Enhancing Graph Transformers for Greener Campus Microgrids

Abstract

The growing integration of distributed energy resources (DERs) within campus microgrids creates a myriad of operational challenges while also creating multiple opportunities for improved reliability, sustainability, and cost efficiency. Existing work has looked at Transformer-based forecasting models (e.g., Informer, Autoformer, FEDformer, Pyraformer) and combined deep learning models using CNNs, LSTMs, and attention. However, these works generally treat forecasting and control independently, utilize static or simplified grid formations, and suffer from scalability issues for real-time implementation. In this paper, we present our unified Spatio-Temporal Graph Transformer (STGT) framework to jointly perform load and renewable generation forecasting, while simultaneously optimizing control decisions for the DERs. The model utilizes dynamic graph attention to learn changing spatial–temporal interactions across campus buildings and DER units, and embeds physical power-flow and reliability constraints directly into the optimization. By incorporating forecasting and control into one integrated model, the framework is able to concurrently schedule batteries, inverters, and flexible loads to minimize operational cost, peak-demand charges, and emissions. Various experiments show that our proposed STGT provides accurate forecasting, reliable stability predictions, and effective multi-objective optimization, demonstrating its applicability for real-time microgrid energy management.