Leveraging power grid topology in machine learning assisted optimal power flow
This work addresses the computational complexity of optimal power flow for power grid operators by evaluating neural network architectures, though it is incremental as it focuses on comparing existing methods rather than introducing new ones.
The paper systematically compares fully connected, convolutional, and graph neural networks for machine learning-assisted optimal power flow, finding that CNN and GNN offer marginal benefits over FCNN for fixed grid topologies, but GNN outperforms others with variable topologies like line contingencies.
Machine learning assisted optimal power flow (OPF) aims to reduce the computational complexity of these non-linear and non-convex constrained optimization problems by consigning expensive (online) optimization to offline training. The majority of work in this area typically employs fully connected neural networks (FCNN). However, recently convolutional (CNN) and graph (GNN) neural networks have also been investigated, in effort to exploit topological information within the power grid. Although promising results have been obtained, there lacks a systematic comparison between these architectures throughout literature. Accordingly, we introduce a concise framework for generalizing methods for machine learning assisted OPF and assess the performance of a variety of FCNN, CNN and GNN models for two fundamental approaches in this domain: regression (predicting optimal generator set-points) and classification (predicting the active set of constraints). For several synthetic power grids with interconnected utilities, we show that locality properties between feature and target variables are scarce and subsequently demonstrate marginal utility of applying CNN and GNN architectures compared to FCNN for a fixed grid topology. However, with variable topology (for instance, modeling transmission line contingency), GNN models are able to straightforwardly take the change of topological information into account and outperform both FCNN and CNN models.