Research on Flight Delay Propagation Prediction Method Based on Transformer

In recent years, flight delay has become a problem that needs to be solved urgently. The problem of flight delay propagation prediction is studied in this thesis, which is the phenomenon of large area flight delay caused by an airport delay. The analysis and prediction of flight delay propagation in advance can assist civil aviation departments in controlling the flight delay rate and reducing the economic loss caused by flight delays. Firstly, the national flight data is cleaned and filtered to construct a dataset of flight chain affected by flight delay. At the same time, errors or missing values in the data are removed, and data related to flight delay are selected for subsequent analysis and modeling. Modeling analysis is performed on the issue of flight delay propagation, and a model of flight delay propagation is constructed to analyze the impact of flight delay propagation in airport network transmission. The model fully considers the interrelationship between flights, treating the flight chain as a whole rather than individual flights. This will be more conducive to exploring and understanding the causes and impacts of flight delay. Subsequently, in terms of constructing the dataset of flight chain structures, the method of data preprocessing and triplet flight chain dataset contains a series of flight chain data that can be used to predict the scope of flight delay propagation. Secondly, according to the features of flight delay, the Transformer model was adjusted and improved to enhance the accuracy and efficiency of prediction. The network architecture is restructured by injecting new convolutional pooling modules into the original network and adjusting the input and output dimensions of the network to make the overall model more tailored to flight delay prediction. Last but not least, the results of the experiments are showed that the modified model has low complexity, shorter time and good real-time performance. It makes prediction and decision faster, improves the real-time performance and better meets the needs of practical applications with an accuracy of 90.3%.