Research On Interpretable Prediction Of Flight Departure Delays Using A Data Fusion Based Machine Learning Combination Model

Flight delay is one of the common problems faced by airlines,which also results in losses for airlines.The causes of flight delays are varied,including mechanical failures,severe weather,air traffic control,passenger delays,and flight scheduling issues,among others.To reduce flight delays,airlines usually take measures such as pre-arranging standby aircraft or crew,improving the robustness of flight scheduling,etc.,to minimize the impact of flight delays.If flight delay can be predicted,it will help airlines better plan flight schedules and reduce losses.To address the issue of predicting flight delays,this study analyzed publicly available flight data in the United States.First,the flight dataset was screened to remove features unrelated to takeoff and to select takeoff flight data from five major US airports.The flight meteorological characteristics were fused with the planned takeoff time,including wind direction,wind speed,visibility,cloud condition,etc.The fused data were then standardized to obtain a flight dataset containing meteorological conditions for training prediction models.Five commonly used algorithms,random forest,light GBM,XGBoost,SVR,and neural networks,were selected for training and prediction.By comparing mean squared error and mean absolute error,training time,interpretability,etc.,light GBM and XGBoost were found to perform better.Forward feature selection and Bayesian method were further used to optimize the models.Subsequently,SHAP analysis was performed on some of the models’ prediction results,and their results were explained.By interpreting the model’s prediction results,decision guidance can be provided to professionals.However,during the overall interpretation process,the mixing of data from the five airports resulted in some trends not being obvious.Therefore,the original dataset was split by airport,and SHAP analysis was used to analyze the prediction results of each airport and its feature importance ranking,further explaining the prediction results,ultimately proving that the problems in the overall analysis were caused by the mixing of data from different airports.According to the combination forecast strategy,the results of the two best-performing models were combined,and by comparing the mean squared error and mean absolute error of different weight calculation strategies,the model with the smallest prediction error was obtained.To make this research more applicable,a flight takeoff delay prediction system was developed.The system was analyzed for requirements,overall architecture design,and functional design.Based on the Gradio framework,the development of the front-end page and related functions was implemented,and it has the function of explaining.Overall,the developed system has some application value.