Graph Neural Network Based Prediction of Data Traffic in Cyber-Physical Smart Power Grids

Abstract

Smart power grids rely on tight integration of physical and cyber layers. Event-driven (ED) data packets are generated in the cyber layer in response to emergency conditions in the grid such as weather conditions, physical sabotage, cyber-attacks, etc. For proper management of the resources at the cyber layer, and hence, timely delivery of ED packets, efficient prediction of ED traffic conditions is required. Since the stochastic arrival process of ED packets is attributed to several factors, a data-driven prediction approach is appealing. However, this is challenged by: (a) unavailability of datasets capturing ED packet arrivals and departures at the cyber layer of the power grid, which are needed to train and test the data-driven models, (b) sparsity of the ED traffic data as emergency conditions are rare, and such sparsity impedes the learning process of data-driven models based on gradient descent, and (c) inability of traditional models, e.g., multilayer perceptron (MLP), long short-term memory (LSTM), convolutional neural networks (CNNs), to present accurate prediction as they fail to capture the interactions among the routers within the cyber layer. To address these challenges, this paper: (a) proposes a method to generate ED traffic based on real emergency reports in the U.S. power grid, (b) proposes a pre-processing method to convert the sparse ED traffic data into dense data, and (c) proposes a topology-aware prediction model based on graph neural network (GNN) to accurately predict the network condition. Our results demonstrate the superior performance of the proposed GNN-based approach.