TCN-Informer-Based Flight Trajectory Prediction for Aircraft in the Approach Phase

Zijing Dong, Boyi Fan, Fan Li, Xuezhi Xu, Hong Sun, Weiwei CaoSustainability 2023 · DOI

TL;DR

This paper introduces a novel hybrid deep learning model, TCN-Informer, for predicting the flight trajectories of General Aviation (GA) aircraft during their approach phase. By combining the feature extraction capabilities of Temporal Convolutional Networks (TCNs) with the long-sequence forecasting power of the Informer model, they achieve significantly improved prediction accuracy, crucial for enhancing flight safety and optimizing air traffic control.

What Problem Does This Paper Address?

Accurate flight trajectory prediction (TP) is essential for air traffic control (ATC) to monitor, track, and manage aircraft, thereby preventing conflicts and ensuring safety. While TP has been studied for commercial aircraft, General Aviation (GA) aircraft present unique challenges. GA operations often occur in less controlled airspace, are more susceptible to various factors like weather and pilot decisions, and the approach phase is particularly critical due to its visual nature and higher accident rates. A key concern during the approach is the decision to "go-around" (abort landing) if conditions are not safe. Unnecessary go-arounds lead to increased fuel consumption and emissions, while delayed or missed go-arounds can be dangerous. Therefore, precise trajectory prediction during this phase is vital for both safety and efficiency, but existing methods struggle with the complexity and data characteristics of GA aircraft.

Methods

The core of this paper is the proposed TCN-Informer architecture, a hybrid deep learning model designed to tackle the challenges of GA aircraft trajectory prediction. The approach involves two main stages:

  1. Feature Extraction with TCN: Temporal Convolutional Networks (TCNs) are employed to extract spatiotemporal features from the historical flight trajectory data. TCNs are chosen for their ability to handle sequential data effectively, maintain stable gradients (avoiding vanishing/exploding gradient problems common in RNNs), and process inputs of arbitrary length. They use dilated causal convolutions to expand their receptive field, allowing them to capture long-term dependencies in the flight data.
  2. Trajectory Prediction with Informer: The features extracted by the TCN are then fed into an Informer model for prediction. The Informer is an efficient variant of the Transformer architecture specifically designed for long sequence time-series forecasting (LSTF). It incorporates several key enhancements:
  3. Adaptive Input Embedding: This captures both local and hierarchical time-series information (e.g., seconds, minutes, hours) through multi-level embeddings.
  4. ProbSparse Self-Attention: This mechanism optimizes the standard self-attention by focusing only on the most important queries, significantly reducing computational complexity from O(L²) to O(L log L) for long sequences.
  5. Distilling Mechanism: This operation downsamples feature maps between layers, compressing dimensions and highlighting dominant features, further reducing computational load and memory usage.

The TCN-Informer model integrates these components into an encoder-decoder framework for trajectory prediction.

Datasets

The study utilizes 3-month Automatic Dependent Surveillance-Broadcast (ADS-B) data collected from GA aircraft between March and May 2021. This data represents normal flight operations and was collected from ground-based receivers. The dataset includes parameters such as timestamp, position (longitude, latitude), altitude, ground speed, and heading. For experimental purposes, the data was divided into training, validation, and test sets at an 8:1:1 ratio.

Key Insights and Ideas

  • Hybrid Approach Synergy: The combination of TCN for robust feature extraction and Informer for efficient long-sequence prediction proves to be more effective than using either model in isolation or standard LSTM. The TCN effectively captures the complex spatiotemporal dynamics, while the Informer leverages these features for accurate forecasting.
  • Importance of Approach Phase: The paper highlights the critical nature of the approach phase for GA aircraft, where manual operations and visual cues increase the risk. Accurate TP here directly impacts safety and the decision-making process for go-arounds.
  • Addressing GA Data Challenges: The proposed method is designed to handle the characteristics of GA flight data, which can be scattered and incomplete, by focusing on feature extraction and robust prediction models.
  • ProbSparse Attention Efficiency: The use of ProbSparse self-attention in the Informer is crucial for handling the long sequences of flight trajectory data efficiently, making the model computationally feasible.

Key Results

The TCN-Informer model demonstrated superior performance across various evaluation metrics (MAE, RMSE, MAPE) compared to baseline models like TCN, LSTM, and Informer.

  • Overall Accuracy: For all three flight subjects tested (Slow-Flight, Basic Visual Flying, Basic Instrument Flying), the TCN-Informer consistently achieved the lowest MAE, RMSE, and MAPE values.
  • Quantitative Improvements:
    • Compared to TCN, MAE was reduced by up to 70.63% and RMSE by up to 73.36%.
  • Compared to LSTM, MAE was reduced by up to 32.01% and RMSE by up to 32.94%.
  • Compared to Informer, MAE was reduced by up to 54.77% and RMSE by up to 64.01%.
  • Dimensional Accuracy: Analysis of individual dimensions (longitude, latitude, altitude) also showed the TCN-Informer outperforming other models, particularly in altitude prediction where it exhibited a stable effect with the lowest error range.
  • Visual Validation: 3D trajectory plots confirmed that the TCN-Informer and LSTM models' predictions closely matched the actual trajectories, with TCN-Informer showing slightly higher accuracy.

Limitations and Future Work

  • Dataset Sensitivity: The authors acknowledge that the proposed model, being a self-supervised learning model, is sensitive to the training dataset. The size and quality of the dataset significantly impact its learning capabilities.
  • Computational Efficiency: While the Informer improves efficiency, the TCN-Informer is still noted to be less computationally efficient than some supervised learning models like TCN and LSTM. This is partly attributed to the model treating flight parameters as unlabeled data, requiring a substantial training sample size.
  • Outliers in Prediction: Despite overall improvements, the proposed architecture can still exhibit outliers in its prediction results.
  • Future Directions:
    • Improving the computational efficiency of self-supervised learning for flight trajectory prediction.
  • Exploring solutions for anomaly detection within the TCN-Informer framework.
  • Investigating more suitable methods for spatiotemporal feature extraction to further enhance efficiency and accuracy.
  • Comparing the proposed model with other flight trajectory prediction algorithms, considering potential differences in labeling and network complexity.

Summary updated

Abstract

Trajectory prediction (TP) is a vital operation in air traffic control systems for flight monitoring and tracking. The approach phase of general aviation (GA) aircraft is more of a visual approach, which is related to the safety of the flight and whether to go around. Therefore, it is important to accurately predict the flight trajectory of the approach phase. Based on the historical flight trajectories of GA aircraft, a TP model is proposed with deep learning after feature extraction in this study, and the hybrid model combines a time convolution network and an improved transformer model. First, feature extraction of the spatiotemporal dimension is performed on the preprocessed flight data by using TCN; then, the extracted features are executed by adopting the Informer model for TP. The performance of the novel architecture is verified by experiments based on real flight trajectory data. The results show that the proposed TCN-Informer architecture performs better according to various evaluation metrics, which means that the prediction accuracies of the hybrid model are better than those of the typical prediction models widely used today. Moreover, it has been verified that the proposed method can provide valuable suggestions for decision-making regarding whether to go around during the approach.
BibTeX
@inproceedings{dong2023tcn,
  author = {Zijing Dong and Boyi Fan and Fan Li and Xuezhi Xu and Hong Sun and Weiwei Cao},
  title = {{TCN-Informer-Based Flight Trajectory Prediction for Aircraft in the Approach Phase}},
  booktitle = {Sustainability},
  year = {2023},
  doi = {10.3390/su152316344},
}
Figure