Improving Flight Trajectory Predictions with

Bayesian-Optimized ConvLSTM Models

Thi-Lich Nghiem1,2,3,4[0000−0002−0378−4977] , Thi-Lan Le3[0000−0001−9541−3905] ,

Viet-Duc Le3 , Pierre Maréchal1[0000−0001−7121−5699] , Daniel
Delahaye2[0000−0002−4965−6815]
1 Toulouse Mathematics Institute, Paul Sabatier University, Toulouse, France
thi-lich.nghiem@univ-tlse3.fr
pierre.marechal@math.univ-toulouse.fr
2 OPTIM Team, French Civil Aviation University, Toulouse, France

delahaye@recherche.enac.fr
3 School of Electrical and Electronic Engineering, Hanoi University of Science and
Technology, Hanoi, Vietnam
lan.lethi1@hust.edu.vn
DUC.LV192777@sis.hust.edu.vn
4 Thuongmai University, Hanoi, Vietnam

Abstract. Flight trajectory prediction is a critical task in aviation, enabling ef-

ficient air traffic management and ensuring safe and seamless flight operations.
Conventional methods for trajectory prediction struggle to capture complex spa-
tiotemporal dependencies and uncertainties inherent in the aviation domain. Within
this study, we introduce an innovative method to predict flight trajectories, named
BayesCoordLSTM. It is a hybrid model that transforms coordinate and applies
Bayesian theorem into the ConvLSTM models. The proposed model leverages the
spatial features gleaned by the Convolutional Neural Network (CNN) architecture
and the temporal dependencies captured by Long-Short Term Memory (LSTM)
to enhance the accuracy of trajectory predictions. By incorporating Bayesian the-
orem, our model provides a probabilistic trajectory forecasts and associated con-
fidence levels while coordinate transformation enhances spatial awareness and
predictive capabilities. The paper presents experimental results demonstrating
the effectiveness of the proposed BayesCoordLSTM-based approach in improv-
ing flight trajectory prediction accuracy, with a focus on Mean Absolute Error
(MAE) and Root Mean Square Error (RMSE) values. The integration of the
Bayesian theorem and Coordinate transformation into ConvLSTM models rep-
resents a substantial advancement in the field of flight trajectory prediction.

Keywords: Air traffic management · Bayesian inference · Coordinate transfor-

mation · Flight trajectory prediction · Uncertainty quantification

1 Introduction
Flight trajectory prediction is vital in modern aviation and air traffic management (ATM).
It involves estimating a flight’s future position, altitude, heading, and speed based on its
historical data and current state. Accurate trajectory prediction is essential for ensuring
2 Thi-Lich Nghiem et al.

safe and efficient air travel, optimizing airspace utilization, and mitigating congestion
challenges [1].
As air transport data volumes increase, novel analysis techniques and operational
strategies are crucial in ATM for more precise trajectory prediction models. However,
uncertainties, including weather and unknown variables from un-modeled actors, re-
main significant challenges in predicting flight trajectories [2]. Researchers have ex-
plored diverse approaches, from physics-based methods relying on flight dynamics to
data-driven methods using machine learning algorithms to analyze historical data [3].
The first approaches take into account factors such as flight dynamics, aerodynam-
ics, propulsion, and environmental conditions (e.g., wind, air density) to calculate tra-
jectories. They are referred to as conventional approaches. Typically, they are analytical
models to forecast future flight positions based on historical data and predefined rules
like Markov statistical model [4], and the Kalman Filter (KF) [5]. Traditional trajectory
prediction methods often struggle to capture intricate spatial and temporal dependencies
among flights, especially in congested and dynamically changing airspace scenarios or
dealing with complex flight interactions [1].
With growing demands on ATM and the need for more accurate predictions, there
is increasing interest in data-driven approaches to overcome traditional method lim-
itations. Uncertainties in predicting aircraft trajectories, stemming from factors like
changing weather conditions, are addressed by adopting a data-driven approach such
as LSTM [6] and statistical learning [1] that accounts for uncertainty, enabling a com-
prehensive representation of possible futures within the framework [7].
Although having significant achievement in trajectory prediction, data-driven ap-
proaches are susceptible to overfitting and lacking the ability to generalize well to
novel and unseen flight patterns [3, 7]. Additionally, information about the aircraft’s
position, altitude, and heading is usually processed in a relative space to a reference
point. Therefore, coordinate transformation becomes crucial for a precise understand-
ing and prediction of aircraft trajectories. In this paper, we enhance flight trajectory
prediction accuracy by integrating Bayesian theory and Coordinate transformation in a
proposed model called BayesCoordLSTM. The primary contributions of this study can
be summarized as follows:

– Coordinate transformation: Coordinate transformation adoption significantly en-

hances the model’s spatial awareness, enabling more accurate predictions by effec-
tively relating positions and headings to a reference point in aviation trajectories.
– Bayesian Integration: By incorporating Bayesian theorem principles, our model
handles uncertainty in predictions, providing probabilistic trajectory forecasts and
confidence levels. This comprehensive approach enhances predictive accuracy, op-
timizes hyperparameters.
– Empirical Validation: We test on real dataset and compare the prediction per-
formance of our proposed model against other existing methods, including 3D-
CNN, CNN-Gated Recurrent Unit (GRU), and combined CNN-GRU with a 3-
Dimensional Convolution (CG3D) models. The analysis highlights the superior
accuracy and robustness of our proposed model in flight trajectory predictions,
demonstrating its practical utility in ATM and aviation operations.
Improving Flight Trajectory Predictions with Bayesian-Optimized ConvLSTM Models 3

The remainder of our paper follows this structure: Section II presents a a brief lit-
erature review on flight trajectory prediction. Section III illustrates our proposed model
and the experimental results of a real flight dataset are presented in Section IV. Finally,
Section V summarizes the conclusions and outlines future research directions.

2 Related works

Flight trajectory prediction is a critical aspect of ATM to forecast a flight’s future path
based on its historical data. Substantial research efforts have been dedicated to develop-
ing effective prediction techniques to enhance flight safety, efficiency, and overall ATM.
Two primary methodologies employed for trajectory prediction are physics-based meth-
ods and data-driven approaches.
Physics-based methods rely on classical mechanics and aerodynamics principles to
model flight motion. These methods use mathematical equations to describe the dy-
namics of a flight, taking into account forces like lift, drag, thrust, and gravity [1]. In
this approach, the trajectory prediction is based on the flight’s dynamics model, mo-
tion characteristics, and the Markov property of its state [1]. This approach is highly
solvable and performs well in short-term predictions, especially for large flights with
relatively stable motion parameters. The next moment’s state of the flight is closely
linked to its current state, following the principles of Markov theory [4]. Physics-based
methods often integrate theories related to flight performance, dynamics, and various
interacting models [1]. Some researchers used KF to estimate future trajectories based
on noisy measurements [5]. While these approaches provide valuable insights into flight
dynamics and are often used to estimate trajectories under controlled and predictable
conditions; it has been documented that as the prediction time span increases, the error
in these state estimation models tends to escalate rapidly [1].
Therefore, in recent years, data-driven approaches have gained popularity, leverag-
ing historical flight data and other relevant variables to learn patterns and relationships.
These patterns are used to create a data-driven model, which is trained to represent
the collective behavior of the trajectories between input features (e.g., past positions,
velocities) and future trajectory predictions using machine learning models such as
LSTM [8]. While LSTM can capture features in the time dimension, it may not ad-
equately address spatial features. In this context, CNN is more suitable for extract-
ing spatial features. Hence, a hybrid CNN and LSTM [9, 10] or combined CNN and
GRU [11, 12] approaches have been widely used in classification and prediction tasks.
In the context of flight trajectory prediction, uncertainty arises play (e.g., weather con-
ditions or unpredictable aircraft behaviors), in the data itself, or in the optimal values
of model parameters. Failure to capture this information can lead to models that lack
the capability to provide probabilistic predictions and associated confidence levels, ulti-
mately reducing the accuracy of predictions. This can have significant implications for
safety and performance in the aviation industry.
To deal with this problem, the Bayesian framework is adaptable to any prior ranges,
avoids expensive computations, implements the LSTM model to develop continuous-
time responses, and can effectively handle large datasets. So, many researchers have
proposed the application of Bayesian inference [13] such as in Graph Transformer [2],
4 Thi-Lich Nghiem et al.

CNN-GRU [14]. This approach not only estimates the model’s uncertainty but also
optimizes predictions to enhance their performance [15, 16].
However, in a 3D space, as aircraft move, information about their position, altitude,
and heading is processed relative to a reference point. Without performing the coordi-
nate transformation, the model may not effectively grasp the real spatial understanding
and cannot depict the complex variations in flight trajectories accurately [17]. This di-
minishes the accuracy of aircraft position predictions and the reliability of those predic-
tions. Therefore, integrating Bayesian methods and Coordinate transformation into the
aforementioned models strengthens their ability to handle flight trajectory predictions.
This combination allows the model to enhance their ability to manage uncertainty. It not
only aids in ATM but also contributes to improved safety and efficiency, and optimizes
model parameters, ensuring better and more stable performance.
By learning from historical data, data-driven approaches can adapt to changing
conditions, improve accuracy and reliability in estimating flight paths. Nevertheless,
challenges such as generalizing in sparse data scenarios, unpredictable weather con-
ditions, dynamic airspace structures, and the need for real-time prediction remains to
be addressed. To tackle these challenges, we resort to the integration of physics-based
knowledge with data-driven techniques that provide better prediction accuracy [3].

3 Proposed model
We propose a hybrid model named BayesCoordLSTM that is presented in Fig. 1. This
model comprises two primary components: a CNN for spatial feature extraction and
a Bayesian LSTM for capturing long-term temporal information. In the context of

Output

FC
Layer
Latitude
Longtitude Bayesian LSTM t+2M
Altitude t+M+1
Flight LSTM LSTM LSTM LSTM
Normalization Predicted
trajectory Coordinate Trajectories
Transformation
R
θ
h CNN
t+M
t
4 -D trajectories

Fig. 1. The BayesCoordLSTM framework proposed for predicted flight trajectory.

this research, the input and the output of our model in the flight trajectory prediction
problem are given as follows:

– Input:
• X ={Xi }, i ∈ [1, N], where N is the number of flight trajectories.
• Xit = Pit , Pit+1 , . . . , Pit+M is a list of M points of flight trajectory and Pit is the
point of the observed flight trajectory i at time t;
• Pit = (longti , latit , altit , vti , φit ) where each point has five components, correspond-
ing with longitude, latitude, altitude, speed, and heading of the flight.
Improving Flight Trajectory Predictions with Bayesian-Optimized ConvLSTM Models 5


– Output: Pit+M+1 , Pit+M+2 , . . . , Pit+2M is the predicted flight trajectories.

For example, assume that we use 100 points to predict five points in the future,
so
that M = 100. For flight i, starting at time t = 1, we have Pit = Pi1 , Pi2 , . . . , Pi100 . The


output is then Pi101 , Pi102 , . . . , Pi105 , corresponding to five points in the future.
Initially, the CNN layer is responsible for extracting high-level spatial features. This
involves a series of steps including convolutional layers to capture important features
and learn relevant filters, ReLU (Rectified Linear Unit) activation for non-linear trans-
formation, and Max pooling for feature compression and to prevent overfitting. Subse-
quently, the low-dimensional features extracted by the CNN are passed on to the LSTM
layer for the extraction of long-term temporal features. Bayesian inference is employed
in this stage to fine-tune the hyperparameters and architecture of the model. Addition-
ally, it helps quantify uncertainty in the model’s weights by sampling them from a dis-
tribution parameterized by trainable variables during each feed-forward operation. This
comprehensive approach ensures efficient feature extraction, tackles model uncertainty,
and optimizes performance. Lastly, a fully connected layer is introduced to make the
final output prediction.
In our proposed model, we outline the dimensional adjustments of both input and
output data for every layer. Initially, the input data comprises a three-dimensional tensor
with dimensions (None, 100, 6), where "None" denotes the number of batch samples
during model training. The detailed information is described in Table 1.

Table 1. Basic parameters of proposed model.

Layers Input size Output size Parameters
Convolutional (None, 100, 6) (None, 100, 32) 1x3 kernels, 32 filters
ReLU (None, 100, 32) (None, 100, 32) -
Max Pooling (None, 100, 32) (None, 50, 32) Window size: 2
Dropout (None, 50, 32) (None, 50, 32) Dropout rate: 0.3
LSTM (None, 50, 32) (None, 50) 50 units
Fully Connected (None, 50) (None, 4) 4 nodes

In the context of Bayesian LSTM, we use the general formula for sampling weights W
and biases b at the t th time step of the N layer:
   
(t) (t) (t) (t) (t) (t)
W(n) = N (0, 1) ∗ log 1 + ρ(w) + µ(w) and b(n) = N (0, 1) ∗ log 1 + ρ(b) + µ(b) ,

where ρ , µ represent the standard deviation and mean of the input feature, respec-
tively.
Bayesian optimization utilizes Bayesian inference to compute the posterior distri-
bution of the objective function, aiding in the selection of hyperparameters for LSTM
models. It iteratively refines the model based on past data, optimizing the output. If re-
sults are unsatisfactory, a reassessment of hyperparameters and architecture is prompted;
otherwise, successful outcomes inform future predictions using the obtained weights.
In our proposed model, Bayesian optimization is pivotal for identifying critical LSTM
hyperparameter values. It can be seen in Fig. 2.
6 Thi-Lich Nghiem et al.

Data Testing set

Hyperparameters &
Training set architecture

Long Short-Term Memory

Bad
Bayesian optimization

Predictability

Evaluate model

Fig. 2. Optimizing hyperparameters using Bayesian theorem.

4 Experimental Results
4.1 Data preparation
Dataset Our analysis uses the flight data is collected from the Hartsfield–Jackson
Atlanta International Airport (ATL) [11], spanning historical records from March 30,
2016, to March 30, 2020. The detailed attributes of each trajectory are outlined in Ta-
ble 2. It is noteworthy that records sharing the same identifier correspond to the same
aircraft, with a collection of such records collectively forming the trajectory Xi . In the
context of this research, we leverage the entirety of the trajectory data sourced from
geographical coordinates linked to the ATL airport. Each trajectory’s track points are
logged at one-second intervals.
Table 2. Flight trajectories’ features.
Features Unit An example of Trajectory
Timestamp Unix 1537009474
Flight ID Icao24 e8044e
Longitude Degree -84.402633978396
Latitude Degree 33.7542572021484
Altitude Feet 12275.82
Speed Knot 260.53122347432
Heading Degree 12.1974800580645
Hour Unix 1537009200

The dataset comprises both static and dynamic data, with the experiment focusing
on dynamic data such as heading, speed, and 4D data (timestamp, longitude, latitude,
and altitude) and data updates occur every 5 seconds. To manage the spatial range of
Improving Flight Trajectory Predictions with Bayesian-Optimized ConvLSTM Models 7

input data and improve forecast accuracy and smoothness, a sliding window approach
with a window size of 100 and a step size of 1 is employed. The data is segmented into
105-second intervals, where the initial 100 seconds provide historical data and the final
5 seconds serve as the predicted range. This results in 21258 trajectories, which are then
divided into training and test sets, including 17006 and 4252 samples in training and
testing sets, respectively.

– Training set: LSTM model is trained using Bayesian optimization to adjust parame-
ters iteratively, minimizing prediction error, and leveraging probability distributions
to represent uncertainty and enhance effectiveness, particularly with outliers.
– Testing set: Trained LSTM model is evaluated on a testing dataset to assess per-
formance on unseen data. Predictions are compared with ground truth values using
metrics like RMSE and MAE to quantify accuracy, helping gauge the model’s abil-
ity to generalize and providing insights into overall performance.

Coordinate transformation In the trajectory prediction framework, both the input and
output trajectories are converted into discretized vectors to ensure numerical stability
in deep learning models. In flight trajectories, latitudes, longitudes, and altitudes span
different value ranges. Achieving uniformity can be accomplished by normalizing these
variables within a range defined by the minimum and maximum values specific to a
particular airport. However, this approach lacks generalizability across all airports due
to varying normalization factors for latitudes and longitudes. As a solution, trajectories
are transformed into a cylindrical coordinate system, aligning flight states relative to
the airport’s position [18]. Given, flight position x f p = (long f p , lat f p , alt f p ) and airport
position xap = (longap , latap , altap ), we can transform coordinate as follows:
!
xfp − xap longfp − longap altfp − altap
R= , θ = arcsin , h= ,
Dmax xfp − xap Amax

where:
– Dmax is the distance extending from the airport and encompassed by the predictive
scope of the framework;
– Amax corresponds to the highest achievable altitude within this predictive domain;
– R is the distance to the airport that has been normalized;
– θ is the angle formed by the relative position vector with the north direction;
In our paper, we choose Dmax = 328084 feet and Amax = 39370 feet for predicting
aircraft trajectories, balancing computational efficiency and practicality. These values
provide sufficient coverage while conserving computational resources and align with
typical aircraft operating ranges and altitudes, ensuring relevance to real scenarios.
Before coordinate transformation, flight trajectory data exhibits significant diver-
sity in terms of position, altitude, and heading, making it challenging to process and
predict. However, after converting the coordinates relative to the ATL airport, all trajec-
tory data becomes relatively consistent and directly related to the reference point. This
enhances predictive capability while reducing complexity in data processing and inte-
gration into prediction models. In Fig. 3, the left image illustrates flight positions before
8 Thi-Lich Nghiem et al.

transformation, which are uneven and complex due to the Earth’s spherical coordinate
system. For instance, flight ’a5e827’ demonstrates this complexity with its trajectory
being particularly challenging to track and predict due to large variances in spatial and
coordinate values. The right image illustrates trajectories simplified and standardized
through cylindrical coordinate transformation, aligning them with the airport. This uni-
form representation enhances the utilization of machine learning models by normalizing
the data, thereby improving their ability to forecast trajectories.

a492c2 a492c2
adf531 adf531
adfece adfece
acde77 acde77
a5e827 a5e827

12000
10000

10000 9000

Altitude
8000 8000
Z
6000 7000

6000
4000

5000
2000

33.84 10

33.82 5
−84.48 33.80 6000 0
e

Y
ud

−84.46 8000
tit

−84.44 33.78 −5
La

10000
Long
itud−84.42
e 33.76
X 12000 −10
−84.40 14000
−84.38 16000 −15

Fig. 3. An example of five flights at ATL airport before (left) and after (right) transform-
ing coordinate.

Normalization In this paper, we use Z-Score normalization, features are transformed

to become standardized as follows: Xnew = (X − min)/(max − min) where Xnew is the
normalized data, obtained by scaling the original data to fit within a standardized range.

4.2 Evaluation metrics

Two common metrics that are Root Mean Square Error (RMSE) and Mean Absolute
Error (MAE) are employed. They are respectively defined by:
s
1 n  bt 2 1 n bt
RMSE = ∑ i i
n t=1
P − P t and MAE = ∑ Pi − Pit .
n t=1

In this equation, n represents the number of predicted points, while Pbit and Pit corre-
spond to the predicted and actual values at time step t of flight i.

4.3 Comparison of experimental results

We evaluate the proposed method on ATL dataset [11] that is presented in Section 3.
In order to compare our proposal method with three models, including 3D-CNN [19],
CNN-GRU [20], and CG3D [11]. Our experimental results are illustrated in Table 3.
The reported RMSE values for the predictive models showcase their respective ac-
curacy in making predictions. Among the models evaluated, the CNN-GRU model ex-
hibited the highest RMSE value of 0.3728, signifying a larger discrepancy between
Improving Flight Trajectory Predictions with Bayesian-Optimized ConvLSTM Models 9

Table 3. Comparing our proposed method with the state of the art methods.
Predicting models RMSE MAE
CNN-GRU [20] 0.3728 0.2164
3D CNN [19] 0.2646 0.1785
CG3D [11] 0.2626 0.1776
Our proposed method 0.2154 0.1624

predicted and actual values. Similarly, its MAE of 0.2164 indicates relatively higher
prediction errors. The 3D CNN model showed improvement with an RMSE of 0.2646
and MAE of 0.1785, indicating a more accurate predictions than the CNN-GRU. No-
tably, the CG3D model achieved an even lower RMSE of 0.2626 and MAE of 0.1776,
positioning it as one of the top-performing models in terms of prediction accuracy.
In our proposed model, it can be attributed to two key factors: the use of coordinate
transformation and Bayesian inference. By implementing coordinate transformation,
which facilitates consistent data normalization, we effectively reduced the impact of
outliers. Furthermore, the incorporation of Bayesian inference helped prevent overfit-
ting by using probability distribution during training, thereby improving the model’s
generalization ability. This combination of techniques enabled our proposed model to
achieve more accurate predictions. As seen in Table 3, our method achieved an RMSE
of 0.2154 and MAE of 0.1624, demonstrating its superior predictive accuracy compared
to the other models.

5 Conclusions and Future works

Within this research paper, we introduced a new approach to predict flight trajectory,
named BayesCoordLSTM. This model efficiently integrates Bayesian optimization into
a hybrid ConvLSTM framework, delivering accurate flight path forecasts with im-
proved precision and efficiency. Our experiments, conducted using real flight trajec-
tory datasets, clearly demonstrate the remarkable superiority of the BayesCoordLSTM
model over existing methods. Across various flight scenarios and conditions, this model
consistently outperforms baseline models in terms of trajectory prediction accuracy.
The utilization of Bayesian optimization not only streamlines hyperparameter tuning
but also yields promising results in terms of predictive performance enhancement. Al-
though our study lays a foundation for predicting flight trajectories, further research
could explore dynamic hyperparameter adaptation techniques during the prediction
process. These methods could involve real-time adjustments to hyperparameters in
response to evolving flight scenarios or changing weather conditions, enhancing the
BayesCoordLSTM model’s adaptability and real-time predictive capabilities.

References
1. W. Zeng, X. Chu, Z. Xu, Y. Liu, and Z. Quan, “Aircraft 4D Trajectory Prediction in Civil
Aviation: A Review,” Aerospace, vol. 9, no. 2, p. 91, Feb. 2022.
10 Thi-Lich Nghiem et al.

2. Y. Pang, X. Zhao, J. Hu, H. Yan, and Y. Liu, “Bayesian Spatio-Temporal grAph tRansformer
network (B-STAR) for multi-aircraft trajectory prediction,” Knowledge-Based Systems, vol.
249, p. 108998, Aug. 2022.
3. H. Li, H. Jiao, and Z. Yang, “AIS data-driven ship trajectory prediction modelling and anal-
ysis based on machine learning and deep learning methods,” Transportation Research Part
E: Logistics and Transportation Review, vol. 175, p. 103152, Jul. 2023.
4. X. Wang, X. Jiang, L. Chen, and Y. Wu, “KVLMM: A trajectory prediction method based
on a variable-order Markov model with kernel smoothing,” IEEE Access, vol. 6, pp. 25 200–
25 208, 2018.
5. T. Zhang, Y. Gao, and C. Zhang, “Short-term 4D trajectory prediction based on KF joint
EKF parameter identification,” Journal of Civil Aviation University of China, vol. 34, no. 5,
pp. 1–4, 2016.
6. P. Jia, H. Chen, L. Zhang, and D. Han, “Attention-LSTM Based Prediction Model for Aircraft
4-D Trajectory,” Scientific reports, vol. 12, no. 1, pp. 1–11, 2022.
7. Y. Pang, X. Zhao, H. Yan, and Y. Liu, “Data-driven trajectory prediction with weather un-
certainties: A Bayesian deep learning approach,” Transportation Research Part C: Emerging
Technologies, vol. 130, p. 103326, Sep. 2021.
8. W. Zeng, Z. Quan, Z. Zhao, C. Xie, and X. Lu, “A deep learning approach for aircraft trajec-
tory prediction in terminal airspace,” IEEE Access, vol. 8, pp. 151 250–151 266, 2020.
9. L. Ma and S. Tian, “A Hybrid CNN-LSTM Model for Aircraft 4D Trajectory Prediction,”
IEEE Access, vol. 8, pp. 134 668–134 680, 2020.
10. Q. Li, X. Guan, and J. Liu, “A CNN-LSTM framework for flight delay prediction,” Expert
Systems with Applications, vol. 227, p. 120287, Oct. 2023.
11. H. Shafienya and A. C. Regan, “4D flight trajectory prediction using a hybrid deep learning
prediction method based on ADS-B technology: A case study of hartsfield–jackson Atlanta
international airport (ATL),” Transportation Research Part C: Emerging Technologies, vol.
144, pp. 1–12, 2022.
12. P. N. Tran, H. Q. V. Nguyen, D.-T. Pham, and S. Alam, “Aircraft Trajectory Prediction With
Enriched Intent Using Encoder-Decoder Architecture,” IEEE Access, vol. 10, pp. 17 881–
17 896, 2022.
13. X. Zhang and S. Mahadevan, “Bayesian neural networks for flight trajectory prediction and
safety assessment,” Decision Support Systems, vol. 131, p. 113246, Apr. 2020.
14. J. Huang and W. Ding, “Aircraft Trajectory Prediction Based on Bayesian Optimised Tem-
poral Convolutional Network–Bidirectional Gated Recurrent Unit Hybrid Neural Network,”
International Journal of Aerospace Engineering, vol. 2022, pp. 1–19, Dec. 2022.
15. M. Abdar, F. Pourpanah, S. Hussain, D. Rezazadegan, L. Liu, M. Ghavamzadeh, P. Fieguth,
X. Cao, Khosravi et al., “A review of uncertainty quantification in deep learning: Techniques,
applications and challenges,” Information fusion, vol. 76, pp. 243–297, 2021.
16. M. Atencia, R. Stoean, and G. Joya, “Uncertainty quantification through dropout in time
series prediction by echo state networks,” Mathematics, vol. 8, no. 8, 2020.
17. S. Varun S., “Deep learning framework for trajectory prediction and in-time prognostics
in the terminal airspace,” Ph.D. dissertation, Purdue University Graduate School, 2022.
[Online]. Available: https://doi.org/10.25394/PGS.21256119.v1
18. G. Cai, B. M. Chen, and T. H. Lee, “Coordinate systems and transformations,” Unmanned
rotorcraft systems, pp. 23–34, 2011.
19. W. Qin, J. Tang, and S. Lao, “DeepFR: A trajectory prediction model based on deep feature
representation,” Information Sciences, vol. 604, pp. 226–248, 2022.
20. H. Shafienya and A. Regan, “4D flight trajectory prediction based on ADS-B data: A com-
parison of CNN-GRU models,” in IEEE Aerospace Conference (AERO), 2022, pp. 01–12.