Research On The Flight Requirement Prediction Using Deep Spatio-temporal Residual Networks

In the field of civil aviation,perceiving the demands of the customers accurately is crucial for the development of civil aviation industry.Accurate flight requirement prediction models not only can help us to perceive market demands in advance but also improve the service quality and income management of enterprises.Traditional flight requirement forecasting models are mostly based on historical data of passengers' orders,in which the timeliness,sensitivity and accuracy are limited.Fortunately,the continuous accumulation of users' ticket query data on ticket websites can effectively remedy these defects.Therefore,we propose to predict the flight requirements by analyzing and mining ticket query data.We first analyze the factors that affect flight requirements from many aspects.The results indicate that there are complex interactions between many air routes in the route network formed by a large number of routes,but the traditional time series analysis methods cannot capture these related information.In recent years,as an important research branch,convolutional neural networks can automatically extract the internal correlation features from input data,so as to reduce the error caused by handcrafted features.Therefore,in this paper,convolution neural networks are used to apply multi-layer convolution operations on the query matrix of the route network,which can capture the spatial dependence from near to long distance,and further model the relations between different routes in the route network.Aiming at the flight requirement forecasting problem in the airline network,we propose a deep-learning-based approach,called DSTRN-FRP,to forecast flight requirements.We firstly transform time series data of users' query volumes into grid maps.Considering two kinds of time dimensions,including query date and flight date,four different time fragments are extracted from the grid maps,and they are used in prediction models.Then we design an end-to-end structure of DSTRN-FRP based on the unique properties of the query data.More specifically,we employ a residual neural network framework to model the properties of temporal closeness and period of users' online query behaviors.For each property,we design a branch of residual convolutional units,each of which models the spatial properties of the query data.DSTRN-FRP learns to dynamically aggregate all the outputs of the four residual neural networks,and assign different weights to different branches and regions.What's more,the aggregation is further combined with the external factors,such as holidays and the day of the week,to predict a period of time series of flight requirements in the future.Experiments are carried out on a real historical query dataset provided by a GDS service provider.The experimental results show that the proposed DSTRN-FRP outperforms other existing forecasting methods,such as ARIMA and LSTM.It indicates that the DSTRN-FRP model can effectively predict the flight requirements.