A Large Graph Dataset for Building Energy Modeling

Summary & Methodology Analysis

The researchers addressed the computational inefficiency of traditional physics-based building performance simulation by developing ArchEGraph, a dataset designed for graph-based machine learning models. The pipeline standardizes geometry by converting polygonal meshes into heterogeneous graphs, utilizing minimal convex partitioning to ensure the building topology is compatible with graph-based spatial operators. This structured approach allows models to capture structural inductive biases, which are necessary for generalizing across various architectural layouts and climate conditions. The dataset maps designs against 64 representative global cities using batch simulations via EnergyPlus to generate aligned zone-level thermal loads. By providing these ground-truth labels, the researchers enable surrogate modeling that avoids the high latency associated with iterative physics-based solvers. Two core tasks are defined: M2G for reconstructing mesh designs and G2E for predicting energy requirements. Performance metrics confirm that specialized graph architectures, specifically F2S-TransConv and F2S-GPS, achieve the highest accuracy for these tasks depending on the dataset subset. However, developers should note that the current data is restricted to office building typologies and relies entirely on simulated data rather than real-world measurement. This reliance on simulation means the models are bound by the fidelity of the underlying EnergyPlus physics models used to generate the dataset.