On-Meter Graph Machine Learning: A Case Study of PV Power Forecasting for Grid Edge Intelligence

A team of researchers has published a groundbreaking study demonstrating the practical deployment of Graph Machine Learning (GML) models directly on smart meters, a key piece of hardware at the 'grid edge.' The paper, 'On-Meter Graph Machine Learning: A Case Study of PV Power Forecasting for Grid Edge Intelligence,' focuses on using Graph Neural Networks (GNNs) to predict solar power generation within a microgrid. The researchers successfully trained and compared two specific models—Graph Convolutional Networks (GCN) and GraphSAGE—and then ported them to run on the meter's limited hardware using the ONNX and ONNX Runtime ecosystem, even developing a custom ONNX operator for the GCN.

The core achievement is moving AI inference from the cloud to the very edge of the electrical grid. The case study used real-world data from a village microgrid to validate the approach, showing both models could execute successfully on the smart meter itself. This enables real-time, localized forecasting without relying on constant cloud connectivity, which is crucial for managing the variable output of solar panels and maintaining grid stability. The work represents a significant step toward truly intelligent, decentralized energy systems where each node can make autonomous decisions.