add oscillator dataset

[chrisn-pik]

Adding a new dataset, consisting of ensembles of oscillator networks.

The data has been published in the following papers:
Toward dynamic stability assessment of power grid topologies using graph neural networks
https://doi.org/10.1063/5.0160915
Towards dynamic stability analysis of sustainable power grids using graph neural networks
https://www.climatechange.ai/papers/neurips2022/16

The data is available on Zenodo: https://zenodo.org/records/8204334

There are three datasets: osc20, osc100, and osctexas. osc20 consists of 10,000 grids with 20 nodes per grid, osc100 consists of 10,000 grids with 100 nodes each. osc20 and osc100 are split 70:15:15.

The task involves nodal regression, where the goal is to predict the probabilistic measure of single-node basin stability for each node. This measure represents the likelihood that the entire grid will return to a synchronized state after being perturbed.

Besides considering osc20 or osc100 individually, training on osc20 and evaluating on osc100 is an interesting out-of-distribution generalization task. For evaluation purposes only, osctexas is a single grid with 1,910 nodes inspired by Birchfield et al. https://electricgrids.engr.tamu.edu/electric-grid-test-cases/activsg2000/ to further test the out-of-distribution generalization.

Synchronization of non-linear oscillators is a crucial phenomenon in many real-world systems, including cognitive functions of brains, pacemaker cells in a beating heart, and the stable operation of power grids. However, exact numerical simulations of large systems of coupled oscillators are exceedingly expensive. The underlying dynamical simulations are computationally very expensive and took roughly 700,000 CPU hours.