Robust Sampling Methods for Accurate Power Flow Time-Sweep Approximations with Graph Neural Networks

Training precise machine learning models, such as neural networks, for steady-state power flow analysis relies on high-quality datasets that closely resemble the target applications. Synthetic data generation often acts as a valuable complement, necessitating the construction of a probabilistic model that accurately captures dependencies among power grid components. In this study, we present user-friendly data generation techniques tailored for power flow time-sweeps. These methods incorporate the underlying time series using adequate distribution estimation and dependency modeling based on copulas. The generated data is used to train graph convolutional neural networks. Experiments are conducted on a large real-world-sized power grid with 2000 buses. Our results demonstrate that the accuracy of the approximations achieved with specific sampling methods strongly depends on the amount of time-series data available. Particularly, with an appropriately chosen sampling method these networks can be trained effectively with relatively few samples while maintaining highly accurate approximations.