Sector traffic flow prediction based on the attention-improved graph convolutional transformer network (AGC-T)

Abstract The improvement of the accuracy and real-time performance of sector traffic flow prediction is of great significance to air traffic management decision-making. Sectors operate under complex spatial structures and time dimensions. Some neural network methods adopt sequence order to gradually transmit information, which makes it difficult to achieve complete parallel training. Not only does it take too long to train, resulting in low training efficiency, but it is also easy to lose the effective correlation information of long sequence data. To this end, a sector traffic flow prediction method based on attention-improved graph convolutional transformer (AGC-T) network is proposed to improve the current traffic prediction problem for sectors. First, the graph structure information and historical traffic data of the sector are input into the graph convolutional network improved based on the attention mechanism to fully capture the spatial relationship with sectors as nodes. Combined with the transformer’s multi-head self-attention mechanism, it can directly focus on the sequence data at any position without gradually transmitting information. Not only does it improve efficiency through parallel training, but the encoder-decoder structure can also mine the information features in the traffic data, focus on the traffic data features of key nodes and more accurately predict sector traffic. Finally, the operation traffic data of sectors in typical areas in central and southern China are taken as an example to analyse the model. The results show that compared with other prediction models, the AGC-T model $RSME$ , $MAE$ and ${R^2}$ are 45.16%, 46.78% and 2.63% higher than the GCN model in the 15-min single-day traffic prediction task, and 41.74%, 35.27% and 1.20% higher than the GRU model. In the single-week traffic prediction task, $RSME$ , $MAE$ and ${R^2}$ are 37.12%, 40.54% and 3.55% higher than the GCN model, and 35.15%, 35.17% and 0.65% higher than the GRU model, respectively, showing better prediction performance. This study will help air navigation service providers (ANSP) to make sector traffic predictions more accurately, thereby implementing more scientific and reasonable traffic management measures.