Application of LSTM neural networks with multivariate numerical analysis to aviation wind gust forecasting

This paper presents a long short-term memory (LSTM) framework developed for predicting wind gusts 1 h in advance at Taiwan Taoyuan International Airport (RCTP) during typhoons. Hourly surface observations were collected from 12 landfalling typhoons (2010–2020) and used to compare three feature-selection strategies: Pearson correlation, recursive feature elimination with cross validation, and random-forest importance. Models were trained on 12-h multivariate histories. A leave-one-typhoon-out cross-validation scheme revealed that the LSTM model with random-forest selection achieved a mean root-mean-square error of 2.33 m/s and mean absolute percentage error of 21.12%. Although these statistics are comparable to those of a 1-h persistence baseline model on average, the proposed model considerably outperformed the persistence baseline model during rapid intensification and decay phases, reducing errors by approximately 45%. Forecast errors generally remained within the ±5 m/s operational advisory threshold. The results of this case study for RCTP suggest that feature selection can be combined with sequence-based deep learning to provide robust decision support for aviation operations during extreme weather events.