Attention-based Deep Learning Model for Flight Delay Prediction using Real-time Trajectory

This paper presents a deep learning model termed LSTM-Attention based Time-dependent Flight-delay Classifier (LATTICE) for real-time flight arrival delay classification. Initially, this model incorporates a comprehensive set of factors influencing flight delays, including weather conditions, flight information, and en-route real-time trajectory data provided by ADS-B technology. Subsequently, LATTICE leverages a full-sequenced LSTM network for the extraction of deep temporal trajectory features and employs an attention network for the allocation of weights and mapping of relevant information. Ultimately, the model utilizes a masking layer to address the challenges posed by varying trajectory lengths, and experimental results demonstrate a significant enhancement in the accuracy of flight delay predictions as a result of these integrated measures. The model classifies incoming flights into On-Time/Late and Early/Punctual/Late. On being evaluated against historical data, it achieves about 91% accuracy and 0.96 AUC at predicting delay, yielding better predictions compared to baseline models. Trajectory inputs improve the prediction by about 15%. The model is real-time via ADS-B technology, robust via adaptive improvement with continuous training, and able to handle both late and early arrivals. This paper demonstrates that the real-time trajectory inferred from ADS-B messages can add significantly to the reliability of delay prediction.