Physics-informed data-driven topology identification in power distribution networks with adversarial robustness enhancement

Abstract

Accurate topology identification is essential for ensuring reliable and stable operation of distribution networks, especially under limited monitoring and frequent switching operations that make timely situational awareness challenging. This work proposes a data-driven topology identification method that embeds physics-informed feature engineering by constructing power flow residual features with respect to a reference topology. To enhance robustness against perturbations, adversarial training is incorporated into the learning process. The proposed method requires only limited microphasor measurement units (μPMUs) deployment, supports both radial and meshed configurations, and enables fast topology inference. Experiments on the 33-, 69- and 118-node systems demonstrate that the proposed method achieves high accuracy and stable performance across varying perturbation strengths.