基于UKDE和XGBoost的航班过站时间动态预测  

Dynamic Prediction of Flight Transit Time Based on UKDE and XGBoost

作　　者：吴薇薇[1] 熊奥萍 唐红武 WU Weiwei;XIONG Aoping;TANG Hongwu(College of Civil Aviation,Nanjing University of Aeronautics&Astronautics,Nanjing 211106,China;Travelsky Mobile Technology Limited,Beijing 100029,China)

机构地区：[1]南京航空航天大学民航学院,南京211106 [2]中航信移动科技有限公司,北京100029

出　　处：《南京航空航天大学学报》2023年第6期1016-1024,共9页Journal of Nanjing University of Aeronautics & Astronautics

基　　金：国家自然科学基金(U2033205,U1933118);南京航空航天大学校企协同育人平台工程实践计划项目(2022QYGCSJ59);南京航空航天大学科研与实践创新计划(xcxjh20220712)。

摘　　要：为了提高机场运行高峰时航班过站时间预测的精度及可靠性,研究了一种结合无偏核密度估计(Unbiased kernel density estimation,UKDE)和极端梯度提升决策树(Extreme gradient boosting,XGBoost)模型的航班过站时间动态预测方法。首先,考虑模型输入变量航班密度的连续性和不确定性变化,利用UKDE法估计机场航班密度,将其作为动态指标输入模型。其次,引入量子粒子群(Quantum particle swarm optimization,QPSO)法优化XGBoost模型。最后,考虑前序航班延误发生前后输入特征的变化,利用初始预测结果对航班密度进行修正,得到二阶段预测结果。研究结果表明:本文方法在高峰时段的预测平均绝对误差为7.365 min,效果优于随机森林(Random forest,RF)、粒子群(Particle swarm optimization,PSO)-XGBoost和XGBoost,修正后的预测结果平均绝对误差减少了3.373 min;模型输入参数按敏感性程度由高到低依次为航班密度、前序航班提前到港时间和延误到港时间。To improve the accuracy and reliability of flight transit time prediction during airport peak periods,a dynamic flight transit time prediction method combining the unbiased kernel density estimation(UKDE)and the extreme gradient boosting(XGBoost)models is proposed.Firstly,considering the continuous and uncertain changes in the flight density of the model input variables,the UKDE method is used to estimate the flight density of the airport as a dynamic indicator.Secondly,the quantum particle swarm optimization(QPSO)method is introduced to optimize the XGBoost model.Finally,the changes in the input characteristics before and after the occurrence of the preceding flight delays are considered,and the initial prediction results are modified to obtain the two-stage prediction results.The final results verify that the average absolute error of the prediction in peak hours is 7.365 min,which is better than those of random forest(RF),particle swarm optimization(PSO)-XGBoost and XGBoost,and the average absolute error of the modified prediction results is reduced by 3.373 min.The model input parameters,in descending order of sensitivity,are flight density,early arrival time of preceding flights and delayed arrival time.

关 键 词：航空运输 时间预测 极端梯度提升决策树 航班过站保障 核密度估计 

分 类 号：V351[航空宇航科学与技术—人机与环境工程]