A Pattern-constrained deep learning model for urban canopy turbulence reconstruction from sparse sensor data

Yong Cao, Peixing Xie, Guoshuo Huang, Wei Wang, Wen-Li Chen, Gang Hu, Shuyang Cao,
A Pattern-constrained deep learning model for urban canopy turbulence reconstruction from sparse sensor data,
Building and Environment,
Volume 285, Part C,
2025,
113535,
ISSN 0360-1323,
https://doi.org/10.1016/j.buildenv.2025.113535.
(https://www.sciencedirect.com/science/article/pii/S036013232501008X)
Abstract: Traditional numerical simulations for estimating the pedestrian-level wind environments are often computationally intensive and time-consuming. While machine learning-based reconstruction offers potential for efficient, high-fidelity estimation using sparse sensor data, challenges persist due to chaotic turbulence, architectural heterogeneity, and sensor sparsity. This study proposes a pattern-constrained generative adversarial network (PCG) to reconstruct instantaneous urban wind fields from sparse, flexibly distributed sensor measurements. The PCG framework integrates a contrastive learning-driven Flow Pattern Extraction Module (FPEM) that encodes aperiodic physical features, enabling the generative model to capture multi-scale turbulent structures. Validated on high-fidelity large-eddy simulation data from Niigata City, the PCG demonstrates significant improvements over baseline models, achieving >15% improvement in most metrics (MSE, R2 and LPIPS). Key innovations include the capacity of FPEM to distill monotonic latent representations from chaotic flow dynamics and the adaptability to varying sensor numbers (16–48 sensors). Results show that the PCG maintains robust reconstruction accuracy even with non-optimal sensor configurations, outperforming conventional approaches in both numerical accuracy and spatial fidelity of flow patterns.
Keywords: Pedestrian-level wind environment; Urban canopy; Deep learning reconstruction; Sparse sensor data