Urban airflow field prediction using sparse sensors: A hybrid data-driven approach with reduced-order modeling

Urban airflow field modeling can provide cities with monitoring solutions to enhance pedestrian/cyclists comfort, improve air quality, and enhance urban wind energy generation yield. In addition, it could support a safe and efficient deployment of urban air mobility (UAM) supporting the next generation of faster and more sustainable means of urban transport reducing the congestion and traffic. However, real-time wind monitoring across urban areas remains a significant challenge due to spatial and temporal variations and uncertainties. This study presents an integrated framework that serves as a surrogate model for the real-time reconstruction and prediction of wind velocity fields around a simplified urban building configuration. The approach relies exclusively on online sensor measurements and employs reduced-order modeling techniques for dimensionality reduction, thereby improving computational efficiency. The offline development of the model integrates three principal components: (1) a bidirectional long short-term memory (BILSTM) network trained on airflow sensor data to capture temporal dynamics, (2) a sparsity-promoting dynamic mode decomposition (SP-DMD) for extracting dominant spatial modes of airflow, and (3) a multi-task learning (MTL) network to map sensor signals. The proposed framework was assessed against high-fidelity CFD simulations of wind flow around a high-rise urban building. The results indicate that the proposed three-stage method reliably reconstructs accurate urban airflow fields in real time under varying levels of sensor noise, and demonstrate its feasibility for sparse-sensor reconstruction under the tested sensor-placement configurations. Moreover, these findings demonstrate that the proposed approach can significantly reduce operational/computational costs and shows strong potential as a surrogate tool for next-generation of urban airflow monitoring.