Semi-supervised State Classification in Distribution Systems with Graph Neural Networks

Andrey Poddubnyy, Minh-Quan Tran, Phuong H. Nguyen, Han Slootweg

Abstract: With the wide penetration of distributed energy resources and renewables in the modern power grids, especially in the distribution and consumption levels there appears a high growth in uncertainty and the number of possible scenarios. Some of them may violate grid constraints and thus, have to be a subject of control. Considering the inherently low level of observability in distribution systems, estimation and assessment of their state becomes a significant challenge, that must be tackled before developing a control strategy. Moreover, the real time state estimation in partially-observable DS can be undetermined or timely infeasible in case of when the small time steps needed. In this paper we propose to use the graph neural network - based state classification, which identifies the possible grid limits violations in the distribution grid. The use of neural network in this case is in line with their main advantages over physical models - they are fast and adaptive. We benchmark several graph neural network approaches, that are based on the topology of the grid, which allows to derive information about unlabeled nodes and increase interpretability - one of the main obstacles on the way to wide neural network utilization in power systems. We set the problem in a semi-supervised manner, which allows us to use less labeled, metered data, as the large parts of modern distribution systems remain not measured. We show that graph neural networks are more precise and explainable in this task compared to the regular ones.