Short-Term Nationwide Airport Throughput Prediction With Graph Attention Recurrent Neural Network | JoVE Visualize

Area of Science:

Artificial Intelligence
Transportation Science
Network Analysis

Background:

Dynamic air traffic demand and limited capacity necessitate accurate airport throughput prediction for operational efficiency and resilience.
Existing research on predicting traffic throughputs or flight delays faces challenges due to the complex spatiotemporal dynamics of air transportation systems.

Purpose of the Study:

To propose and evaluate a novel deep learning model, Graph Attention Neural Network stacking with a Long Short-Term Memory unit (GAT-LSTM), for predicting short-term airport throughput across a national air traffic network.
To explore and compare different graph modeling methods (airport-based and OD-pair graphs) within the GAT-LSTM framework.
To analyze the influence of temporal versus spatial patterns on airport throughput prediction and assess model interpretability.

Main Methods:

Development of a GAT-LSTM deep learning model integrating Long Short-Term Memory (LSTM) layers for temporal feature extraction and a graph attention mechanism for spatial dependency capture.
Exploration of two graph modeling approaches: airport-based graphs and Origin-Destination (OD)-pair graphs.
Empirical testing using real-world air traffic data from 65 major Chinese airports over a 3-month period in 2017, with performance comparison against state-of-the-art models.

Main Results:

Temporal patterns were identified as the dominant factor in predicting airport throughputs over the national air traffic network, compared to spatial patterns.
The proposed GAT-LSTM model and the standard LSTM model demonstrated strong prediction accuracy across the network.
The GAT-LSTM model exhibited enhanced performance, particularly for airports with higher throughputs, and its parameters revealed learned spatiotemporal correlations.

Conclusions:

The GAT-LSTM model provides an effective approach for short-term airport throughput prediction, accurately capturing complex spatiotemporal dynamics in air traffic networks.
Temporal dynamics play a more significant role than spatial dependencies in national airport throughput prediction.
Model interpretability analysis offers valuable insights into the underlying topology and dynamics of the air traffic system.