A deep learning and parallel simulation methodology for air traffic management
Kim, Young Jin
MetadataShow full item record
Air traffic management is widely studied in several different fields because of its complexity and criticality to a variety of stakeholders including passengers, airlines, regulatory agencies, air traffic controllers, etc. However, the exploding amount of air traffic in recent years has created new challenges to ensure effective management of the airspace. A fast time simulation capability with high accuracy is essential to effectively explore the consequences of decisions from the airspace design phase to the air traffic management phase. In this thesis, two key components for enabling intelligent decision support are proposed and studied. To accelerate fast time simulations, a time-parallel simulation approach has been studied and applied to air traffic network simulation in addition to exploitation of spatial parallel simulation. This approach splits the simulation time axis into time intervals and simulates the intervals concurrently potentially achieving a high level of parallelism. This approach requires a way to ensure that the distributed simulation takes into account dependencies across time periods. A methodology to address this issue is proposed. The proposed time-parallel algorithm works seamlessly with the spatial parallel approach. In particular, the synchronization algorithm used for the spatial parallel simulation is integrated with the time-parallel simulation algorithm. In this thesis, an efficient algorithm spanning these aspects of the distributed simulations is proposed and implemented. The implemented simulation is tested in a variety of scenarios and balances time and spatial parallelism to improve speed up. As another aspect, to predict the future scenarios more accurate, it is necessary to feed the appropriate input vales to the simulation program. This input can be acquired by learning the previous patterns in data, statistically. Recent improvements in machine learning and artificial intelligence research enable an accurate prediction of the future state variables in the air traffic network system. Recurrent neural network is one type of algorithm which can effectively model sequential state variables. In that sense, a recurrent neural network approach is proposed for modeling the input of each simulation scenario. By utilizing a large amount of historical flight and weather data, the proposed recurrent neural network model learns the best parameters in the model to predict the future status of the airports in the National Airspace System (NAS). In particular, airports’ daily capacity in the future is a key input variable for the NAS simulation model. The proposed model is trained to accurately predict the airports’ daily capacity. Based on real world air traffic data, the improvements in the performance and the accuracy of both techniques have been investigated and presented. The proposed approaches show significant improvements for supporting air traffic management decision making.