Event-Aware Graph Neural Networks for Real-Time Optimal Power Dispatch in Distribution Microgrid

Overview

Distribution microgrid operation is becoming increasingly constrained by renewable variability, distributed energy resources and contingency events. The classic dispatch method is still AC optimal power flow (ac-OPF), which can take hundreds of milliseconds to dispatch, which is not real time.

Problem Statement

While graph neural network (GNN) surrogates can greatly shorten OPF computing time, they normally learn a system state-to-dispatch-action mapping that is contingency independent. The physics-informed methods that are currently available also use soft constraints which are used to promote, but not guarantee, feasibility.

Methodology

We propose an Event-Aware Graph Neural Network (EA-GNN) that combines the following modules: disturbance-conditioned graph attention, conservative-dispatch learning and feasibility restoration. Contingencies are represented by a three-dimensional descriptor: Line outage, load surge magnitude, and generator loss indicator. The descriptor supports the information flow and the zero-shot adaptation to a different disturbance for different contingency classes without retraining. Training is a combination of supervised learning from the interior point OPF solutions and the penalties for power balance and generator limits, which are informed by the physics. A lightweight differentiable projection ensures by-construction satisfaction of generator operating limits. Based on complexity analysis, the asymptotic speed up compared to sparse interior-point AC-OPF is O(N^1.5).

Results

The dispatch is computed by EA-GNN in 5.3ms and 7.8ms on the IEEE 14-bus system and IEEE 33-bus feeder, respectively, which is 74.8× and 81.2× faster than AC-OPF, respectively. The power-balance residuals were close to baseline, and EA-GNN had no constraint violations, whereas AC-OPF had 21 violations.

Conclusion

EA-GNN brings contingency-conditioned graph attention, by-construction feasibility guarantees and enhanced scalability for real-time dispatch.