4D trajectory lightweight prediction algorithm based on knowledge distillation technique.

Area of Science:

• Artificial Intelligence
• Aerospace Engineering
• Computer Science

Background:

• Current 4D trajectory prediction methods face challenges in multi-factor feature extraction and high computational costs.
• Real-time air-traffic management requires efficient and accurate trajectory prediction frameworks.

Purpose of the Study:

• To develop a lightweight prediction framework for real-time air-traffic management.
• To address limitations in feature extraction and computational expense of existing methods.

Main Methods:

• A hybrid Residual Convolutional Block Attention Module-Temporal Convolutional Network-LSTM (RCBAM-TCN-LSTM) architecture was proposed.
• A teacher-student knowledge distillation mechanism was employed, using RCBAM as the teacher and TCN-LSTM as the student network.
• Historical ADS-B trajectory data was preprocessed using cubic spline interpolation and sliding window techniques.

Main Results:

• The distilled RCBAM-TCN-LSTM model demonstrated 40%-60% reductions in MAE, RMSE, and MAPE.
• The model showed a 4%-6% improvement in R-squared (R²) across various prediction horizons.
• Computational complexity was significantly reduced while maintaining high prediction accuracy.

Conclusions:

• The proposed method effectively balances high-precision spatiotemporal modeling with lightweight deployment.
• The framework enables real-time air-traffic monitoring and early warning on standard hardware.
• This offers a scalable solution for enhancing air-traffic control safety and efficiency.