KIT | KIT-Bibliothek | Impressum | Datenschutz

Physics-informed geometric deep learning for inference tasks in power systems

Jongh, Steven de 1; Gielnik, Frederik 1; Mueller, Felicitas 1; Schmit, Loris 1; Suriyah, Michael 1; Leibfried, Thomas 1
1 Institut für Elektroenergiesysteme und Hochspannungstechnik (IEH), Karlsruher Institut für Technologie (KIT)

Abstract:

In this paper, geometric deep learning techniques are applied to learn approximate models for power system estimation and calculation tasks. Nine different graph neural network architectures from literature are compared for this purpose. The underlying graph and known physical algebraic equations are taken into account during training, which allows to learn inductive, approximate algorithms for state- and powerflow estimation. The learned models are applied on randomly generated synthetic electrical medium voltage grids that are generated based on typical grid properties. It is shown, that the learned models are able to estimate the system states with high accuracy and that the trained neural networks can be applied to previously unseen grid topologies. Different sensor configurations and assumed random sensor noise and failures are taken into account. The trained neural networks are able to estimate the states with high accuracy despite of high sensor failure rates as well as noise that is added to the system.


Originalveröffentlichung
DOI: 10.1016/j.epsr.2022.108362
Scopus
Zitationen: 4
Dimensions
Zitationen: 5
Zugehörige Institution(en) am KIT Institut für Elektroenergiesysteme und Hochspannungstechnik (IEH)
Publikationstyp Zeitschriftenaufsatz
Publikationsmonat/-jahr 10.2022
Sprache Englisch
Identifikator ISSN: 0378-7796
KITopen-ID: 1000148721
Erschienen in Electric Power Systems Research
Verlag Elsevier
Band 211
Seiten Article no: 108362
Nachgewiesen in Dimensions
Web of Science
Scopus
KIT – Die Forschungsuniversität in der Helmholtz-Gemeinschaft
KITopen Landing Page