Delay Prediction Based on Deep Stacked Autoencoder Networks

Abstract

Accurate and timely flight delay information is essential for overall coordination of airports, airlines and air traffic management. However, because of the complexity of factors influencing flight delay, limitations like weak generalization ability, narrow application scope and unsatisfying prediction accuracy, could be found in existing flight delay prediction technology. Thus, a new method based on deep stacked autoencoders networks is proposed to predict flight delay in a future period, which totally considers its relationship with time and space and obtains information from high-dimensional data. A stacked autoencoder is adopted to train networks, deriving the characteristics of flight delay information from massive data by unsupervised learning and optimizing all the networks' parameters with backpropagation method. Future flight delay situation is predicted based on real data which was obtained from flight delay information of American airports announced in FAA website. The algorithm model reveals the evolution rule of flight delay in space-time variation and proves to be effective and superior after being compared with the performance of traditional neural network. Results from plenty of experiments have implicated that the prediction accuracy with deep stacked autoencoders is above 90%, which exceeds main delay prediction method over 5%, proving the efficiency of this method.