Research on the Air Traffic Flow Prediction Using a Deep Learning Approach

In order to improve the accuracy and robustness of the air traffic prediction, in this paper, a recurrent 3D convolutional neural network (R-3DCNN) based model is proposed to consider the spatial and temporal air traffic transitions comprehensively. A new data representation, i.e., traffic situation graphics (TSG), is firstly proposed to illustrate the traffic flow situations in a single instant. A TSG is generated by splitting the 3D earth space with fixed grid map and flight levels. Motivated by the applications of deep neural network, the 3D CNN and long short-term memory (LSTM) blocks are introduced to extract high-level features (spatial and temporal) from an TSG sequence. The proposed TSG also allows us to consider some real-time factors to enrich the input information. Thus, the model input is determined by combining the traffic situations on different flight levels with areas affected by other real-time factors, such as the adverse weathers, important activities and general aviation flights. The model output is the traffic flow on different flight levels for studied airspaces at next prediction instant. The busiest routes in China are used to conduct evaluation experiments. To determine the influence of temporal dependencies, the length of input sequence is set to 30, 60 and 90 minutes before the prediction instant to select optimal architecture of the proposed model. By evaluating the prediction results with three statistical factors, we can draw the conclusion that the proposed model can obtain accurate and stable prediction results of air traffic flow with distribution on different flight levels.