Determining Observable Islands in Power Systems: A Graph Machine Learning Approach
M. Stanila (TU Delft - Electrical Engineering, Mathematics and Computer Science)
M. Skrodzki – Mentor (TU Delft - Electrical Engineering, Mathematics and Computer Science)
Jerry Guo – Mentor (Alliander)
B. Mamudi – Mentor (TU Delft - Electrical Engineering, Mathematics and Computer Science)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
State estimation (SE) is an important tool in power system management, alleviating the electrical traffic congestion problem by estimating the voltage angle and magnitude at every bus using an incomplete set of measurements.
Observability analysis represents a sub-problem in the SE pipeline. It determines if the system is fully observable. If the network is not fully observable, many algorithms will split it into observable subnetworks.
Machine learning (ML) methods have been sparsely previously used to solve observability analysis, as they tend to focus on full SE. Prior ML work solves the problem through a binary result of observable or unobservable, without dividing the grid into islands. Moreover, the data used in training those models, or the generation algorithm that created it were never made public.
We present an algorithm that randomly assigns measurements to components of a given network and computes its observable islands used for generating a training dataset. We then propose a graph machine learning architecture that can identify observable islands in a power grid from its topology and measurements. Our architecture follows the variational autoencoder (VGAE) structure, consisting of three graph convolutional layers and a graph attention layer.
The architecture was trained and evaluated on the IEEE 14, 30, 39, 57 and 118-bus cases with consistent results across all cases, alongside a general model trained on a dataset consisting in a combination of all the network sizes.
Moreover, a direct comparison with the existing literature shows that our solution produces fairly accurate results, with 5 out of 6 test cases being correctly identified as fully observable or unobservable.