A hybrid machine learning framework for wind pressure prediction on buildings with constrained sensor networks

Accurate and efficient prediction of wind pressure distributions on high-rise building façades is crucial for mitigating structural risks in urban environments. Conventional approaches rely on extensive sensor networks, often hindered by cost, accessibility, and architectural limitations. This study proposes a novel hybrid machine learning (ML) framework that reconstructs high-fidelity wind pressure (HFWP) coefficient fields from a limited number of sensors by leveraging dynamic spatiotemporal feature extraction and mapping. The methodology consists of four key stages: (1) low-fidelity pressure field reconstruction from limited sensor data using constrained QR decomposition, (2) dimensionality reduction of both low-fidelity wind pressure and HFWP reconstructions to extract dominant spatiotemporal features, (3) dynamic mapping of the reduced-order representations using a long short-term memory network, and (4) prediction of the high-fidelity pressure field reconstruction over time. The proposed approach, which predicts the time history of high-fidelity pressure coefficients for various wind directions, is validated using wind tunnel data, with case studies on multiple façades—including the windward, right-side, and leeward surfaces—under various constrained sensor placement scenarios. The proposed methodology is also evaluated against alternative ML models, demonstrating superior accuracy in reconstructing the full pressure field. The results highlight the robustness and generalization capability of the model across different wind directions and sensor configurations, making it a practical solution for real-time wind pressure estimation in structural health monitoring and digital twin applications.