AbstractIn determining wind loads when designing buildings—especially high-rise buildings—wind speed is a critical factor. With increased progression of city high-rise construction, accurate prediction has become ever more important. Many factors, such as local climates, roughness of surrounding conditions, and terrain, may affect wind speed. But among all factors considered, terrain has a very large influence. Difficulties in reflecting the influence of terrain on wind speed include (1) collecting topographic data information; (2) quantitatively evaluating terrain features; and (3) expressing specific area relationships with surrounding observation stations. This paper poses means to solve these limitations by (1) using satellite imagery to gather topographic information; (2) evaluating terrain quantitatively with convolutional neural network (CNN)–based encoders; and (3) employing a graph neural network (GNN) to estimate the wind speed relationship of an arbitrary place to that of observation stations. Herein, a machine learning model utilizing the aforementioned means was proposed. To support this model, an experiment was conducted using observed wind speed data in Korea. Throughout the experiment, previous studies were used to compare results with ratification of the posed model. Given the submitted findings, further research and investigation to establish correlation with other experimental case results are proposed.IntroductionDue to urbanization, an increase in the number of high-rise buildings constructed is projected. Given their vulnerability to wind loads, precise determination becomes increasingly important. As is, wind load is determined from the basic wind speed of a site and varies with regional climate. Because wind velocity pressure is proportional to the square of wind speed, small discrepancies or errors in determining basic wind speed may result in large differences in wind load. When considering gust effects and vortex-induced vibrations, these differences can be even greater (Bahmani et al. 2014; Bernardini et al. 2015; Kwon et al. 2015; Lou et al. 2015; Alinejad and Kang 2020; Ouyang and Spence 2020; Cui and Caracoglia 2020; Jeong et al. 2021). Because vortex shedding is affected by building shape, characteristic length, and wind speed, with a fixed shape and length of target structure, the wind speed is a main governing factor.For tall and slender buildings with rectangular shapes, more eminent vortex shedding can occur due to two-dimensional air flow at the midsection. Tall buildings exceeding 30 stories are governed by across-wind loads induced by vortex shedding (Kang et al. 2019). Including gust effects, vortex-induced wind load, and resonance of the structure, the maximum displacement by along-wind load is proportional to 2.1 power of mean wind speed (or basic wind speed), and that by across-wind (vortex-induced vibration) is proportional to 3.1 power of mean wind speed (Tamura 2020). Underestimation or overestimation of wind loads can be directly related to safety or cost issues even for low-rise buildings and wind turbines in mountains. Thus, accurate determination of basic wind speed is significant in wind engineering.Basic wind speeds in the form of maps or tables are provided in national design codes and standards. These speeds are derived from wind speed data collected at observation stations. The raw data of observed wind speeds involve not only climate characteristics, but also effects of surrounding site conditions (called surface roughness) and topography at the station. The effects of roughness and topography on wind speed profiles and mean wind speeds are explained in Engineering Sciences Data Unit (ESDU) publications ESDU 84011 (ESDU 2012) and ESDU 91043 (ESDU 2000). ESDU 84011 (ESDU 2012) presents wind speed profile modification depending on roughness and varying roughness in upwind site. ESDU 91043 (ESDU 2000) provides simplified and detailed methods to consider the effect of topography. For use in design codes, standardization of terrain effects and modification of the wind speed data is carried out. After modification, an extreme value analysis to estimate the expected maximum wind speed for the target return period is conducted. Accurate estimations require sufficient data recorded over decades. Determination of basic wind speed on a national scale requires significant judgment and effort from engineers. Thus, the results of the process may differ depending on the engineers.One of the most challenging issues for determination of basic wind speed is homogenization of site conditions. In design codes, basic wind speed is defined as mean wind speed (averaging time varies with design codes) at 10 m above the ground in open terrain. Thus, wind speed data from observation stations should be modified for homogenized site conditions. There are code-specified regulations for modification of wind speed considering neighboring surface roughness and topography. However, it is difficult to apply the regulations to real-world circumstances due to its complexity. At this stage, judgment for classification of surface roughness or determination of topography factor can vary depending on engineers.Additionally, selection of data is important. The traditional method uses wind speed data from the nearest observatory station regardless of terrain features. However, modification equations to account for terrain features may not be accurate in three-dimensional complicated terrains. Referring to wind speed data from stations with similar terrain features can result in better estimation (Lee et al. 2022).To address the aforementioned issues, this study attempts to improve the method of determination of basic wind speed using machine learning. Terrain features from satellite images are evaluated by convolutional neural network (CNN)–based encoders, and particularly, graph neural network (GNN) is used to express the relationship between stations.Existing Research on Prediction of Wind Speed and Its LimitationsAs mentioned previously, wind speed is significantly affected by surrounding site conditions, such as surface roughness and topography. Most prior studies were conducted to figure out the effect of topography, for which wind tunnel tests and computational fluid dynamics (CFD) analysis were commonly used (Bowen and Lindley 1976; Carpenter and Locke 1999; Wang et al. 2014; Wani et al. 2021). However, the research was mainly focused on two-dimensional (2D) hill or escarpment effects. For more complex conditions combined with surface roughness, Abdi and Bitsuamlak (2014) conducted CFD analysis to examine the effect of roughness change; however, CFD analysis has limitation of requiring a large amount of computational time and cost.Research comparing national basic wind speeds has been conducted in the past. Almaawali et al. (2008) compared basic wind speeds in Oman determined by Gumbel and Gringorten methods. Lakshmanan et al. (2009) updated India’s basic wind speed map by using long-term hourly wind data. However, research conducted by Almaawali et al. (2008) and Lakshmanan et al. (2009) did not mention the standardization of terrain conditions in modification of wind speed data. On the other hand, Ballio et al. (1999) presented basic wind speeds in Italy based on the distance from the sea and altitude. Jeong et al. (2014) suggested an effective height concept to modify wind speed data for standardized terrain conditions. Although terrain conditions were considered in these studies, how complex three-dimensional (3D) terrain in the real world was quantified and classified by authors is not clearly known.For problems difficult to quantify or solve by human intuition, artificial intelligence has begun to be used. Especially, machine learning is rapidly developed and applied to various fields. Koo et al. (2015) adopted the method of artificial neural networks to predict wind speed based on distance according to the type of topography and discussed the relationship between them. By studying each topography type, Koo et al. (2015) indirectly confirmed its importance when predicting wind speed. This study had the following limitations: (1) wind speed data were categorized roughly into three types of coast, plain, and mountain; (2) the data used were only numerical data in terms of distance between two points, observed wind speed data, etc.; and (3) only 6 years of the observed data were used.Lee et al. (2022) proposed the method of evaluating and predicting topography similarity using satellite images and a CNN-based neural network. The model suggested by Lee et al. (2022) did not acquire a direct relationship between wind speed and topography. The topography features of satellite images were learned through a CNN, and the 100-year return period wind speed of a certain region was predicted. The model adopted a two-staged method that first extracts and learns topographic features of images using autoencoder (AE) to distinguish topography features from satellite images and estimate wind speeds at specific locations by linearly combining learned topographic information with location information of the station. In this study, the selection of observation station was significant for accuracy, and three methods to select a reference station for basic wind speed prediction were proposed as follows: (1) selecting the nearest station from the target site (select one); (2) using wind speed data from nearby k-number of stations [k-nearest neighbor (k-NN)]; and (3) selecting k-number of stations considering distance and terrain similarity. The method of simply utilizing the data of the nearest station showed a considerably large error, and the method of utilizing the data of the nearest k-number of stations increased in accuracy as k increased, but still showed an error of 3.26 to 3.47  m/s. In case that terrain similarity was additionally considered, the predicted basic wind speed (average 3.18  m/s) showed higher accuracy than that of the previous two methods.Because the aforementioned methods are two-staged methods using CNN, they require user intervention. Lee et al. (2022) also proposed an end-to-end method that treats from input to output as a single neural network without a segmented network, and to this end, multilayer perceptron (MLP) was used. The predicted wind speed through the MLP method was compared with observed data, and the accuracy was significantly higher than that of all other previous methods (average 2.92  m/s). Therefore, it is considered that the end-to-end method using a single neural network is more efficient for predicting the basic wind speed using satellite images. However, because the MLP method only derives one value of 100-year return period basic wind speed, limitation exists in that multiyear data of specific location cannot be predicted.This study proposes an improved end-to-end model that complements the aforementioned limit. The main characteristics of the model proposed in this study are as follows: •The process of extracting features from satellite images to the process of predicting wind speed is a single deep learning model (end-to-end model).•Graphical representation of observation stations are in the form of unstructured data using GNN.•Wind speed records for the last 19 years were used for estimation in lieu of a basic wind speed based on the probability model.•Multiyear annual maximum wind speed can be predicted.Limitations of Existing Basic Wind Speed Determination MethodExisting Basic Wind Speed Determination ProcessCurrent design codes and standards reference wind speed for the region to be used in all site conditions as the basic wind speed. It is standardized for specific site conditions. In design codes and standards such as ASCE 7-22 (ASCE 2022), Architectural Institute of Japan (AIJ) AIJ-RLB (AIJ 2015), Eurocode 1 (CEN 2003), Korean Design Standard [KDS 41:2019 (MOLIT 2019)], and ISO 4354 (ISO 2009), basic wind speed is defined as wind speed at 10 m above the ground in flat open terrain. Because design codes use different averaging times and return periods, the basic wind speed determined varies depending on the code or standard referenced.The general process for determining basic wind speed is shown in Fig. 1. Modification of wind speed for site conditions is essential. Although it is desirable to build observation stations at locations not influenced by surface roughness and topography, inevitable mountainous areas and cities conjoined with tall buildings need to be addressed. For these cases, engineering judgment is used in determination of classification, and wind speeds are modified based on empirical equations for roughness categories.In the Commentary of ASCE 7-22, schematic explanations of combined conditions of roughness categories were presented as shown in Fig. 2. However, the spatial length conditions provided are strict and do not necessarily represent real-world circumstances. Roughness conditions in the real-world can vary with azimuth and distance. In the Commentary of Korean Building Code [KBC 2016 (MOLIT 2016)] (previous version of KDS 41:2019), roughness categories are determined from a 45° fanwise upwind surface up to 40H or 3 km. If roughness conditions are combined, the conservative design would require the utilization of a smoother surface roughness category.In addition to surrounding surface effects, the site height of observatory stations needs also to be considered given that hills or escarpments may increase wind speed. Design codes and standards typically employ a topography factor to account for these effects. However, difficulty exists in applied determination because the land may be too wide, continuous, and complicated. As a result, wind speed modification by an effective height was used in KDS 41:2019 based on research conducted by Jeong et al. (2014). Effective height Ze is defined by the following equation: (1) Ze=(Ha−H¯a)−(H¯−Z0k)where Ha = height from sea level of observatory station; H¯a = mean height from sea level to surrounding area; H¯ = mean height of buildings in surrounding area; Z0 = roughness length; and k = Karman constant ≒ 0.4.Wind speed roughness categories can be converted to that of roughness Category C (flat open terrain) using effective height and the wind speed profile factor Kzr. By definition, a wind speed profile is the ratio of wind speed at target height and target roughness to wind speed at 10 m height and roughness Category C. Wind speed above the atmospheric boundary layer is assumed to be unaffected and constant regardless of surface roughness. Thus, observed wind speed at an effective height V(Ze) can be converted to wind speed at 10 m height and roughness Category C, per equation VC,10. Derivation was provided by Jeong et al. (2014) (2) VC,10=V(Ze)Kzr=V(Ze)1.67(ZgZe)αwhere V(Ze) = wind speed at effective height; and Zg = gradient height.ASCE 7-22 and KDS 41:2019 classify the surface roughness into four categories of A to D, and the description for each category is similar (recently, ASCE has excluded Category A). However, unlike KDS, which uses a 10-min average wind speed, the ASCE method is based on a 3-s gust (or 3-s average) wind speed, and thus it is difficult to apply Eq. (2) itself. Therefore, the preceding equation for 10-min average wind speed should be converted to 3-s gust wind speed. Although there is much research regarding the relationship between the 3-s gust and 10-min average wind speed, it is reasonable to refer to Kim and Ha (2015), who proposed a gust factor that converts the 10-min average wind speed to 3-s gust wind speed because Kim and Ha (2015) used the same database of the Korea Meteorological Administration (KMA).In this paper, the gust factor was proposed based on wind speed data observed throughout the Korean peninsula for 40 years, and the factor on surface roughness Category C as shown in Eq. (3). Therefore, according to ASCE, the wind speed at 10 m above the ground at the surface roughness Category C can be calculated by multiplying G by VC,10 calculated in Eq. (3) (3) G=V3V600={1.691(V600<16  m/s)−0.0156V600+1.9428(16  m/s

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *