Publikationen - Netze für Erneuerbare Energien

Markus de Koster,
"Power Quality State Estimation in Distribution Systems Based on Neural Networks",
Masterarbeit TH-Köln, 30.Okt.2023

This thesis explores the suitability of physics-aware neural networks for power quality state estimation in distribution grids. For that, power quality data was generated in a simulated environment and used for training and evaluation of dierent neural network models. Comprehensive data analyses were carried out, focusing on optimal representations and processing of power quality data for neural network applications. Comparative assessments with traditional fully connected architectures demonstrated the superior capabilities of physics-aware models, which utilize the physical grid structure as a regularization mechanism.
Despite increased computational complexity, eective methods were identified to address challenges posed by deep architectures and sparse connectivity. Best performing models achieved a mean squared error loss of 4 × 10−6, signicantly outperforming traditional models with a loss of 1.1 × 10−5. The results strongly indicate that physics-aware neural networks are suitable for power quality state estimation tasks. Promising avenues for expanding developed models include incorporation of physical laws in the learning process, potentially further constraining the network to physically possible states.


> Abschlussarbeit: PDF-Dokument (1.8 MB)
Öffentliche Version mit reduzierter graphischer Auflösung.


> Swarm Grid 

> Projekt QUIRINUS-Control

Voltage Predictions
Voltage predictions for a single randomly chosen iteration of the test set at 50 Hz.
Physics-aware neural networks (PANN) and dense neural networks (DNN) predict values for all nodes based on inputs at three nodes with added measurement noise.

E.Waffenschmidt, 3.Nov.2023