Wind-Induced Pressure Prediction on Tall Buildings Using Generative Adversarial Imputation Network

Area of Science:

• Structural Engineering
• Computational Fluid Dynamics
• Data Science

Background:

• Wind tunnel testing is crucial for assessing building wind loads and ensuring structural safety.
• Pressure sensor failures during wind tunnel tests lead to data loss, hindering accurate cladding pressure assessment.
• Existing data imputation methods struggle with accurate predictions for multiple missing time-series data points.

Purpose of the Study:

• To investigate the efficacy of machine learning for predicting missing wind pressure data in tall buildings.
• To propose and validate a generative adversarial imputation network (GAIN) for reconstructing time-series pressure coefficient data.
• To compare GAIN's performance against traditional imputation techniques like K-nearest neighbor and multiple imputations by chained equations.

Main Methods:

• Utilizing machine learning, specifically a generative adversarial imputation network (GAIN), for data imputation.
• Training and validating the GAIN model on wind pressure data from tall buildings.
• Comparative analysis of GAIN against K-nearest neighbor and multiple imputations by chained equations.

Main Results:

• The GAIN model demonstrated superior performance in predicting missing wind pressure coefficients.
• GAIN achieved the minimum average mean-squared error (0.016) and maximum average R-squared error (0.961).
• The model exhibited minimum average variance and standard deviation, indicating accurate and stable predictions.

Conclusions:

• Machine learning, particularly GAIN, offers a robust solution for imputing lost wind pressure sensor data.
• Accurate data imputation enhances the reliability of wind loading assessments for tall buildings.
• The proposed method contributes to ensuring structural integrity and longevity under wind environmental conditions.