Enhancing flight delay prediction using machine learning techniques

Abstract

Flight delays remain a persistent challenge within the U.S. aviation sector, impacting both operational efficiency and passenger satisfaction. This study presents a comprehensive machine learning framework to accurately predict flight arrival delays in minutes using an enriched dataset comprising historical flight records and weather conditions. Multiple models were developed and evaluated, with particular emphasis on deep learning approaches. A fine-tuned Long Short-Term Memory (LSTM) network outperformed other models, achieving a high coefficient of determination (R2 = 0.9385), a low Mean Absolute Error (MAE) of 10.7 minutes, and a Root Mean Squared Error (RMSE) of 15.3 minutes. Comparative models such as XGBoost and Artificial Neural Networks also demonstrated strong performance, validating the robustness of the feature set. To ensure interpretability and model transparency, Local Interpretable Model- Agnostic Explanations (LIME) were applied, revealing that departure delay, time-of-day, and specific carriers were among the most influential features. The final LSTM model was deployed through a user centric application, enabling real time delay prediction based on user flight details.

KEY WORDS: Scheduled Time of Arrival, Actual Time of Arrival, Delays, Airport Artificial Neural Network, Machine Learning, LSTM.