Research map / Probabilistic and scientific modeling / Molecules and proteins
FlightBERT++: A Non-autoregressive Multi-Horizon Flight Trajectory Prediction Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FlightBERT++, a non-autoregressive encoder-decoder for multi-horizon flight trajectory prediction using binary encoding with differential prediction and a differential-prompted decoder.
Problem
Multi-horizon flight trajectory prediction in air traffic control suffers from error accumulation and low inference efficiency when performed autoregressively, and binary encoding representations produce high-bit outlier errors.
Approach
Predicts multi-horizon flight trajectories in one pass via a horizon-aware context generator (horizon embeddings concatenated with a trajectory-level embedding), predicts differential values encoded in binary, and uses a differential-prompted decoder that feeds differential-sequence embeddings as a prompt.
Key results
- FlightBERT++ achieves MAE of 0.0124 (longitude) at horizon 15 versus 0.0159 for FlightBERT, per Table 1.
- Mean time cost of 6.81 ms versus 201.41 ms for FlightBERT at horizon 15, the fastest among deep learning models.
- Ablation shows removing the differential prompt mechanism raises horizon-15 longitude MAE from 0.0124 to 0.0153.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FlightBERT++ instead of LSTM+Attention | Flight trajectory prediction · Flight trajectory test set, horizon 15 | MAE (longitude): 0.0125 -> 0.0124 | Helped |
| use FlightBERT++ instead of LSTM+Attention | Flight trajectory prediction · Flight trajectory test set, horizon 15 | MAE (latitude): 0.0122 -> 0.0117 | Helped |
| use FlightBERT++ instead of LSTM+Attention | Flight trajectory prediction · Flight trajectory test set, horizon 15 | MAE (altitude): 9.05 -> 7.43 | Helped |
| use FlightBERT++ instead of LSTM+Attention | Flight trajectory prediction · Flight trajectory test set, horizon 15 | MDE: 2.04 -> 1.97 | Helped |
| use non-autoregressive FlightBERT++ instead of LSTM+Attention | Flight trajectory prediction · Flight trajectory test set, horizon 15 | MTC (ms): 14.43 -> 6.81 | Helped |
| add the ASA module rather than use a sum operator (ablation) | FlightBERT++ · Flight trajectory test set, horizon 15 | MAE (longitude): 0.0123 -> 0.0124 | No clear effect |
| add the ASA module rather than use a sum operator (ablation) | FlightBERT++ · Flight trajectory test set, horizon 15 | MDE: 2.05 -> 1.97 | Helped |
| add the differential-prompted mechanism (ablation) | FlightBERT++ · Flight trajectory test set, horizon 15 | MAE (longitude): 0.0153 -> 0.0124 | Helped |
| add the differential-prompted mechanism (ablation) | FlightBERT++ · Flight trajectory test set, horizon 15 | MDE: 2.53 -> 1.97 | Helped |
Surprising findings
- The FlightBERT++ model has about 44M parameters, far larger than the LSTM and Transformer baselines, yet it is the fastest deep learning model due to non-autoregressive decoding.
Limitations
- Evaluated on a single proprietary ATC dataset from one region, with no public release.
- Model size (44.13M parameters) is much larger than the baselines.
- Differential operation may lose geographical and kinematic features, as the authors acknowledge.
Experimental setup
Datasets: real-world flight trajectory dataset from an industrial ATC system (8643 trajectories, 20-second intervals)
Benchmarks: in-house ATC flight trajectory test set
Metrics: MAE MAPE RMSE MDE MTC
Baselines: LSTM Transformer Kalman-Filter FlightBERT LSTM+Attention
Builds on: FlightBERT Transformer LSTM+Attention binary encoding representation
Closest papers by meaning
- Aeolus: A Multi-structural Flight Delay Dataset (NeurIPS 2025 · similarity 0.85)
- Effective and Efficient Representation Learning for Flight Trajectories (AAAI 2025 · similarity 0.84)
- Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM (CVPR 2025 · similarity 0.83)
- C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction (AAAI 2025 · similarity 0.82)
- AutoTraces: Autoregressive Trajectory Forecasting via Multimodal Large Language Models (CVPR 2026 · similarity 0.82)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.