Flight Delay Prediction For Civil Aviation Data Analysis Applications

With the rapid development of China’s civil aviation industry,there is a strong contrast between the expanding flight size of airlines and the limited available airspace resources.The contradiction between the continuously growing air transport demand and the comprehensive support capacity of civil aviation is becoming more and more prominent,resulting in the continuous low punctuality rate of China’s civil aviation.Accurate and effective flight delay prediction not only provides guarantees for normal travel of passengers,but also has great significance for the construction of civil aviation system.In the face of large-scale flight data,the current mainstream machine learning in the task of flight delay prediction has the problem of low accuracy and reliability,while deep learning methods such as Recurrent Neural Network do not comprehensively consider the effect of various correlation effects in the flight operation process,and thus the prediction results are often unsatisfactory.Therefore,after comprehensive analysis of the correlation effects,a flight delay prediction model based on Convolutional Neural Network is proposed in this thesis,and the problem of low recall due to imbalanced data is solved from the algorithm level and data level.The main work of the thesis includes:(1)Combining the factors that may affect the flight during operation,this thesis considers the effect of weather on flight delay.On this basis,it is proposed that flight delay is affected by various correlation effects,including flight string delay,airport delay and periodic delay.The experiments validate that fusing both weather information and flight correlation information in the data pre-processing stage can effectively improve the evaluation metrics such as accuracy and recall of the model.(2)A flight delay prediction model based on One-dimensional Convolutional Neural Network is designed.The spatial pyramidal pooling layer is used for multi-level feature extraction on the same feature vector,which not only solves the problem of irregular samples length after fusing correlation information,but also improves the performance of the network slightly.On this basis,the dense connection structure and attention module are used to improve the network,and the network proposed in this thesis is optimal in all evaluation metrics such as accuracy and recall compared with other mainstream methods.(3)After an in-depth analysis of the low recall of the model,this thesis discusses the problem of imbalanced flight samples from the algorithm level and data level.At the algorithm level,a cost-sensitive loss function is used to balance the weights of imbalanced data and difficult and easy samples.At the data level,Variational AutoEncoder and Generative Adversarial Network are used to augment the delayed samples to balance the flight data.The experiments show that the cost-sensitive loss function has a certain improvement on the performance of the model,and the use of deep generative networks to balance the data can greatly improve the recall of the model without excessive loss of accuracy compared to the traditional sampling methods.