Deep learning based short-term air traffic flow prediction considering temporal–spatial correlation

Abstract

In order to improve the accuracy and stability of air traffic flow prediction, an end-to-end deep learning-based model is proposed in this paper. By analyzing spatial correlations of adjacent areas and temporal correlations of historical traffic on given area, we firstly apply the gridded map method (including the flight levels) to encode the whole air traffic flow situation into a new data representation, i.e., traffic flow matrix (TFM). By the proposed data representation, inherent features of air traffic flow and their transition patterns on different cells and flight levels can be represented comprehensively. Learning from the powerful ability of convolutional neural network and recurrent neural network on modeling spatial and temporal correlations, the ConvLSTM module is proposed to build a trainable model for air traffic flow prediction. Since the output of the proposed model is also the TFM and shows the overall air traffic situation of studied regions, we call the proposed model as an end-to-end one. Experimental results on real data show the superior performance over existing approaches on prediction accuracy and stability. Furthermore, the proposed model can also predict the flow distribution on different flight levels in our application, which promotes the level of air traffic management. By analyzing the distribution of prediction errors on different cells, flight levels and predicting instants, we can draw the conclusion that spatial and temporal transition patterns of flight flow in air traffic system are fully learned by the proposed model. With the proposed model, more efficient air traffic flow measures are expected to be fulfilled to improve the operation efficiency.