Abstract
Accurate Air Traffic Flow Prediction (ATFP) is critical for enhancing airspace safety and efficiency. Achieving high-precision ATFP requires simultaneous consideration of traffic dynamics recurrence and spatiotemporal interaction. However, modeling both dimensions concurrently is challenging, as it demands that the model account for multiple latent factors. To address this, a novel deep learning-based model is proposed that simultaneously accounts for these two dimensions. First, key features with spatial representation value are distilled from the dual perspectives of causal influence and temporal correlation. Second, a sparse slope attention mechanism is designed to capture similar dynamic patterns in historical traffic flows. Finally, multi-scale temporal features are extracted from traffic flows, and a nonlinear activation function is employed to generate dynamic calibration factors, thereby reducing prediction biases across different date distributions. Experimental results on real-world air traffic datasets demonstrate that the proposed approach not only achieves superior performance in the ATFP task but also validates the effectiveness of each technical module in enhancing overall predictive accuracy.
Introduction
Predicting air traffic flow is a complex task involving multi-dimensional spatiotemporal interactions. Effective prediction must account for two primary factors:
- Spatiotemporal Interaction: Events in one sector, such as altitude changes to avoid weather, create ripple effects in adjacent sectors with varying time lags.
- Dynamics Recurrence: While traffic patterns often repeat (e.g., daily peaks), they are rarely identical in volume, leading to "magnitude bias" in traditional models.
Many existing models struggle when a weekend peak shares the same "shape" as a weekday peak but at a lower volume. Furthermore, the indiscriminate fusion of trajectory data (heading, altitude, etc.) often introduces noise that degrades performance.
Methodology: The TDDTNet Architecture
To address these challenges, the researchers developed TDDTNet, a model comprising three specialized modules designed to isolate signal from noise and focus on recurring dynamics.
Dual-Perspective Distiller (DPD)
Instead of processing all trajectory data, DPD utilizes Graph Attention Networks and multi-head attention to evaluate features from both a causal perspective (immediate impact) and a temporal correlation perspective (lagged impact). This distillation process identifies the most influential factors, keeping the input signal clean.
Traffic Dynamics Perception (TDP)
This module decomposes traffic data into trend and seasonal components. It introduces Sparse Slope Attention (SSA), which calculates the slope of data patches rather than absolute values. Because slopes are invariant to additive shifts, the model can recognize identical dynamic patterns regardless of total volume.
Prior Knowledge Calibrator (PKC)
PKC acts as a final refinement stage, using categorical data—such as hour of day and day of week—to generate calibration factors that scale predictions to match real-world social cycles.
Key Insights: Adaptive Patching and Shape Recognition
A core innovation of TDDTNet is its move away from fixed time windows. The model uses an Autocorrelation Function (ACF) to adaptively select patch sizes that align with the natural periodic cycles of the specific airspace.
By focusing on slope similarity, the model becomes robust to daily volume fluctuations. The researchers also discovered a "feature paradox": incorporating all trajectory factors (latitude, longitude, heading, altitude) actually increased error. Their distillation process proved that using only the most dominant factors provides a more potent signal for the neural network.
Experimental Results
TDDTNet was tested on five real-world datasets from Frankfurt (EDDF) and Jinan (ZSJN) airports. It consistently outperformed modern baselines like Informer and iTransformer:
- Accuracy: On the Frankfurt All-Altitude dataset, TDDTNet achieved a 15.38% reduction in RMSE and a 19.61% reduction in MAPE compared to Informer.
- Efficiency: The SSA mechanism reduced training time by 8.42% and memory usage by 20.35% compared to standard Sparse Attention.
- Robustness: The model maintained high precision even at long-term forecasting horizons of up to 96 hours.
Conclusion
TDDTNet represents a significant advancement in ATFP by prioritizing the "shape" of traffic dynamics over raw magnitude. While future work is needed to handle data imputation for missing sensor points, the current model provides reliable decision support for air traffic controllers and airspace management.
Summary updated