Real-time Arrival Traffic Management Using Machine Learning Approach

In recent years,the demand for air transportation has grown rapidly while the flow of air traffic is increasing.Due to the limited airspace resources,ensuring the efficiency and safety of air traffic management becomes more and more challenging.As an important part of the civil aviation management system,the existing terminal area flow management strategies are no longer able to meet the growing demand for air traffic flow.To solve this problem,the machine learning method was used to study the terminal area flow prediction and flow management.The short-term prediction of terminal area arrival flow is the basis of terminal area flow management.It can provide decision-making support for terminal area air traffic management and improve the efficiency of the flow control.After comparing the advantages and disadvantages of the existing traffic prediction methods and combining their advantages,an arrival traffic prediction framework based on air situation analysis module and machine learning model is proposed.In the first step,the information of the air situation is processed to create a feature vector for a machine learning model to predict whether a single aircraft will enter the terminal area.Then,the predicted number of arrival aircraft is added to produce a rough flow value.Finally,an additional machine learning model is trained to correct the error,and the correction value of the traffic prediction is obtained.Experiment shows that the minimum error of this method is only 0.35 frames/15 minutes.Another topic of this research focuses on the air traffic flow management.Firstly,according to the predicted traffic pressure,it is supposed to evaluate whether it is necessary to implement traffic flow management.If the traffic flow management is required,a deep reinforcement learning model for the real-time traffic management will be used.The state information of the aircraft is used as the input parameters of the model.The speed adjustment of the aircraft is used as the action output of the model.The optimization objective and constraint conditions are designed as the reward function.In the simulation environment,the model and the environment are interactively operated to train the network parameters of the deep reinforcement learning model.The trained model can select the optimal action under different conditions.The simulation experiment shows that the flight normality rate of the model after optimizing the parameters can reach more than 80%.