Conditional Diffusion Model to Predict Structural Wind Loading Using Data from Sparse Sensors
Abstract
Extreme local wind pressure causes damage to the envelopes and claddings of high-rise buildings. Accurate prediction of pressure time-series benefits the wind-resistance design and disaster precaution. The conditional diffusion model (CDM) is introduced to reconstruct time-resolved surface pressure fields from sparse measurements. A wind-tunnel pressure time-series from a high-rise building model instrumented with 400 taps is used for training and testing. Three incidence angles (0°, 15°, and 45°) are examined. The model performance is quantified by statistical indicators like root-mean-square error and determination coefficient R 2 as well as wind loading factors like aerodynamic coefficients, pressure coefficients, instantaneous loading patterns, and probability density distributions. CDM outperforms generative adversarial networks at training tap numbers 1 to 100. CDM has a lower requirement of training data. Using 10 and 20 training taps out of 400 taps, R 2 exceeds 0.9 and 0.95, respectively. When further increasing the training tap number to 40 and 60, the overall errors change slightly, while the differences in model performance at various vertical levels and inflow incidences are reduced. CDM also reproduces the dominant unsteady behavior of aerodynamic coefficients and instantaneous pressure patterns but remains mildly conservative for localized extremes using very few sensors.
