Building and Environment, 294, 114384
Environment|구충완
Operational day-ahead load forecasting using graph-based data-driven urban building energy modeling
2026International Journals
Research AreaEnvironment
Professor구충완
AuthorsKim, J., Yeom, S., Ann, S., Seo, S., Koo, C., and Hong, T
Publication Year2026
Journal (Volume, Issue, Pages)Building and Environment, 294, 114384
AbstractThis study proposes a day-ahead urban building energy modeling (UBEM) framework that employs graph neural networks (GNN) to forecast campus-scale electricity demand. Addressing heightened electrification and outage risks during extreme heat events, the framework formulates multi-building load forecasting as a spatially contextual learning problem, explicitly modeling inter-building interactions through graph representations. Implemented at a university campus using three years of hourly electricity data from 24 buildings, the framework integrates observed and forecasted weather variables with building-level solar radiation features derived from EnergyPlus simulations. Buildings are represented as nodes in static distance- and similarity-based graphs, with spatiotemporal dependencies learned via graph neural networks, including graph convolutional network (GCN), graph attention networks (GAT), and GraphSAGE layers, coupled with a gated recurrent unit (GRU). A weather-aware attention mechanism and graph-based imputation module address forecast uncertainty and missing data under non-stationary conditions. The selected GraphSAGE–GRU model achieves an R² of 0.962 and weighted mean absolute percentage error (WMAPE) of 9.93 %, outperforming multi-building and single-building baselines. Distance-based graphs consistently outperform similarity-based graphs, and a GraphSAGE–GRU architecture offers a favorable trade-off between accuracy and computational cost. Robustness analyses show WMAPE increases only 1.5 percentage points under 40 % node-level missingness. Moreover, the framework maintains stable performance over one-year operational simulations and demonstrates improved peak-load tracking during severe heatwaves. Overall, this study presents an empirically validated Graph-based UBEM framework and offers guidance on graph construction, model selection, and training protocol for campus-scale day-ahead load forecasting.
Urban Building Energy ModelingGraph Neural NetworksGated Recurrent NetworksDay-Ahead Load ForecastingPower Outages
PrevA systematic review on the integrations of CFD and artificial intelligence for the future perspectives of built environment: Multi-scale approaches from room to cityNextVirtual reality-based experimental analysis of personality and cognitive traits on task performance and safety in novice tower crane operators