Predicting low-occurrence wind speed at the pedestrian levels of simplified arrays: Artificial neural network models versus statistical methods

Abstract

Low-occurrence wind speed in urban areas has a profound impact on the daily activities of people. Although several statistical methods were developed for estimating the low-occurrence wind speeds, their accuracy and flexibility still need improvements. Based on the inference from the favorable outcomes of statistical models in which low-occurrence wind speeds can be estimated by the statistics including high-order moments, in this study, the artificial neural network (ANN) models were designed to estimate the low-occurrence wind speeds based on the moments. After the sensitivity tests for the dimensions, hidden layers, epochs and training data ratios of the ANN models, the optimal setting was applied to predict the low-occurrence wind speeds of two building array cases including square and staggered layouts. The applicability of the ANN models was also evaluated for the extrapolation method between two building array cases and compared with the traditional statistical methods. It was found that the ANN models with the higher-order statistic input showed better performance than those with the second-order statistic input. The results obtained from the interpolation method are more accurate than those from the extrapolation method. Compared with traditional statistical methods, the ANN models show greater superiority in the prediction accuracy. According to the comprehensive evaluation of the ANN models, although they require ample training datasets connecting the statistics and low-occurrence wind speeds, it is proved to be a promising approach for estimating the low-occurrence wind speeds in urban wind environment.