Within this project a total of 10 peer-reviewed journal publications have been published: Two times each in Scientific Reports, IEEE Access, Chaos and Physical Review Research and once in Nature Communications and New Journal of Physics. All but Chaos are gold-standard open access journals and for the two Chaos publications, we ensured green open access by depositing the published manuscript version both on arXiv and the institute's webservers. These cover the full range of anticipated results, from open data measurements over cascading failures to stochastic tools. Here, we discuss two highlights in more detail: An open database of power grid measurements published in Nature Communications 11, 6362 (December 2020) and a machine-learning approach published in IEEE Access 8, 2020.
In the Nature Communications article, we collected power grid data from 17 locations across three continents and covering 12 synchronous areas - regions containing different power plants and consumers that are connected and operate under the same frequency. This was a major collaborative effort involving scientists from six institutions (Forschungszentrum Jülich, Queen Mary University of London, Karlsruhe Institute of Technology (KIT), Technical University Dresden and Istanbul University). We used these experimental recordings to test theoretical predictions on how the size of a synchronous area influences its stability. In particular, we found that smaller areas tend to be much more volatile than larger areas in their fluctuations in frequency and we quantified this empirical observation with stochastic modelling.
Furthermore, by simultaneously measuring frequencies in several locations within a synchronous area, we also observed that whilst on longer time scales of minutes or more, the frequencies were identical everywhere, on a shorter time scale of seconds, substantial differences between locations were observed. We quantified the time needed for two locations to fully synchronize.
Meanwhile, in the IEEE Access article, we demonstrated how open power system data can be used to fuel modern machine learning applications. In particular, we investigated how a weighted-nearest-neighbour approach can be used to obtain precise forecasts of the power-grid frequency. We showcased how our new predictor consistently outperforms the daily profile (daily averages of the trajectory). The choice of weighted nearest neighbours as our forecasting method allowed us to identify patterns in the power grid frequency time series and advance our understanding of the underlying power system and the frequency dynamics.
