Graph Neural Network‐Based Prediction of Building Energy Consumption

ABSTRACT

This study develops a framework for predicting building energy consumption using Graph Neural Networks (GNNs). A two‐story reference building with seven thermal zones was represented as a graph, where each zone was a node and thermal or spatial interactions were edges. EnergyPlus simulations, driven by real hourly weather data from Tehran, London, and New York, generated the training dataset. The GNN was implemented in PyTorch Geometric and trained on these outputs. The proposed model predicts cooling and heating loads with high accuracy. It achieved mean absolute errors (MAE) of 0.8–1.0 kWh, a 15%–40% reduction compared with baseline models: Random Forest and XGBoost. Performance was stable across climates: the lowest errors appeared in London's temperate conditions, while slightly higher deviations occurred in the more extreme climates of Tehran and New York. At the zone level, accuracy was strongest in low‐load areas, such as corridors and stairwells, while offices with greater internal gains showed modest deviations during peak usage. Residual analysis indicated that most predictions deviated within ± 1.5 kWh, with only rare larger errors under severe weather. Sensitivity tests confirmed that outdoor temperature, solar radiation, and internal loads strongly affected predictions; for example, a 1°C rise in outdoor temperature increased cooling demand by about 5%–8%. Overall, this study demonstrates that GNN‐based models are fast, accurate, and scalable surrogates for building energy simulation. By capturing inter‐zone thermal interactions, they bridge physics‐based and data‐driven approaches. Future work should emphasize real building data, HVAC system integration, and improved interpretability for reliable, climate‐resilient energy management.