Generative AI applied for synthetic data in PMU

Abstract

The growing deployment of Phasor Measurement Units (PMUs) has enhanced power system observability but introduced new challenges related to data privacy, incompleteness, and measurement quality. To address these issues, this paper proposes a data-driven methodology for generating and completing PMU phasor measurements using Generative Artificial Intelligence. Specifically, we employ Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) trained on real-world PMU datasets to learn the underlying empirical data distributions without assuming predefined statistical models. The proposed deep generative models are evaluated against traditional statistical techniques based on Gaussian Copulas using a suite of distributional similarity metrics, including Kullback–Leibler (KL) divergence, Hellinger distance, Maximum Deviation Nearest Neighbor (MDNN), and the Kolmogorov–Smirnov (KS) test. The GAN model achieved the best distributional fidelity, with KL divergence as low as 0.0106 and Hellinger distance of 0.0435 for voltage signals. In a synthetic data reconstruction task with 0.5% missing values, the GAN reduced the percentage root mean squared error (PRMSE) to 0.52% for voltage and 2.19% for current—significantly outperforming baseline methods. Moreover, the GAN was able to augment the dataset from 1489 to 5000 samples while preserving key statistical properties, as validated by empirical distribution tests. These results demonstrate that deep generative models not only offer superior accuracy but also provide statistically consistent synthetic PMU data, making them a robust alternative to conventional methods for enhancing power system datasets.