Parts | Days | Selection | Search | Updates | Downloads | Help
DY: Fachverband Dynamik und Statistische Physik
DY 57: Networks: From Topology to Dynamics II (joint session DY/SOE)
DY 57.8: Talk
Friday, March 22, 2024, 11:45–12:00, BH-N 128
Network Science and Beyond -- Can Network Measures capture Mechanisms of Desynchronization in Complex Networks? — •Christian Nauck — Potsdam Institute for Climate Impact Research, Germany
This study addresses the fundamental question of how network function emerges from topology, particularly in nonlinear oscillator networks. While traditionally network measures have been discovered, recent advances in Machine Learning (ML), notably Graph Neural Networks (GNNs), provide an alternative for predicting network function. Through a comprehensive literature review, we identify 46 network measures, integrating them with conventional ML (NetSciML) to predict dynamic stability in power grids. Our findings reveal that a complete set of measures rivals GNNs in performance on the same ensemble, offering advantages such as reduced data requirements, shorter training times, and enhanced interpretability. However, NetSciML falls short in predicting stability across varied grid sizes, suggesting that GNNs employ a distinct and potentially more mechanistic approach. This underscores GNNs' potential to overcome challenges faced by current network science-based methods, providing novel solutions for desired outcomes.
Keywords: Machine Learning; Network Science; Graph Neural Networks; Oscillator networks; Power grids