Artificial expansion of power quality datasets using generative adversarial networks

Modern power grids are experiencing a rapid shift from traditional power plants towards highly decentralized and volatile energy generation by renewable sources. Generation units, such as wind turbines and solar panels are often integrated using inverter technology. This shift leads to power grids, which are more often operating under low-inertia condition. This can lead to an increase in Power Quality Disturbances. In order to react to such developments, advanced power quality monitoring systems are necessary. Current research into algorithms working with this monitoring data often employ machine learning. However, the training data sets used in this research can be small and/or sparse. In order to enhance future research into machine learning-based power quality monitoring, we investigate a possible approach for training data expansion. This approach is based on Generative Adversarial Networks, a modern data generation technique. A specialized Generative Adversarial Network (GAN) is tested on a mathematically modeled reference data set. The results show promising examples, but also uncover several challenges resulting from the mechanics of GAN themselves. Further research is needed to overcome these challenges and enable GAN to be used to their full potential.