Research On Flight Delay Prediction Based On Multi-model Fusion

In recent years,as the concept of ‘intelligent transportation’ has been proposed and its applications have been developed extensively,the frequency of the airplane in the civil transportation has been continually increased.Air traffic has become one of the most common choices for people on the long-distance trips.Thus,how to process flight information data with Machine Learning and Artificial Intelligence properly,which has a profound influence on the stability of society and the normal order of work and life.With the high attention on the related applied technologies of machine learning,the research on the transportation field has brought forth new fields,especially on the aspect of the flight delay forecast.Thus,there is a high application value on accurate flight prediction with machine learning technologies.So far,there are various technologies for flight prediction with diverse advantages.The mainstream technologies include convolutional neural networks,data mining,supervised learning,etc.However,due to the diversity of the influence factors in flight prediction,flight delays tie up limited resources and disrupt the original flight plans,which is especially serious in the jumbo airport.Such various factors would lead to the inaccurate and unstable prediction of flight delays,and certain limitations on the feature selection of prediction information at the same time.Thus,how to achieve and optimize the model applied to flight delay prediction has become one of the current research hotspots.In order to satisfy the requirement of the high accuracy of the flight delay predictions in the transportation industry,after the study on the techniques and algorithms used at the current stage,this thesis proposed a prediction model for flight delay based on the multi-model integration.This model integrates the traditional single model including the Random Forest model,XGBoost model,and Light GBM model,and gives play to the sensitivity advantages of each model to different dimensional features.Based on the analysis of flight information data and the study of Stacking algorithms,this thesis optimized the model on the time weighting and base learner weighting and integrated these models with optimized Stacking algorithms.Finally,the comprehensive comparison on the experiment result has been carried out by RMSE(root mean square error)and MAPE(mean absolute percentage error).On this basis,the original data sample has been pre-processed in this experiment.Besides,in the data set,the dirty data cleaning and missing data manual completion were carried out.One-Hot Encoding technology has been used to deal with the discrete values,to ensure the consistency of the data.The noisy data with large deviation values have been removed.The feature selection has been carried out according to the featured marker built,to sieve out the characteristics including flight number information,departure information,weather information,and special information.In the training process of the experiment,base learner weight control has been added,to guarantee the stability of the training results in the base learner layer,and improve the robustness of the model.Finally,in this thesis,the flight delay prediction result of this experiment was compared with that of the traditional Stacking algorithms and a single model.It proved that the algorithm model proposed in this thesis has a better performance in the stability and accuracy in the scenario of flight delay prediction,and it could provide a complete solution to the problem of flight delay prediction.