Database INMET's for 54 conventional stations in Brazil

Wind speed data are collected from INMET's network of conventional stations and made available by the project “Meteorological Database for Teaching and Research (BDMEP)” ( https://bdmep.inmet.gov.br/ ). Originally, data were collected at 3 synoptic times (0000, 1200, and 1800 UTC) and then averages of these three times across the database were calculated. Therefore, the temporal sampling of the analyzed series is daily, and the period of analysis was 60 years, starting on January 1, 1961, and ending on December 31, 2020.
Among the existing multiple imputation techniques, the bootstrap expectation maximization (EM) algorithm was chosen, which imputes the missing observations by estimates generated by the EM algorithms implemented in several bootstrap samples of the data. To assess the quality of the imputed data, the data set from each station was divided into 3 groups according to the percentage of missing data: Group 1 (0-10%); Group 2 (11-20%); Group 3 (21-30%). The evaluation of the imputation method was carried out through descriptive analysis, using the boxplot graph and measures such as minimum, maximum, first and third quartile, median, mean and standard deviation.
To verify equality between the means and medians of the imputed faulty data, Student's t tests and Wilcoxon were used, respectively. When the p-value associated with each test was greater than the significance level, the equality hypothesis was rejected. Both tests were performed at a 5% significance level. For the analysis of a comparison between the two climatologies, only the data referring to the meteorological stations in which the statistical tests detected equality of means and medians between the imputed and faulty data were used.
We applied a filter to exclude time series with a number of failures greater than 30%, leaving, therefore, 54 weather stations. The states of Acre (AC), Rondônia (RO), Amapá (AP), Mato Grosso do Sul (MS), Alagoas (AL), Espírito Santo (ES), Piaui (PI), Pernambuco (PE), Alagoas (AL) and Santa Catarina (SC) do not have meteorological stations that passed the criteria established in the study.

风速数据来自巴西国家气象研究所（INMET）的常规气象站网络，并由“教学与研究气象数据库（Meteorological Database for Teaching and Research，BDMEP）”项目（https://bdmep.inmet.gov.br/）提供。原始数据在3个天气观测时次（synoptic times）（协调世界时00:00、12:00和18:00）采集，随后计算数据库中这三个时次的平均值。因此，分析序列的时间采样（temporal sampling）为日尺度，分析时段为60年，从1961年1月1日至2020年12月31日。
在现有多重插补技术（multiple imputation techniques）中，选择了自助期望最大化（EM）算法（bootstrap expectation maximization (EM) algorithm），该算法通过在数据的多个自助样本中实现的EM算法生成的估计值来插补缺失观测值。为评估插补数据的质量，每个站点的数据集根据缺失数据百分比分为3组：第1组（0-10%）、第2组（11-20%）、第3组（21-30%）。插补方法的评估通过描述性分析进行，使用箱线图（boxplot）及最小值、最大值、第一四分位数、第三四分位数、中位数、均值和标准差等统计量。
为验证插补数据与故障数据之间的均值和中位数是否相等，分别采用Student t检验（Student's t test）和Wilcoxon检验（Wilcoxon test）。当每个检验的p值（p-value）大于显著性水平（significance level）时，拒绝相等假设。两项检验均在5%的显著性水平下进行。在两种气候态（climatologies）的比较分析中，仅使用那些统计检验检测到插补数据与故障数据之间均值和中位数相等的气象站数据。
我们应用过滤器排除故障数超过30%的时间序列，最终保留54个气象站。阿克里州（AC）、朗多尼亚州（RO）、阿马帕州（AP）、南马托格罗索州（MS）、阿拉戈斯州（AL）、圣埃斯皮里图州（ES）、皮奥伊州（PI）、伯南布哥州（PE）、阿拉戈斯州（AL）和圣卡塔琳娜州（SC）没有符合本研究设定标准的气象站。