Prediction and characteristic analysis of flight arrival delay

Abstract

To break the black box feature of XGBoost model and enhance its persuasiveness, an interpretable flight delay prediction model based on SHAP was proposed. Firstly, based on the fusion of flight history data and weather data, outliers were processed and features were selected by recursive feature elimination method. Secondly, a flight delay duration prediction model was constructed, and genetic algorithm was used for parameter optimization, then it was compared with commonly used models at present. Finally, based on the prediction of flight delay duration and the SHAP model, the importance of features was analyzed from two perspectives: overall features and the interrelationships between the features. The experimental results show that the XGBoost model optimized by genetic algorithm has higher prediction, with a decrease of 8.94% in MAE, 19.85% in RMSE, and 6.15% in MAPE, with higher accuracy compared to other models. The SHAP model can break the black box characteristics of the XGBoost model and enhance its interpretability, which provides some support for reducing flight delay duration.

Keywords

• air transport management; delay prediction; limit gradient lifting; parameter optimization; interpretation; feature selection