Learning Urban Climate Dynamics via Physics-Guided Interactions

Core Innovations

• Physics-Guided: Embedding domain knowledge (e.g., Clausius-Clapeyron equation) into deep learning architectures.
• Multi-Task Learning: Joint estimation of 2-m air temperature (T), specific humidity (q), and dew point temperature (td).
• Transformer Architecture: Capturing nonlinear dynamic changes and long-term dependencies in urban climates.
• Physical Consistency: Avoiding physically impossible predictions common in black-box models.

Abstract

Beyond "Black-Box" Prediction

Urban climate prediction typically relies on slow numerical models or data-driven machine learning (e.g., LSTMs). While the latter is faster, it often ignores physical laws, leading to results that lack explainability or physical consistency.

This study proposes UCformer, a novel multi-task, physics-guided Transformer architecture. It not only learns data mappings but also enforces physical priors to constrain the learning process. Experiments show that UCformer excels in predicting urban air temperature, humidity, and dew point, outperforming standard Transformers and other baselines by unifying accuracy with physical rationality.

Model Architecture

UCformer

Combines Attention mechanisms with physics-constrained loss functions to ensure the model captures temporal dynamics while adhering to thermodynamic laws.

Validation

State-of-the-Art

Tests across multiple urban datasets demonstrate that UCformer not only achieves lower error rates but also exhibits more robust performance under extreme weather conditions.