基于深度时空卷积网络的民航需求预测
Deep spatio-temporal convolutional networks for flight requirements prediction

摘要 在线机票预订网站上的用户查询量变化是真实的民航市场需求变化的反映.通过对机票查询数据进行分析，可以准确地预测航班需求，以利于民航业做出快速的市场反应.提出了一种基于深度时空卷积神经网络的民航需求预测模型（DSTCN-FRP），将用户查询量时间序列数据转换成航线网格图，设计多层卷积神经网络来捕捉用户需求与查询数据之间的时间和空间依赖，同时加入节假日等外部因素，最后得到未来一段时间内的民航需求量.在某在线订票网站的真实查询数据集上进行了实验，结果表明:DSTCN-FRP模型优于其他现有的预测方法,其MAE比其他方法降低了15%~50%，RMSE降低了12%~28%.
Abstract：The changes of users’ query volume in online fight ticketing systems indicate the changes of requirements in civil aviation market. By analyzing users’ online query behaviors, we can accurately predict flight requirements, which is very conducive for airlines and agencies to take effective marketing actions immediately. In this paper, we propose a deep-learning-based approach, called DSTCN-FRP, to forecast flight requirements.We first transform time series data of users’ query volumes into grid map, then design multi-layer convolution neural network to capture the time and space dependency between user requirements and query data. In addition, we further add external factors, such as weather and day of the week, to predict a period of time series of flight requirements in the future. Experiments on a real-world users’ query dataset collected from an online ticketing site demonstrate that the proposed DSTCN-FRP outperforms other existing forecasting methods, where its MAE falls by 15% to 50% than other methods and RMSE falls by 12% to 28%.