A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Neural-network based wind pressure prediction for low-rise buildings with genetic algorithm and Bayesian optimization
Highlights Two novel approaches, genetic algorithm-based neural network (GANN) and Bayesian optimization-based neural network (BONN), are proposed to improve the efficiency of NN for wind pressure predictions. The influences of hyperparameters, the number of data pairs, and the NN structures on the prediction performances of NN for building dimensions and roof slopes are evaluated. The proposed two approaches are adopted in the wind pressure coefficient predictions to achieve much higher efficiencies.
Abstract Low-rise wood buildings, which occupy over 95% of all residential structures in the U.S., suffer from wind hazards, such as hurricanes, winter storms, and tornadoes. Realistic wind pressures are critical to obtaining the actual responses of buildings for structural design and damage prediction. Wind tunnel experiments provide realistic wind pressure distributions on buildings for structural analysis. However, it is time-consuming and costly to experiment with every building under different wind scenarios. An effective alternative way is to use the neural network (NN) to predict the wind pressures for buildings with different geometries and under various wind directions based on the available databases. However, this method could have a high computational cost due to the challenges of determining an optimal NN structure with suitable hyperparameters. In the present study, the influences of the hyperparameters (namely the input variable and transfer function), the number of data pairs, and the NN structures on the performances of NNs are evaluated when predicting wind pressure coefficients based on building dimensions and roof slopes. To overcome the tedious trial-and-error procedure for optimal NN selection, two new approaches, genetic algorithm-based neural network (GANN) and Bayesian optimization-based neural network (BONN), are proposed. The BONN is demonstrated to be the most efficient, saving 88–94% computational time compared with the traditional trial-and-error neural network (TENN), while the GANN costs about twice the time of BONN. The much higher efficiencies of the proposed two approaches in wind pressure predictions help broaden the application of NN.
Neural-network based wind pressure prediction for low-rise buildings with genetic algorithm and Bayesian optimization
Highlights Two novel approaches, genetic algorithm-based neural network (GANN) and Bayesian optimization-based neural network (BONN), are proposed to improve the efficiency of NN for wind pressure predictions. The influences of hyperparameters, the number of data pairs, and the NN structures on the prediction performances of NN for building dimensions and roof slopes are evaluated. The proposed two approaches are adopted in the wind pressure coefficient predictions to achieve much higher efficiencies.
Abstract Low-rise wood buildings, which occupy over 95% of all residential structures in the U.S., suffer from wind hazards, such as hurricanes, winter storms, and tornadoes. Realistic wind pressures are critical to obtaining the actual responses of buildings for structural design and damage prediction. Wind tunnel experiments provide realistic wind pressure distributions on buildings for structural analysis. However, it is time-consuming and costly to experiment with every building under different wind scenarios. An effective alternative way is to use the neural network (NN) to predict the wind pressures for buildings with different geometries and under various wind directions based on the available databases. However, this method could have a high computational cost due to the challenges of determining an optimal NN structure with suitable hyperparameters. In the present study, the influences of the hyperparameters (namely the input variable and transfer function), the number of data pairs, and the NN structures on the performances of NNs are evaluated when predicting wind pressure coefficients based on building dimensions and roof slopes. To overcome the tedious trial-and-error procedure for optimal NN selection, two new approaches, genetic algorithm-based neural network (GANN) and Bayesian optimization-based neural network (BONN), are proposed. The BONN is demonstrated to be the most efficient, saving 88–94% computational time compared with the traditional trial-and-error neural network (TENN), while the GANN costs about twice the time of BONN. The much higher efficiencies of the proposed two approaches in wind pressure predictions help broaden the application of NN.
Neural-network based wind pressure prediction for low-rise buildings with genetic algorithm and Bayesian optimization
Ding, Zhixia (author) / Zhang, Wei (author) / Zhu, Dongping (author)
Engineering Structures ; 260
2022-03-26
Article (Journal)
Electronic Resource
English
Artificial Neural Network-Based Prediction of Wind Pressure Coefficients on Buildings
| Springer Verlag | 2021
Machine learning based algorithms for wind pressure prediction of high-rise buildings
| SAGE Publications | 2022