Research On The Key Technologies Of Dynamic Data-driven Flight Delay Prediction

With the development of society and the increasing of air traffic flow, flight delays have become aurgent problem of the civil aviation sector. Since Flight delays can be caused by many reasons, mostlyuncertain, the problem hasn’t been solved effectively. Further more, there is a chain reaction in flightdelays, that is, when one flight delays, it may affect the punctual arrival or departure of the next flight,and then propagate to more downstream flights indirectly. If we can predict the large area flight delaysin a hub airport, measures can be taken timely to reduce flight delays propagation and improve theimage of air services. Flight delays prediction can be seen as the state estimation problem of dynamicsystem, moreover, the whole process of flight operation is monitored and there is adequate real-timedata. Therefore, this dissertation attempts to applie the dynamic data-driven approach to predict flightdelays, and the key issues in the application are disscused. The main contributions of this dissertationare summarized as follows:(1) State space model is applied in the modeling of flight delays system, to obtain the predictionmodel for dynamic data-driven flight delays prediction. Based on the analysis of delay propagationevents sequence, the state space models for multi-task flight delays of single aircraft and the flightdelays of continuous arrival/departure flights are established respectively, so as to provide theprediction models.(2) Support vector regression is applied to estimate the key parameters in flight delays statespace model. Feature extraction, normalization and outlier removal are carried out on the historicrecords of flight operation to obtain the available data sets that can be used to model regression. Afterthe searching of the optimum parameters of support vector machine, air delay and scenes delayregression models are trained and tested, therefore the values of air delay and scenes delay can becontinuously updated by retrain the regression models according to the new flights operation records.(3) Based on the comprehensive study on the data assimilation problem, the filtering algorithm isproposed to the dynamically integrate the priori estimates of flight delays and the observations. Theposteriori estimations of delay state can be produced by the iteration of prediction and update. Toprovide a basis for the selection of data assimilation methods, experiments are carried out to comparethe prediction effect and efficient of the discussed filtering algorithms on all sixteen kinds of possibledynamic system with different nature.(4) As a supplement, an estimation method based on finite mixture model is proposed to calculate the flight delay probability distribution, and the optimal estimates of all the parameters in the modelare calculated by the genetic EM algorithm. Therefore, appropriate data assimilation method can beselected according to the distribution characteristics of the state variables and observation variables.The advantage of the genetic EM algorithm compared to the traditional EM algorithm is also verifiedin the calculation process of flight delays probability distribution model.(5) Architecture of dynamic data-driven flight delay prediction system is presented based on theabove researches. To verify the effectiveness of the proposed solution, flight delays of multi-tasksingle aircraft and continuous arrival/departure flights are predicted respectively. The predictionaccuracy is tested, and the effects of noise covariances and the number of continuous flights on theprediction accuracy are also verified at the same time.