TLDR: UrbanGraph is a novel framework that uses dynamic, heterogeneous graphs informed by physics to predict urban microclimates with greater accuracy and efficiency. It addresses limitations of existing models by explicitly encoding physical processes like shading and evapotranspiration, and modeling complex, time-varying spatial dependencies among urban elements. Evaluated on a new high-resolution dataset, UrbanGraph improves prediction accuracy by up to 10.8% and reduces computational cost by 17.0% over baselines, demonstrating significant advancements in urban climate forecasting.
Predicting the climate within cities, known as urban microclimate prediction, is becoming increasingly important. This is because it directly impacts how much energy buildings consume and the health risks faced by the public. However, current methods often struggle to accurately capture the complex physical interactions, spatial relationships, and changing conditions over time that define a city’s climate.
Traditional approaches, like detailed physics-based simulations (e.g., Computational Fluid Dynamics or CFD), are very accurate but require immense computing power, making them impractical for large-scale, long-term predictions. Data-driven methods, while more efficient, often simplify the urban environment too much. For instance, some models can only predict conditions at specific points, ignoring how different areas interact. Others use fixed grid structures that can’t effectively capture the irregular shapes and varied elements of a city.
Graph Neural Networks (GNNs) offer a more natural way to model these spatial relationships. Yet, many existing GNNs treat all interactions uniformly, failing to distinguish between different physical processes like the cooling effect of trees versus the shade cast by buildings. They also typically use a fixed graph structure, which can’t represent how physical processes evolve as environmental conditions change throughout the day.
Introducing UrbanGraph
To overcome these limitations, researchers have developed UrbanGraph, a new framework designed for urban microclimate prediction. UrbanGraph is unique because it integrates physics-informed principles with dynamic, heterogeneous spatio-temporal graphs. This means it not only understands the physical laws governing urban climate but also represents the city as a complex network where different types of urban elements (nodes) interact through various types of relationships (edges) that change over time.
The framework explicitly encodes key physical processes: the cooling effect of vegetation through evapotranspiration, the impact of shading from buildings and trees, and the movement of heat and moisture through convective diffusion. By doing so, UrbanGraph can model the intricate spatial dependencies among diverse urban entities and how these interactions evolve temporally.
How UrbanGraph Works
UrbanGraph consists of two main parts: a physics-informed graph representation and a spatio-temporal dynamic relational graph network.
The **physics-informed graph representation** is crucial. It transforms continuous physical fields into a discrete, computationally efficient graph structure. At any given moment, the connections (edges) in this graph are reconstructed based on current environmental conditions. There are five distinct types of relationships:
- **Static Edges:** These represent unchanging spatial and semantic relationships.
- Semantic Similarity Edges: Connects nodes that are functionally similar.
- Internal Contiguity Edges: Links parts within large, continuous urban features like building clusters or tree groups.
- **Dynamic Edges:** These explicitly model time-varying physical processes, with connections updated hourly.
- Shading Edges: Model the cooling effect of shadows, with connections from shading objects (buildings, trees) to ground nodes, considering solar angles and shadow lengths.
- Vegetation Evapotranspiration Edges: Represent the local cooling effect of trees, with a dynamic radius of influence based on solar radiation.
- Convective Diffusion Edges: Simulate wind-driven convection, adjusting the ‘effective distance’ between nodes based on wind speed and direction.
The **spatio-temporal dynamic relational graph network** is designed to process these complex graph structures. It uses a specialized architecture that includes Feature Encoders, a Spatial Graph Encoder (using Relational Graph Convolutional Networks or RGCNs to capture spatial dependencies from heterogeneous edges), a Spatio-Temporal Evolution Module (using Long Short-Term Memory or LSTM networks to model temporal dynamics), and a Prediction Head to generate multi-step predictions.
Also Read:
- Unlocking Urban Solar Potential: AI Framework Assesses Facade PV Installation
- Forecasting Ground Movement: A Multi-Modal Transformer Achieves Unprecedented Accuracy
Performance and Impact
UrbanGraph was evaluated on UMC4/12, a high-fidelity, physics-based simulation dataset covering diverse urban configurations and climates. The results were impressive: UrbanGraph improved prediction accuracy (measured by R²) by up to 10.8% and reduced computational cost (FLOPs) by 17.0% compared to all other baseline models. The study specifically highlighted that the heterogeneous graph mechanism contributed 3.5% to these gains, while the dynamic graph mechanism contributed 7.1%.
The model also demonstrated strong robustness, consistently maintaining low errors throughout a 12-hour prediction period, even during times of significant climate fluctuation. Furthermore, UrbanGraph showed strong generalization capabilities, performing well across various microclimate and thermal comfort variables like Air Temperature (AT), Wind Speed (WS), Relative Humidity (RH), Mean Radiant Temperature (MRT), and Physiological Equivalent Temperature (PET).
This research not only provides a powerful new tool for urban microclimate prediction but also introduces the UMC4/12 dataset as the first high-resolution benchmark in this field, which will help accelerate future research. While UrbanGraph explicitly encodes known physical processes, future work aims to explore adaptive graph learning methods that can automatically discover and optimize graph structures from data, potentially uncovering new urban interaction patterns.
For more detailed information, you can read the full research paper here: URBANGRAPH: PHYSICS-INFORMED SPATIO-TEMPORAL DYNAMIC HETEROGENEOUS GRAPHS FOR URBAN MICROCLIMATE PREDICTION.
