<![CDATA[Fog is one of the most hazardous weather phenomena for aviation operations. Dense fog can reduce visibility to below 1,000 meters, potentially causing flight delays, cancellations, and even aviation incidents. To date, fog prediction, particularly at Budiarto Airport, still relies on manual analysis by weather forecasters, making it prone to subjectivity and delays in information delivery. This study proposes and compares two deep learning architectures: the Gated Recurrent Unit (GRU) as an efficient recurrent model, and Time Series Transformer (TST) based on self-attention as a state-of-the-art model for METAR (Meteorological Aerodrome Report) data-based fog event prediction. The METAR data is initially processed using a sliding window technique before becoming a ready-to-use dataset. The dataset comprises 153,838 METAR records from the Budiarto–Curug Meteorological Station spanning from September 2015 to February 2026, which were processed through a METAR code parsing pipeline, BMKG rule-based median imputation, Min-Max normalization, and the construction of a 9-1 sliding window dataset. Experimental results on the test data demonstrate that TST 9-1 delivers the best performance with a Root Mean Squared Error (RMSE) of 0.452427 and a three class classification accuracy (No Fog, Light Fog, Dense Fog) of 88.21%, significantly outperforming GRU 9-1, which achieved an RMSE of 0.883981 and an accuracy of 72.97%. The main novelty of this research lies in the comparative study of GRU and TST architectures for METAR based fog prediction at airports, combined with a sliding window technique and the conversion of visibility regression into a multi class classification of fog events. This research contributes a fog prediction modeling framework capable of processing time-series data sequentially and more effectively, which can serve as a foundation for the development of an accurate, automated early warning system for fog events in airport environments.]]>
