Aircraft Trajectory Prediction Using Social LSTM Neural Network

In this paper, we propose a Social Long Short-Term Memory (SLSTM) Neural Network for aircraft trajectory prediction. This model builds an LSTM network for each aircraft and uses a merging layer to merge the hidden states of its neighboring LSTMs. Then, it uses the merging results and trajectory information as its input for the next time step. Finally, the model is trained by maximizing the probability of real trajectory and evaluate the prediction effect. The experiment is conducted with the flight trajectory dataset over the San Francisco Bay Area in 2006. The evaluation shows that our model has the smallest error from 17 to 18 o'clock when the airspace's flight trajectory density is the highest. The average horizontal error per point is about 282 meters, and the average vertical error per point is about 10 meters.