Predictive Event-Driven Transmission–Distribution System Coordination via Graph Neural Networks and Proactive Dispatch

Abstract

This paper proposes a predictive event-driven framework for transmission–distribution system (TS-DS) coordination that leverages graph neural networks (GNNs) with multi-head temporal attention to forecast imminent congestion, voltage violations and contingency risks. A proactive dispatch layer solves a chance-constrained optimal power flow (CC-OPF) problem with moment-based ambiguity sets, pre-positioning generation and activating flexible demand before predicted violations materialise. The central novelty is a closed prediction–optimization loop in which calibrated, spatially resolved GNN event probabilities adaptively set the per-element chance-constraint risk levels in proportion to predicted severity—a coupling absent from prior GNN-OPF and distributionally robust dispatch work. Supporting contributions include: (i) a spatiotemporal GNN combining graph attention convolution with hierarchical TS-DS pooling and causal temporal attention and (ii) a risk-aware event-driven triggering protocol balancing event probability against expected violation severity. Cantelli's inequality provides a distribution-free SOC reformulation of individual chance constraints. Experiments on IEEE 118-bus/6 × 33-bus (316 buses) and Polish 2383-bus/20-DS (3743 buses) systems over five random seeds demonstrate $73.2\%\pm 1.8\%$ voltage violation reduction, $68.0\%\pm 1.6\%$ congestion reduction and $11.5\%\pm 0.4\%$ operational cost reduction versus state-of-the-art reactive methods, with sub-5-second computation on both test systems.