Research On Airport Delay Prediction Based On Regional Residuals And LSTM Network

The demand of air transportation is increasing as time goes by,people’s requirements for the quality of aviation services are becoming more and more strict.At the same time,with the rapid changes of information technology,deep learning technology has matured and efficiently served people in various fields.Therefore,with the goal of airport delay prediction and the basis of deep learning algorithms,an airport delay prediction method based on region residual and long short term memory(RR-LSTM)network is proposed.The main work of the thesis is as follows:Firstly,because the traditional data processing method does not meet the requirements of civil aviation data characteristics and long short term memory(LSTM)network,based on the full analysis of flight data and meteorological data,a hybrid coding method and time series construction method are proposed to prepare for the feature extraction work of the network.The experimental verification shows that the proposed hybrid coding method can improve the performance of data in the network,maximize the feature extraction ability of the network,and improve the prediction accuracy of the network.Then,an airport delay prediction method based on LSTM network is proposed.This method uses LSTM network to learn the time dependence of airport delay data,which make the physical meaning of the model is more in line with the characteristics of airport delay events and improve the prediction accuracy,to further optimize the air traffic control decisions by airlines and other departments.The super-parameter values of the LSTM network are discussed through experiments,and the optimal time step and hidden layer number based on the existing computing resources are determined.Compared with the traditional network experiments,it shows that the LSTM network has strong data processing capability.The network universality verification is carried out on the actual operational data of several airports,and the airport objects that the method can be effectively applied were determined.Finally,due to the gradient disappearance problem in the LSTM network,the prediction accuracy cannot be improved.In order to better solve the gradient disappearance problem of LSTM network,the back propagation process of multi-layer LSTM network is deduced and proved.Then,based on theory and practice,an RR-LSTM network is proposed,and the network consists of three parts,including an LSTM network module,an region residual module,and a one-way pooling module.By adding the regional residual module,the model can be more stable and achieve the purpose of deep training;the purpose of one-way pooling is to perform dimensional transformation to connect the regional residual module and the LSTM network module,and the number of introduced parameters can be reduced as much as possible.By comparing the prediction accuracy of adding weather information and flight information,the importance of meteorological information for airport delay prediction is qualitatively explained.