Event Classification in Active Distribution Grids Using Physics-Informed Graph Neural Network and PMU Measurements

Power distribution systems are prone to experience several events such as sudden load or distributed energy re-source (DER) fluctuations, or asset failures that can impact their reliability and resilience. Event classification is one of the important initial steps in event analysis to determine remedial actions that mitigate the consequence of incidents. Therefore, developing an efficient event classification method is important for operators to monitor and manage network performance in real-time, increasing the situational awareness. In this paper, a physics-informed graph neural network (GNN) model is proposed for event classification. This model uses phasor measurement unit (PMU) data such as voltage magnitude and angles from a limited number of nodes, as an input to a graph convolution network (GCN). Besides the temporal data captured by PMUs, the physical configuration of the network is also provided to the GCN to consider the spatial relationship among the limited installed PMUs. The proposed physics-informed approach is examined on the modified IEEE 34-bus distribution grid with added DERs and with considering noisy and missing PMU samples. The proposed GNN model is compared with other methodologies such as decision tree and logistic regression. The proposed model shows a superior performance according to the classification metrics.