In this article we investigate the influence of meteorological data selection for the training of a neural network that predicts wind patterns. Different sets of meteorological information, distinct regions of the globe, various sizes for the observed area and different time windows are considered both for training and evaluation of the predicted winds. NCEP reanalysis 2 data was used to feed the neural network, which was based on a spatio-temporal architecture considering convolutions and recurrences. Besides achieving acceptable quality wind speed forecasts, the results for the contrasting cases highlight an integrative methodology for data selection when exploring machine learning in the geosciences.

在本文中，我们研究了气象数据选择对预测风模式的神经网络的训练的影响。不同的气象信息集、全球的不同区域、观测区域的不同大小和不同的时间窗口都被考虑用于预测风的训练和评估。NCEP 再分析 2 数据用于馈送神经网络，该网络基于考虑卷积和递归的时空架构。除了实现可接受的质量风速预测外，对比案例的结果还突出了在地球科学中探索机器学习时数据选择的综合方法。