基于注意力循环网络的4D轨迹预测模型

针对传统的轨迹预测方法很难获取轨迹的时空特征、实现高精度和实时预测等问题,提出了一种基于注意力机制的4D轨迹预测模型ARTP(Attentional Recurrent Trajectory Prediction)。首先,采用正则化方法对各飞行轨迹进行重构,得到等时间间隔的无噪声高质量飞行轨迹;其次,使用长短期记忆(Long Short-Term Memory, LSTM)对飞机飞行轨迹进行时空特征提取;最后,利用注意力机制来捕获飞行历史轨迹中的多层次周期性。该模型有效地利用了周期性的性质来增强LSTM的活动性预测。在真正的广播式自动相关监视系统(Automatic Dependent Surveillance-Broadcast, ADS-B)历史轨迹数据上进行实验和同类方法进行对比,ARTP模型的均方根误差比CNN-LSTM模型低21.04%。实验结果表明,基于注意力机制的飞机轨迹预测模型能够取得更高精度的预测结果。

A 4D trajectory prediction model based on attentional recurrent network

For the problem that it is difficult for traditional trajectory prediction methods to obtain the spatio-temporal characteristics of trajectory and achieve high precision and real-time prediction,a new 4D trajectory prediction model ARTP(Attentional Recurrent Trajectory Prediction) based on attention mechanism is proposed.First,the regularization method is used to reconstruct each aircraft trajectory to obtain the high quality noise free flight trajectory with equal time interval.Then the Long Short-Term Memory(LSTM) is used to extract the spatio-temporal feature of the aircraft trajectory.Finally,the attention mechanism is exploited to capture the multi-level periodicity of the aircraft history trajectory.The model effectively takes advantage of the periodicity property to enhance the activity prediction of LSTM.The experiment is conducted on real Automatic Dependent Surveillance-Broadcast(ADS-B) historical trajectory data and ARTP is compared with similar methods.It is shown that the root mean square error(RMSE) of ARTP model is 21.06% lower than that of CNN-LSTM model.The result demonstrates that the aircraft trajectory prediction model based on attentional mechanism model can achieve higher accuracy.