An intrinsic dynamics capture network for long-term airspace traffic prediction

Area of Science:

• Aviation management and artificial intelligence
• Time series forecasting
• Complex systems analysis

Background:

• Global air traffic growth necessitates improved safety and sustainable management.
• Existing short-term forecasting models struggle with long-term air traffic prediction accuracy.
• Challenges include declining model dynamics learning and positional information loss in long sequences.

Purpose of the Study:

• To develop a novel architecture for accurate long-term airspace traffic prediction.
• To address the limitations of existing models in capturing long-term dynamics and positional information.
• To enhance aviation safety and airspace management through improved forecasting.

Main Methods:

• Proposed IDCformer architecture with Trend and Seasonal Extraction (TSE), position-aware Patch Time Series Transformer (PatchTST), and Local Self-Attention (LAT) modules.
• TSE module stabilizes data and extracts long-term dynamics.
• Position-aware PatchTST integrates convolutional positional signals to prevent temporal order loss; LAT refines local fluctuations.

Main Results:

• IDCformer demonstrates superior predictive performance compared to state-of-the-art models on real-world air traffic data.
• The architecture successfully captures intrinsic airspace flow dynamics for long-term forecasting.
• Incorporating external data as additional input features further enhances IDCformer's long-term prediction accuracy.

Conclusions:

• The IDCformer architecture offers a significant advancement in long-term airspace traffic prediction.
• Its ability to leverage intrinsic dynamics and external information makes it a powerful tool for aviation management.
• The findings highlight the potential for more robust and sustainable air traffic control systems.