applications like flood prediction or water resource management.
However, modeling such processes using analytical or conceptual approaches is made
difficult by both complexity of process relations and heterogeneity of processes. It
was shown manifold that unknown or assumed process relations can principally be
described by computational methods, and that system models can automatically be
derived from observed behavior or measured process data. This study describes the
development of hydrological process models using computational methods including
Fuzzy logic and artificial neural networks (ANN) in a comprehensive and automated
manner.
Methods We consider a closed concept for data driven development of hydrological models
based on measured (experimental) data. The concept is centered on a Fuzzy system using rules of
Takagi-Sugeno-Kang type which formulate the input-output relation in a generic structure like
Ri : IFq(t) = lowAND-¦THENq(t+-³t) = ai0 +ai1q(t)+ai2p(t--³ti1)+ai3p(t+-³ti2)+-¦.
The rule’s premise part (IF) describes process states involving available process information,
e.g. actual outlet q(t) is low where low is one of several Fuzzy sets defined over variable q(t).
The rule’s conclusion (THEN) estimates expected outlet q(t + -³t) by a linear
function over selected system variables, e.g. actual outlet q(t), previous and/or
forecasted precipitation p(t --³tik). In case of river catchment modeling we use head
gauges, tributary and upriver gauges in the conclusion part as well. In addition, we
consider temperature and temporal (season) information in the premise part. By
creating a set of rules R = {Ri|(i = 1,-¦,N)} the space of process states can be
covered as concise as necessary. Model adaptation is achieved by finding on optimal
set A = (aij) of conclusion parameters with respect to a defined rating function
and experimental data. To find A, we use for example a linear equation solver and
RMSE-function.
In practical process models, the number of Fuzzy sets and the according number of rules
is fairly low. Nevertheless, creating the optimal model requires some experience.
Therefore, we improved this development step by methods for automatic generation of
Fuzzy sets, rules, and conclusions. Basically, the model achievement depends to a
great extend on the selection of the conclusion variables. It is the aim that variables
having most influence on the system reaction being considered and superfluous ones
being neglected. At first, we use Kohonen maps, a specialized ANN, to identify
relevant input variables from the large set of available system variables. A greedy
algorithm selects a comprehensive set of dominant and uncorrelated variables. Next,
the premise variables are analyzed with clustering methods (e.g. Fuzzy-C-means)
and Fuzzy sets are then derived from cluster centers and outlines. The rule base is
automatically constructed by permutation of the Fuzzy sets of the premise variables.
Finally, the conclusion parameters are calculated and the total coverage of the input
space is iteratively tested with experimental data, rarely firing rules are combined
and coarse coverage of sensitive process states results in refined Fuzzy sets and
rules.
Results The described methods were implemented and integrated in a development
system for process models. A series of models has already been built e.g. for rainfall-runoff
modeling or for flood prediction (up to 72 hours) in river catchments. The models required
significantly less development effort and showed advanced simulation results compared to
conventional models. The models can be used operationally and simulation takes only some
minutes on a standard PC e.g. for a gauge forecast (up to 72 hours) for the whole Mosel
(Germany) river catchment. |
|
|
|
|
|