Synthesizing Distribution Grid Congestion Data Using Multivariate Conditional Time Series Generative Adversarial Networks

Distribution grid congestion is a significant obstacle to integrating distributed energy resources, leading to voltage instability and overloading grid elements. Existing probabilistic models cannot directly generate realistic multivariate time series data with grid bottleneck characteristics. Generating multivariate time series data that capture photovoltaic and load patterns across correlated buses while performing power flow calculations is inherently complex. These challenges, data compliance issues, and the need for more training data suggest the exploration of Artificial Intelligence methods and the generation of edge test data. This paper introduces Multivariate Conditional Time-series Generative Adversarial Networks (MC- TimeGAN), designed for the conditioned generation of synthetic load and photovoltaic generation profiles. MC-TimeGAN simulates severe grid con-gestion scenarios by purposefully manipulating the respective labels passed to the model and provides data augmentation. Applying this methodology to a validated benchmark dataset for distribution grid shows a significant and realistic increase in grid congestion. Evaluation by power flow calculations, using the dataset generated by MC- TimeGAN shows increase the mean transformer load by 8% and the mean line load by up to 14% compared to the original data. Together with dimensionality reduction techniques, we demonstrate that synthetic data are similar to original data.