基于时空图卷积神经网络的离港航班延误预测

Departure flight delay prediction based on spatio-temporal graph convolutional networks

对于日益频发的机场航班延误,精准的航班延误预测是最重要的防范措施之一。通过谱图卷积将机场网络从不规则的图结构转换为规则的网络结构,利用图卷积神经网络(GCN)和门控线性单元(GLU)捕获网络中的时空相关性并组成时空卷积块,提出可以预测离港航班延误状况的时空图卷积神经网络(STGCN)。遴选美国51座枢纽机场构建机场网络,并预测未来一段时间内的机场离港准点率以检验STGCN用于预测航班延误的可行性。结果表明:当预测窗口为1天时,STGCN预测结果的平均绝对误差(MAE)相对于历史平均(HA)法、长短期记忆循环神经网络(LSTM)、堆栈自编码器(SAEs)分别下降了18.19%、10.45%、6.24%;当预测窗口为2天时,MAE分别下降了9.93%、3.96%、4.37%;当预测窗口为3天时,MAE分别下降了7.02%、2.47%、9.20%。实例证明STGCN相比传统模型能够显著提升航班延误预测的精度,并为机场制定延误决策提供参考指导。

Accurate flight delay prediction is one of the most important preventive measures for the increasingly frequent airport flight delays.The airport network is transformed from irregular topological structure to regular network structure by spectral convolution.The graph convolutional network(GCN)and gated linear unit(GLU)are used to capture spatio-temporal correlation in the network and form spatio-temporal convolutional blocks.A spatio-temporal graph convolutional neural network(STGCN)is proposed to predict the departure flight delay.51 major airports in the United States are selected to construct the airport network,and the on-time departure rate is predicted to carry out the example verification.The results show that when the prediction window is 1 day,the mean absolute error(MAE)of STGCN decreased by 18.19%,10.45%and 6.24%,respectively compared with the historical average(HA)method,long short-term memory(LSTM)and stacked autoencoders(SAEs).When the forecast window is 2 days,MAE decreased by 9.93%,3.96%and 4.37%;When the forecast window is 3 days,MAE decreased by 7.02%,2.47%and 9.20%.The example proves that STGCN can improve the accuracy of delay prediction and provide reference and guidance for airport delay decision.