UrbanGraph: Physics-Informed Spatio-Temporal Dynamic Heterogeneous Graphs for Urban Microclimate Prediction

TL;DR

UrbanGraph introduces a physics-informed dynamic heterogeneous graph framework to predict urban microclimates by encoding multiple physical processes as time-varying edges and employing a spatio-temporal heterogeneous GNN. The method combines a physics-informed graph representation with a three-layer RGCN and LSTM-based temporal evolution to achieve accurate, efficient predictions across six microclimate targets on a high-fidelity ENVI-met–generated dataset. Key contributions include five edge types (semantic, internal contiguity, shading, vegetation evapotranspiration, convective diffusion), a dynamic heterogeneous adjacency, and a fusion-based prediction pipeline that attains state-of-the-art performance while reducing FLOPs. The work provides a new high-resolution benchmark (UMC4/12) for spatio-temporal urban modeling and suggests avenues for adaptive graph learning to further enhance physical fidelity and scalability.

Abstract

With rapid urbanization, predicting urban microclimates has become critical, as it affects building energy demand and public health risks. However, existing generative and homogeneous graph approaches fall short in capturing physical consistency, spatial dependencies, and temporal variability. To address this, we introduce UrbanGraph, a physics-informed framework integrating heterogeneous and dynamic spatio-temporal graphs. It encodes key physical processes -- vegetation evapotranspiration, shading, and convective diffusion -- while modeling complex spatial dependencies among diverse urban entities and their temporal evolution. We evaluate UrbanGraph on UMC4/12, a physics-based simulation dataset covering diverse urban configurations and climates. Results show that UrbanGraph improves $R^2$ by up to 10.8% and reduces FLOPs by 17.0% over all baselines, with heterogeneous and dynamic graphs contributing 3.5% and 7.1% gains. Our dataset provides the first high-resolution benchmark for spatio-temporal microclimate modeling, and our method extends to broader urban heterogeneous dynamic computing tasks.

Figures (13)

• Figure 1: The UrbanGraph framework. (a) Overall pipeline: Geospatial data is converted into a graph structure, which is processed by a spatio-temporal GNN with climate and time characteristics to generate high-resolution predictions. (b) The physics-informed dynamic graph concept: Nodes represent urban entities, while edges—representing physical interactions like building shading and convective diffusion—are dynamically reconfigured over time to reflect changing environmental conditions.
• Figure 2: An illustrative overview of the five edge types used in our graph representation. Dynamic edges are derived from physical processes like shadowing and wind, while static edges are based on spatial proximity, feature similarity, and object integrity.
• Figure 3: The overall architecture of the UrbanGraph model.(a) The end-to-end framework, which processes historical time-series data, weather context, and a sequence of dynamic heterogeneous graphs. At each timestep, an RGCN Block extracts spatial features from the corresponding graph, which are then fused with temporal features by an MLP Fusion Layer. An LSTM layer propagates the temporal state to the next timestep.(b) The detailed structure of the RGCN Block, which aggregates messages from neighbors across different relation types ($R1-R5$) and combines them with the node's self-features.(c) The architecture of a standard LSTM layer used for capturing temporal dependencies.
• Figure 4: Model performance analysis. (a) Training and validation R² convergence curves for the UrbanGraph model. Shaded areas represent the confidence interval. (b) Hour-by-hour R² comparison between UrbanGraph and baselines on the test set, with error bars indicating standard deviation.
• Figure A1: Visualization of the input data for two sample sites from the UMC4/12 dataset. The top row shows a dense, mixed-use urban area, while the bottom row depicts a residential area with more green space. Each column represents a different data layer: (left) Tree Height, (middle) Land Cover Type, and (right) Building Height.