⚡ electrical engineering

Unlocking air traffic flow prediction through microscopic aircraft-state modeling

This paper introduces AeroSense, a novel state-to-flow framework that enhances short-term air traffic flow prediction in terminal airspace by directly modeling instantaneous microscopic aircraft states from ADS-B data, thereby overcoming the limitations of traditional aggregated time-series approaches and achieving superior accuracy, especially during high-density traffic.

Original authors: Feng Hong, Bin Wang, Anqi Liu, Jiangtao Zhao, Yanyong Huang, Peilan He, Guiyuan Jiang, Yanwei Yu, Tianrui Li

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Feng Hong, Bin Wang, Anqi Liu, Jiangtao Zhao, Yanyong Huang, Peilan He, Guiyuan Jiang, Yanwei Yu, Tianrui Li

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how many cars will arrive at a busy parking lot in the next 15 minutes.

The Old Way (The "Traffic Report" Approach)
Most current systems try to predict this by looking at a long video tape of the past few hours. They count the cars that passed by every 15 minutes in the past and use that history to guess the future. It's like trying to guess the weather tomorrow by only looking at a graph of yesterday's temperature, ignoring the fact that a storm front is currently moving in.

The problem with this method is that it treats the traffic as a smooth, blurry line. It forgets the individual cars. It doesn't know if a specific car is speeding up, slowing down, or turning around. It also gets confused when the number of cars changes suddenly, because it's used to looking at a steady stream of data.

The New Way: AeroSense (The "Live Camera" Approach)
The paper introduces a new system called AeroSense. Instead of looking at a history tape, AeroSense looks at a single, high-definition snapshot of the sky right now.

Think of it like a security guard standing on a tower with a powerful telescope. Instead of counting how many cars passed by in the last hour, the guard looks at every single plane currently visible and asks:

Where is it right now?
How fast is it going?
Is it turning toward the airport or away from it?
Is it close to the edge of the "controlled zone"?

How It Works (The Magic Recipe)
The researchers built a computer brain that treats the sky not as a flow of numbers, but as a dynamic group of friends who are constantly joining and leaving the party.

The Snapshot: At any given second, the system grabs the "state" of every plane in the area. This includes their location, speed, and even what the pilot intends to do (like a dial on the cockpit showing where they want to go).
The Group Chat: The system uses a special type of AI (called "masked self-attention") to let these planes "talk" to each other. It figures out which planes are influencing each other. For example, if Plane A is slowing down, it might cause Plane B behind it to slow down too. The system learns these relationships instantly.
The Counting Machine: Once the system understands the group, it doesn't just average them out. Instead, it acts like a counter. It realizes that "Traffic Flow" is simply the sum of all these individual planes. If there are 50 planes heading toward the airport, the prediction is 50. If 5 more join the group, the prediction instantly updates to 55.
Two Different Eyes: The system has two separate "heads" (predictors). One head focuses on the low-altitude area near the runway (where planes are landing), and the other focuses on the high-altitude area (where planes are cruising). They work separately because the rules are different for each zone.

Why It's Better
The paper tested this against the old "history tape" methods using real data from a major airport.

Accuracy: AeroSense was much better at predicting traffic, especially when the sky was crowded. It caught the "bumps" and sudden changes that the old methods missed.
No History Needed: Because it looks at the current situation so clearly, it doesn't need to remember the last 24 hours of data. It can make a prediction the moment it gets the data.
Robustness: When traffic got chaotic (like during a morning rush), the old systems made big mistakes. AeroSense stayed calm and accurate, like a guard who knows exactly what every person in the crowd is doing.

The Bottom Line
The paper argues that to predict the future of air traffic, you shouldn't just look at the history of the flow. You need to understand the microscopic state of every single aircraft right now. By treating the sky as a living, changing group of individual planes rather than a blurry line of data, AeroSense unlocks a much clearer and more accurate view of what's coming next.

Technical Summary: Unlocking Air Traffic Flow Prediction through Microscopic Aircraft-State Modeling

Problem Statement
Short-term air traffic flow prediction in terminal airspace (TA) is critical for proactive air traffic management (ATM), yet existing approaches predominantly rely on a macroscopic time-series forecasting paradigm. In this conventional framework, individual aircraft trajectories are aggregated into macroscopic flow sequences or spatio-temporal tensors before prediction. The authors argue that this aggregation obscures fine-grained information essential to traffic dynamics, such as aircraft kinematics, boundary interactions, and control intent. Furthermore, these methods typically require historical look-back windows and are sensitive to the discretization artifacts and information loss inherent in compressing continuous, variable-cardinality aircraft states into fixed time-series representations. The core challenge addressed is whether instantaneous microscopic aircraft states can provide sufficient information for accurate future flow prediction without relying on historical flow sequences.

Methodology: The AeroSense Framework
To address these limitations, the paper proposes AeroSense, a state-to-flow modeling framework that establishes an end-to-end mapping from instantaneous microscopic aircraft states to future regional traffic flows.

State-to-Flow Paradigm: Instead of treating airspace as a time series, AeroSense models the instantaneous airspace situation at time $t$ as a dynamic, unordered set of aircraft states, $S_t = \{s_1, \dots, s_{N_t}\}$ . The cardinality $N_t$ varies as aircraft enter and exit the airspace.
Microscopic State Representation: Each aircraft state vector $s_i \in \mathbb{R}^{18}$ $s_{i} \in R^{18}$ is constructed from ADS-B data and encodes five distinct components:
- Location: Latitude, longitude, and altitude.
- Kinematics: Ground speed, vertical speed, and heading angle.
- Control Intent: Dialed airspeed and altitude (from Mode Control Panel).
- Boundary Interactions: Distance to airspace boundaries (AP and AR), approach factors (cosine similarity between velocity and vector to airspace center), and airspace inclusion indicators.
- Temporal Context: Cyclical embeddings for hour and minute.
Model Architecture:
- Variable-Cardinality Handling: The dynamic set is padded to a fixed maximum capacity ( $N_{max}=120$ ) and processed via a shared Multi-Layer Perceptron (MLP) to generate latent embeddings.
- Interaction Modeling: A masked self-attention mechanism captures inter-aircraft dependencies. A mask matrix ensures that padding states (dummy entries) do not influence the attention weights, allowing the model to focus solely on valid aircraft interactions.
- Aggregation: The authors employ SumPooling (rather than Mean or Max pooling) to aggregate the embeddings. This choice is motivated by the nature of traffic flow as an accumulative quantity dependent on the number of aircraft, preserving the scale of the set in the global representation.
- Decoupled Prediction Heads: Two independent prediction heads estimate future traffic volumes for the Approach Airspace (AP) and Airspace Control Region (AR) separately, acknowledging their distinct flow dynamics.
Training: The model is optimized using a multi-task Huber loss function over both airspace targets.

Key Contributions

Paradigm Shift: The work reframes air traffic flow prediction from a macroscopic time-series problem to a microscopic state-to-flow learning problem, aligning the learning representation with the true physical state of traffic (a dynamic set of aircraft).
AeroSense Framework: Introduction of a framework that eliminates the need for historical look-back windows and manually specified hyperparameters, naturally accommodating varying traffic density and enabling real-time streaming inference.
Physically Grounded Representations: The design incorporates boundary interactions and control intent as inductive biases, enabling the model to reason about traffic evolution based on flight dynamics rather than solely statistical correlations.
Architecture Design: The specific combination of masked self-attention for interaction modeling and SumPooling for scale-sensitive aggregation is shown to be critical for performance.

Experimental Results
The framework was evaluated on a large-scale real-world dataset from an international airport terminal airspace (March–October 2025), predicting 15-minute-ahead traffic volumes for AP and AR regions.

Performance vs. SOTA: AeroSense consistently outperformed representative State-of-the-Art (SOTA) time-series baselines (including Autoformer, FEDformer, TimesNet, and iTransformer). In the high-density AR region, AeroSense reduced the Mean Absolute Error (MAE) by approximately 22.7% compared to the strongest competitor (TimesNet).
Ablation Studies:
- Removing the masked self-attention mechanism significantly degraded performance, confirming the necessity of modeling inter-aircraft interactions.
- Removing boundary interaction states led to clear performance drops, validating the importance of geometric cues over absolute GPS coordinates alone.
- Replacing SumPooling with Mean or Max pooling caused substantial performance degradation, particularly in high-density scenarios, reinforcing that traffic flow prediction relies on the count of aircraft rather than average states.
Robustness: AeroSense demonstrated superior robustness during high-volatility periods (e.g., morning peaks and late-night transitions), where conventional time-series models exhibited sharp error spikes. The model maintained a smoother error profile, achieving Pareto-optimal performance in dayparting multi-objective evaluations.
Comparison with Set-Based Baselines: While generic set-based models (DeepSets, SetTransformer) performed better than time-series baselines, AeroSense achieved the lowest errors, indicating that the specific airspace-aware design (state representation and decoupled heads) provides additional value beyond generic permutation-invariant modeling.

Significance and Claims
The paper claims that instantaneous airspace situations provide an effective alternative to conventional time-series-based forecasting paradigms. By directly modeling microscopic aircraft states, AeroSense preserves fine-grained dynamics and mitigates information loss caused by aggregation. The authors assert that this approach not only improves average predictive accuracy but also offers practical benefits for deployment: it simplifies real-time streaming inference by removing the need for persistent historical trajectory storage and long look-back windows. The findings suggest that future air traffic management systems could benefit from shifting focus from aggregated flow statistics to instantaneous, state-aware modeling of the airspace.

More like this