<p>With the advancement of AI, the importance of data-driven modeling in the meteorological field has grown significantly. However, observational data obtained solely from ground-based measurement instruments such as AWS, which provide only a limited set of variables, have inherent limitations in predicting complex weather phenomena. These limitations become more pronounced for events like easterly winds, where high-altitude atmospheric variables exert a strong influence. Therefore, model-generated data that include large-scale variables and upper-atmosphere pressure information are highly useful, but their massive volume poses accessibility challenges for individual researchers. In this study, we propose an approach that leverages the advantages of large-scale, vertically layered variables by using ERA5 data, while reducing them to a core set of essential variables for computational feasibility. Based on this framework, we develop a deep learning–based model for predicting easterly winds, which are closely related to natural disasters. Specifically, we propose a cascade two-stage LSTM method that incorporates both the inverted configuration of input data to capture vertical inter-layer relationships and the sequential characteristics across layers. As a prototype study, we conducted experiments on East Asia, including Korea. Comparative evaluations against LSTM-FC and CNN demonstrate the superiority of our method.</p>