Received December 31, 2021, accepted January 24, 2022, date of publication February 4, 2022, date of current version

February 17, 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3149231

Aircraft Trajectory Prediction With Enriched

Intent Using Encoder-Decoder Architecture
PHU N. TRAN , HOANG Q. V. NGUYEN, DUC-THINH PHAM , AND SAMEER ALAM
School of Mechanical and Aerospace Engineering, Air Traffic Management Research Institute, Nanyang Technological University, Singapore 639798
Corresponding author: Sameer Alam ([email protected])
This work was supported in part by the National Research Foundation, Singapore; and in part by the Civil Aviation Authority of Singapore
under the Aviation Transformation Program.

ABSTRACT Aircraft trajectory prediction is a challenging problem in air traffic control, especially for
conflict detection. Traditional trajectory predictors require a variety of inputs such as flight-plans, aircraft
performance models, meteorological forecasts, etc. Many of these data are subjected to environmental
uncertainties. Further, limited information about such inputs, especially the lack of aircraft tactical intent,
makes trajectory prediction a challenging task. In this work, we propose a deep learning model that performs
trajectory prediction by modeling and incorporating aircraft tactical intent. The proposed model adopts the
encoder-decoder architecture and makes use of the convolutional layer as well as Gated Recurrent Units
(GRUs). The proposed model does not require explicit information about aircraft performance and wind data.
Results demonstrate that the provision of enriched aircraft intent, together with appropriate model design,
could improve the prediction error up to 30% at a prediction horizon of 10 minutes (from 4.9 nautical miles
to 3.4 nautical miles). The model also guarantees the mean error growth rate with increasing look-ahead
time to be lower than 0.2 nautical miles per minute. In addition, the model offers a very low variance in
the prediction, which satisfies the variance-standard specified by EUROCONTROL (EU Organization for
Safety and Navigation of Air Traffic) for trajectory predictors. The proposed model also outperforms the
state-of-the-art trajectory prediction model, where the Root Mean Square Error (RMSE) is reduced from
0.0203 to 0.0018 for latitude prediction, and from 0.0482 to 0.0021 for longitude prediction in a single
prediction step of 15 seconds look-ahead. We showed that the pre-trained model on ADS-B data maintains
its high performance, in terms of cross-track and along-track errors, when being validated in the Bluesky Air
Traffic Simulator. The proposed model would significantly improve the performance of conflict detection
systems where such trajectory prediction models are needed.

INDEX TERMS Aircraft trajectory prediction, 4D trajectory, machine learning, encoder-decoder, convolu-
tion neural network, recurrent neural network.

I. INTRODUCTION conflict alert (STCA), etc [2]. For example, improvement in

A. AIRCRAFT TRAJECTORY PREDICTION the accuracy of trajectory prediction in the look-ahead time of
Aircraft trajectory prediction is a crucial component of any approximately 4-8 minutes could potentially help to reduce
air traffic control (ATC) system. Aircraft trajectory is defined STCA’s nuisance alerts, which in turns enhances the system
as ‘‘a four dimensional (e.g., latitude, longitude, altitude and overall efficiency. In the Trajectory-Based Operations (TBO)
time) description of an aircraft’s flight path’’ [1]. Trajectory concept [3], which is envisioned as the key enabler for the
prediction refers to the estimation of a flight’s future tra- next generation of air traffic management, trajectory predic-
jectory within a look-ahead time (prediction horizon) [1]. tion is an essential building block for the Decision Support
Accurate aircraft trajectory prediction not only improves sit- Tools (DST) for ATCOs.
uational awareness of air traffic control officers (ATCOs), Traditional aircraft trajectory predictors compute the air-
but also provides necessary inputs for other ATC functional- craft future trajectories using diverse information such as
ities such as departure and arrival management, monitoring flight-plan, aircraft performance, meteorological forecast,
aids, medium-term conflict detection (MTCD), short-term ATCOs and flight crews intents.
Major sources of errors in traditional approaches include
The associate editor coordinating the review of this manuscript and diversity in aircraft performance, uncertainties in input data,
approving it for publication was M. Shamim Kaiser . ATCOs intents, and in longitudinal progress (Figure 1a) [4].

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 10, 2022 17881
 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

Prediction errors also resulted from limited accuracy of ordinary least square and neural network were employed
numerical integration algorithms and simplified aircraft’s in [17] for the task of predicting altitude of the aircraft, which
equations of motion [5]–[7]. Among the factors that hin- showed that both regression approaches perform significantly
der the trajectory prediction, uncertainties in the intents of better than the point-mass model. In [18], the authors used
ATCOs is recognized as the most significant one [4], which Generalized Linear Model (GLM) to predict aircraft’s arrival
are influenced by the dynamics of traffic flow, weather con- time at significant points, and the model can also be used for
ditions, operational constraints, etc. Figure 1b illustrates the spacing the aircraft in continuous descent operation. Multiple
flown trajectory of flights from Kuala Lumpur International regression models approach was also considered for trajec-
Airport to Singapore Changi Airport. Although those flights tory prediction in [19], in which aircraft trajectories in Termi-
were planned to follow a fixed route, there was a high nal Maneuvering Area (TMA) were classified by their distinct
variances in the realizations of the planned routes making patterns, and different regression models were developed for
trajectory prediction a challenging task. different groups of trajectory to predict the aircraft arrival
In the past few decades, the rapid advances in machine time. Such approach, i.e., trajectory clustering combined with
learning algorithms have offered new approaches to improve multiple predictive models, were also employed for trajectory
the performance of computational models in solving classical prediction in [20]–[23]. Other recent studies on the employ-
problems in air traffic management [8]. Approaches to the air- ment of different machine learning algorithms for trajectory
craft trajectory prediction are also being transformed by the prediction include Bayesian deep neural networks [24], [25],
paradigm shift towards data-driven methods using machine variational inference [26], conditional generative adversar-
learning techniques [9], [10]. As mentioned earlier, the lack ial network [27], deep Gaussian process [28]. In addition,
of information about aircraft intent is a primary source of a hybrid machine learning-physics approach was recently
trajectory prediction errors. In this study, we aim to model proposed [29], in which an estimation algorithm (Residual-
and incorporate aircraft intent in a deep learning model for Mean Interacting Multiple Model) was introduced to improve
trajectory prediction. The aircraft intent used in our work is a machine learning models by accounting for the motion of
modeled from the historical flown trajectories of the aircraft, the aircraft.
which in turn is derived from the Automatic Dependent- Deep learning has improved the state-of-the-art in many
Surveillance Broadcast (ADS-B) data. In this work, we make problems that challenge the conventional machine learning
use ADS-B data as the only data source, without other salient methods such as speech recognition, visual object recog-
features such as wind and weather information. As for short- nition, object detection. For processing sequential inputs,
term trajectory prediction, the effect of wind and weather is recurrent neural network (RNN) is often a better approach.
considerably stable and is reflected in the actual track data, However, traditional RNN has problems of learning long-
which can be learned by the proposed model as latent vectors. term dependencies and vanishing gradient. Long Short Term
In the following section I-B a review of the related work Memory [30] (LSTM), a type of RNN, was introduced to
is presented followed by details of research objectives and overcome these problems and have been used to advance the
research contribution in Section I-C. state of the art for many difficult problems such as hand-
writing recognition and generation, language modeling and
B. RELATED WORK translation. A simpler variant of LSTM, Gated Recurrent
Aircraft trajectory prediction is specifically important to air- Unit (GRU) [31] was introduced to deal with the vanishing
craft separation assurance, which includes conflict detection problem and long-term depencies, which combines the forget
and resolution. Conflict detection algorithms can be classi- and input gates into a single ‘‘update grate,’’ and provides
fied into nominal, worst-case and probabilistic approaches the similar performance on certain tasks such as speech sig-
[11], [12]. Among these, the probabilistic approach is a nal modeling or natural language processing, with LSTM.
trade-off between detection accuracy and computational time These RNN models can be used as modules inside encoder-
required under various uncertainties. Probabilistic conflict decoder sequence-to-sequence architecture [32] for problems
detection requires as input the predicted positions of the with both sequential input and output. This encoder-decoder
aircraft with cross-track and along-track errors, by which one architecture was used for solving human mobility trajec-
can estimate the probability of the distances (both vertical and tory prediction [33] or for the vehicle trajectory predic-
lateral) between any two aircraft to be lower than the standard tion problem [34] and was shown to outperform other non
separations [13], [14]. Therefore, the estimation of trajectory encoder-decoder methods. RNN was also combined with
prediction errors plays a key role in the provision of an graph model in a structural-RNN framework in [35], which
accurate conflict detection. The increasing availability of can learn the trajectory patterns of different agents (pedes-
aircraft historical trajectory data (e.g. ADS-B) and advanced trians, bicycles or cars). The approach based on 4D graph
techniques in trajectory mining and pattern analysis [15], [16] (two dimensions for instances and their interactions, one
have been contributing to the enrichment of trajectory data, for time series and one for high-level categorization), and
which allows further improvements of prediction models. showed a improvement of 20% of accuracy over previous
Various machine learning algorithms have been considered models. LSTM and attention mechanism [36] were integrated
for aircraft trajectory prediction. Regression methods using within an encoder-decoder architecture in [37] to perform

17882 VOLUME 10, 2022

P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

FIGURE 1. (a) Cone of uncertainty in aircraft future position. The color gradient indicates degree of uncertainty, darker color means higher
uncertainty. Uncertainty in aircraft position increases with look-ahead time. (b) High variances in the actual flown tracks of flights from
Kuala Lumpur International Airport (WMKK) to Singapore Changi Airport (WSSS). The thick red curves are planned routes.

geo-sensory time series prediction. The attention mechanism also demonstrated to be superior to regression methods for
included local spatial, global spatial and temporal attentions trajectory prediction task [45]. Dynamic physical constraints
(multi-level attention-based recurrent neural network), which was also taken into consideration within a LSTM network
outperformed nine baseline models. [38] proposed a Hierar- framework in [46], which showed the capability of integrat-
chical Spatial-Temporal Long-Short Term Memory Network ing both long-term dependencies and dynamic constraints
(HST-LSTM) for location prediction, in which a hierarchical in trajectory prediction. Another emerging approach is the
extension was used for embedding the users’ visiting context. vision-based approach, in which aircraft trajectory prediction
The results showed the effectiveness of the approach on is achieved using convolutional LSTM on a series of traffic
real-world dataset. A combination of attention mechanism scenario images [47]. Such approach is benefited from the
and LSTM network for prediction of human location was automated feature extraction offered by the convolutional
proposed in [39], which extended the LSTM network with layers and the input dimension is independent of the scenario
attention-based sequence-to-sequence for learning human complexity (e.g., the number of aircraft).
semantic trajectories, which was shown to be beneficial to We observed that different approaches using deep learn-
modelling semantic trajectories and prediction of human ing, specifically recurrent neural network, had been recently
movement patterns. [40] proposed two variants of RNN, employed for aircraft trajectory prediction, as those meth-
namely Constrained State Space RNN (CSSRNN) and Latent ods very well fit the spatio-temporal property of trajectory
Prediction Information RNN (LPIRNN), that can incorporate data. Nevertheless, there were limited work that give special
unique trajectory constraints, of which normal LSTM models consideration to model the aircraft intent as input features
are incapable. The proposed variants showed slightly better for deep learning model. Furthermore, some important eval-
results than that of ordinary LSTM networks. uation metrics such as along-track error, cross-track error,
As aircraft trajectories can be naturally modeled as time and the error growth rates overtime had not been properly
sequences of aircraft states, many studies employed sequence compared with available standard, such as the standard spec-
learning techniques, such as RNN, GRU, and LSTM, for ified by EUROCONTROL [2]. In this work, we propose an
trajectory prediction. A LSTM network was proposed in [41] approach overcoming these limitations.
for trajectory prediction and was shown to outperform other
methods such as Markov Model and weighted Markov C. OBJECTIVES & SUMMARY OF CONTRIBUTIONS
Model. GRU neural network was also employed for trajectory The objective of this research is to propose a machine
prediction in [42], which showed accuracy improvements learning model for aircraft trajectory prediction that takes
compared to LSTM network. A combination of LSTM and into consideration the aircraft intent, in a prediction horizon
convolutional layer was investigated in [43], in which the of 1 to 10 minutes.
model can predict aircraft trajectory between any arbitrary The proposed model adopts the Encoder-Decoder
two airports at low variance. Such combination of LSTM architecture [32], which can take a sequence as input and
and convolutional layer was also employed in [44], where generate another sequence as output. This makes Encoder-
the prediction accuracy was shown to be increase by 21% Decoder architecture a suitable choice for trajectory predic-
comparing to using LSTM only. The LSTM network was tion because a trajectory can be naturally expressed as a

VOLUME 10, 2022 17883

 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

time-sequence of spatial positions. Furthermore, the Encoder explanation of different loss functions to achieve the learning
part of the model can extract latent features automatically, objectives is presented. Finally, hyper-parameters and special
which in turns is used by the Decoder to produce a prediction. treatments to improve training robustness in the experiments
The input to the model is the aircraft current trajectory, is discussed.
which is defined as a time sequence consisting of current
position of the aircraft and its k last positions. The output A. AIRCRAFT INTENT MODELING
of the model, or the predicted trajectory, is a time sequence Aircraft intent refers to the list of waypoints that the aircraft
of n positions of the aircraft into the future. Given that all is set to traverse. Intent modeling is to express the relative
trajectories have equal time interval between any two consec- relationship between the aircraft current position and these
utive positions, n also determines the look-ahead time or the waypoints in a form suitable for the learning model. This
prediction horizon. Model architecture is described in detail relative relationship is encoded in a 10-dimensional vector,
in Section II. which is referred to as intent vector It , where the subscript t
In this research, the trajectory prediction is limited to en- indicates the time step.
route phase. The main contributions of this paper include: Figure 3 shows an aircraft at the current time in the context
• An effective way to model aircraft intent for the integra- of three waypoints: previous, current, and next waypoints.
tion into the trajectory prediction model. Experiments The aircraft passed the previous waypoint, is passing by
show that our intent modeling allows the trajectory pre- the current waypoint, and then heading to the next way-
dictor to improve performance at different prediction point. An intent vector includes the four distances c1 and c2
horizons. To the best of our knowledge, such modeling (c1 , c2 ≥ 0), d1 and d2 (d1 , d2 ≥ 0) and the two angles β1 and
of aircraft intent for deep learning prediction model has β2 (−π ≤ β1 , β2 < π). Here, c1 and c2 are the cross-track
not been reported previously. distances from the aircraft to the current and the next airways,
• An encoder network architecture that effectively learns respectively. d1 and d2 are the distances from the aircraft to
patterns in the behavior of aircraft without the need of the current and the next waypoints, respectively. β1 and β2 are
explicit information about the aircraft performance. the angles measured from the instantaneous moving direction
• A decoder network architecture that fuses information of the aircraft to the lines connecting the aircraft to the current
about aircraft future intent into the knowledge about its and the next waypoints, respectively. The intent vector also
recent behavior, and effectively predicts future trajectory includes the lateral directions from the aircraft to the current
of the aircraft. and the next waypoints. The ten elements of an intent vector
• A loss function which consists of two components, are described in Table 1.
namely position loss and dynamics stability loss. The
high accuracy of the prediction model is achieved by TABLE 1. Elements of the intent vector. See Figure 3 for a visual
minimizing this loss. explanation.

• The prediction horizon (i.e., look-ahead time) is inde-

pendent of network architecture and can be dynamically
changed even after the model has been trained.
• The proposed model requires limited amount of infor-
mation. In specific, only historical tracks of the aircraft
(ADS-B data) is necessary for model training.
The aforementioned features enable the proposed model
to outperform existing models using the same approach and Physically, d1 and d2 indicate the progress of actualizing
achieves prediction accuracy that is comparable to the stan- the intent by the aircraft, while c1 and c2 imply the lateral
dard specified by EUROCONTROL for aircraft trajectory deviation of the flown trajectory from the planned route.
prediction [2]. The rest of this paper is organized as follows. In addition, β1 and β2 observe the track angle evolution
Section II elaborates learning model for trajectory prediction. with respect to the three waypoints. These information helps
Data source used in this work and features engineering are the model to discriminate between turning and non-turning
presented in Section III. Experiment setup, results and dis- (maintaining heading) behaviors. The intent vector is dynam-
cussion are presented in Section IV. Finally, conclusions are ically computed during prediction to reflect the changes in the
drawn in Section VI. three waypoints as the aircraft is moving and passing them
one by one. The computation of the intent vector is performed
II. LEARNING MODEL in a local Cartesian coordinates system (see Section III-C2).
This section elaborates the learning model for aircraft Note that the intent modeling is still valid if the current and the
trajectory prediction that incorporates the aircraft intent. next airways form a straight line with intermediate waypoints.
An overview of the proposed model is illustrated in Figure 2.
First, the approach for aircraft intent modeling is presented. B. THE PREDICTION TASK
Second, a description of the data and data preparation steps Let pidenotes the aircraft
 position at the discrete time step i,
(1) (2) (3) (1) (2) (3)
is provided. Next, the learning model architecture, with pi = pi , pi , pi where pi , pi , pi are the coordinates

17884 VOLUME 10, 2022

P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

FIGURE 2. Overview of the learning model. (a) Context of trajectory prediction problem. The proposed model performs trajectory prediction in the spatial
context of three waypoints (previous, current and next waypoints). Note that the altitude dimension is not included in the figure for a simple
presentation. (b) Conceptual illustration of the prediction model. The model is trained using the train dataset (upper part) and the trained model
performs predictions using the test dataset (lower part).

C. MODEL ARCHITECTURE
The learning model adopts the encoder-decoder architec-
ture [32], which particularly suits our problem for several
reasons. First, the encoder-decoder architecture allows to
conveniently handle the difference in length of the input and
output data. In fact, the model takes as input a fixed-length
vector of aircraft recent positions and produces a variable-
length vector of aircraft future trajectory. Second, the internal
structure of the encoder and that of the decoder are indepen-
dent of each other. Thus, they can be designed to perform
different tasks. As such, the encoder is trained to recognize
hidden information and patterns in the behaviors of the air-
craft. These knowledge is output by the encoder as a fixed-
length context vector. After that, the decoder dynamically
FIGURE 3. Aircraft intent modeling. Intent of the aircraft are encoded by concatenates the context vector and the intent vector, and
the distances c1 , c2 , d1 , d2 and the angles β1 , β2 , in the spatial context of
three waypoints.
produces predicted trajectory. The overall architecture of our
model is illustrated in Figure 4. Details about the encoder and
the decoder are described below.
of the aircraft in a three-dimensional Cartesian reference
frame. Assuming that pt is the aircraft position at the cur-
rent time t, the last k positions are [pt−k , pt−k+1 , . . . , pt−1 ]. 1) ENCODER
The next n positions, i.e., future trajectory, of the aircraft is The encoder is designed to learn and recognize hidden pat-
defined as Tt n = [pt+1 , pt+2 , . . . , pt+n ] (Figure 2a). Given terns in the recent trajectory of the aircraft. Those hidden
the current and the last k positions of the aircraft, the task is to patterns carry information about aircraft dynamics character-
predict its future trajectory. In other words, we train a model istics under different circumstances, which are essential to the
that receives a vector [pt−k , pt−k+1 , . .. , pt−1 , pt ] as an input prediction of future trajectory. To achieve this, the encoder
and produces a prediction T̂t n = p̂t+1 , p̂t+2 , . . . , p̂t+n consists of one one-dimensional convolution (Conv1D) layer
as the output. The model also receives intent vector as an and one fully connected layer, each followed by a ReLU
input. The learning objective is to minimize the difference activation (Figure 4).
between the prediction T̂t n and the actual trajectory Tt n . In the Convolutional layer makes use of convolution opera-
final evaluation settings, we set k = 9 and n = 40. This tor in place of general matrix multiplication [48, p. 224].
means the model receives an input sequence of 10 last posi- Some important features of convolutional layers are weights
tions (including the aircraft current position) and produces an sharing and their ability to capture the local connectivity.
output sequence of 40 positions. These features make convolutional layers a natural choice

VOLUME 10, 2022 17885

 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

FIGURE 4. Encoder-decoder architecture for trajectory prediction. The decoder is unrolled for illustration of n sequential prediction steps (only three
steps are visible in the figure).

T̂t n = p̂t+1 , p̂t+2 , . . . , p̂t+n . At each prediction step,

 
for patterns recognition in time sequence data like aircraft
trajectory. in general, the GRU receives the current input and the pre-
The output from the encoder, which is defined as the vious hidden state, and produces a GRU output. This GRU
context vector, has a length of lc . Here, lc is a hyper-parameter output is then employed in two ways. First, it is concatenated
and determined experimentally. The context vector carries with the context vector and then input to a fully connected
essential information about the current aircraft dynamics and layer in order to produce the aircraft position prediction for
other uncertainties such as wind effect. As the prediction the current step. Second, it is used as the GRU hidden state
horizon considered in this work is relatively short, changes for the subsequent prediction step (see Figure 4).
in the wind field and other uncertain factors are insignificant. The input to the GRU at every step is illustrated in Figure 4.
As a result, the learned information in the context vector can At a prediction step i, for instance, the GRU input is from the
be used in the prediction of the aircraft future locations. concatenation of four vectors:
The parameters of the encoder are presented in Table 2. • the last predicted position of the aircraft p̂i−1 ,
• the last predicted velocity of the aircraft v̂i−1 = p̂i−1 −
2) DECODER p̂i−2 ,
In the proposed model, the decoder is a Recurrent Neural Net- • the current intent vector as described in Section II-A, and
work (RNN). Particularly, RNN is a class of neural networks • the context vector produced by the encoder.
where connections between nodes form a directed graph The concatenation of these four vectors forms a GRU input
along a sequence. RNN is designed to recognize sequential vector of length (lc + 16), where lc is the length of the context
patterns for inferring the next likely outcome. Figure 5a vector (Table 2). At the first prediction step, the GRU hidden
shows a structure of a RNN. The vanilla RNN has issues of state h0 is initialized to the context vector. The details of
gradient vanishing and exploding when dealing with long- decoder is given in Table 2.
term series. Gated Recurrent Unit (GRU), a type of RNN, was
introduced in [31] to overcome these issues in capturing long- TABLE 2. Parameter settings used in our experiment.
term dependencies. The structure of a GRU cell is demon-
strated in Figure 5b, and its internal computational logic is
described by
zt = σ (Wz · [ht−1 , xt ])



 r = σ (W · [h , x ])

t r t−1 t
(1)


 ĥt = tanh (W · [rt ∗ ht−1 , xt ])
ht = (1 − zt ) ∗ ht−1 + zt ∗ ĥt


where Wz , Wr , W are the coefficient matrices, σ is the 3) LOSS FUNCTION

sigmoid function and ∗ is the element-wise multiplication The model is trained by minimizing the loss that has two
operator. components, namely position loss LtP and dynamics stability
The decoder is designed to perform n-step sequential pre- loss LtS .
diction. Figure 4 illustrates the unrolled GRU network of The position loss LtP measures the deviation of the pre-
the decoder that predicts the future trajectory of n steps, dicted trajectory T̂t n from the actual one Tt n . It is determined

17886 VOLUME 10, 2022

P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

FIGURE 5. a) An unrolled recurrent neural network; b) Internal structure of a Gated Recurrent Unit [31].

by the RMSE of the Euclidean distances between correspond- of parameters 2∗ that minimizes the total loss:
ing points in T̂t n and Tt n , i.e., X
2∗ = arg min L = arg min LtP + αLtS


(5)
v
u n 2 2 Tt
uX
LtP = t d2 p̂t+i , pt+i ,


(2)
III. DATA SOURCE & FEATURES ENGINEERING
i=1
In this section, we present the data source and features
where d(· , ·) yields the Euclidean distance between two engineering, including data pre-processing steps and aircraft
positions. intent reconstruction from the trajectory data.
The dynamics stability loss, as the name suggests, reflects
the stability of the aircraft dynamics. It captures the variation A. INTENT RECONSTRUCTION
in the predicted aircraft velocity. Let v̂i = p̂i − p̂i−1 is the As described in Section II-A, aircraft intent refers the planned
predicted velocity of the aircraft at a time step i. The dynamics route of the aircraft. In this work, only ADS-B data is used,
stability loss of the n-step prediction horizon is determined by without using filed flightplans of the flights presented in the
v data. Thus, we reconstruct the planned routes, or the intent,
u n
uX of the flights from their flown track data.
LtS = t d2 v̂t+i , v̂t+i−1


(3) Intent reconstruction is performed at two levels. On the
i=1 first level, given a set of all designated significant points,
we identify those points that are sufficiently close to the
Note that when i = 1, v̂t+i−1 ≡ v̂t = pt − pt−1 , which is the
flown track of a flight, and consider them as the planned
current velocity of the aircraft and can be determined from
route. Here, designated significant points (or waypoints) are
the model’s input. The total loss is then defined as
specified geographical locations in the airspace used to define
Lt = LtP + αLtS (4) flight routes, and they can be collected from the Electronic
Aeronautical Information Publication (eAIP) of the related
where α is a weight coefficient that determines the contribu- Flight Information Regions (FIRs). For convenience, we refer
tion of the dynamics stability loss to the total loss. to the intent on this first level as ordinary intent.
The inclusion of the dynamics stability loss is beneficial ATCOs often vectors aircraft from their original plans due
in several manners. First, minimizing this loss helps the to reasons such as conflict resolution, congestion manage-
model to learn the aircraft dynamics effectively. Physically, ment, flows optimization, etc. In those situations, however,
the dynamics stability loss measures the fluctuation in the the ATC clearances (in the form of voice instructions) given
predicted aircraft velocity in every time interval. As the tra- to the flight crews are not captured in the ADS-B data. On the
jectory predictor is designed for en-route phase, during which second level of intent reconstruction, we also identify missing
the aircraft velocity is relatively stable, it is desirable that this trajectory change points (i.e., turn location) that were resulted
velocity fluctuation to be insignificant. Second, experiments from ATC clearances, in addition to the designated significant
shows that the second term in the right-hand side of Equa- waypoints. We refer to the intent on this second level as
tion 4 has similar effect of a regularization term in typical enriched intent. Note that to facilitate the enriched intent,
context of machine learning. That means, at an optimal value we first need to identify all turn segments in the trajectory
of α, minimizing the dynamics stability loss contributes to data.
reducing over-fitting and improving generalization of the
model. Here, α is treated as a hyper-parameter that needs to 1) TURN SEGMENT IDENTIFICATION
be determined empirically. We determine the turn segments, noted as St , and non-turn
Let 2 denotes the set of all trainable parameters of the segment, Ss , by the geometric curvature and along a trajec-
model. The learning objective is to determine the optimal set tory. Recall that the curvature of a curve κ and its relationship

VOLUME 10, 2022 17887

 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

FIGURE 6. (a) Radius of a space curve in the calculation of curvature. (b) The determination of additional trajectory change
point for enriched intent. The addition trajectory change point is randomly located in the green circle.

with the curve radius R are given by: where dis(wpi , S) is the closest Euclidean distance from way-
point wpi to a segment S, and W is the set of all waypoints
d TE 1
κ= = (6) in the airspace.
ds R
b: ENRICHED INTENT
where TE is the unit tangent vector of the trajectory, s is
To identify an additional trajectory change point (Figure 6b),
the arc length, and R is the radius (Figure 6a). We define a
first, we determine the time tc at which the curvature reaches
turn segment as a trajectory segment where all the points in
its maximum value during the turn. Second, we find the posi-
the segment have R ≤ 40 NM. This radius threshold was
tion of the aircraft at tc assuming the aircraft was following
determined empirically to reduce turns mis-detection.
the pre-turn direction. Finally, we randomly choose a point
within a noise circle of radius 1 NM, whose center is placed
at the position found in the second step (the green circle
in Figure 6b). The enriched intent for a trajectory can be
expressed by:
IE = IO ∪ {TCPi } (8)
where each TCPi is an identified trajectory change point.
An example of ordinary and enriched intents for a trajectory
is demonstrated in Figure 7.

B. DATA SOURCE AND PRE-PROCESSING

The data used in this work is Automatic Dependent
Surveillance–Broadcast (ADS–B) provided by the Civil Avi-
FIGURE 7. Example of reconstructed ordinary intent (left) and enriched
intent (right) for a flown track. ation Authority of Singapore. The data covers a geographical
region of latitude from N0◦ 00 to N25◦ 00 , longitude from
2) ORDINARY AND ENRICHED INTENTS RECONSTRUCTION E85◦ 00 to E125◦ 00 (Figure 8), and dated from March 15th
Intent reconstruction is performed following the steps below. 2019 to April 30th 2019. We consider all the aircraft flown
tracks in the mentioned region during the chosen time win-
a: ORDINARY INTENT dow, and this results in a total of 16,884 flight trajectories.
We choose designated significant waypoints to include in the The following pre-processing steps were applied.
flightplan of a flight. For non-turn segment, Ss , a waypoint
is included when it is less than 2 NM apart from the flight’s 1) DATA FILTERING AND TRAJECTORY RE-SAMPLING
flown track. For turn segment, St , this threshold is 7 NM. The The following conditions were applied for filtering data to
ordinary intent for a trajectory can be written as: remove noises:
• A trajectory needs to have at least one hour of data in
IO = {wpi ∈ W | dis(wpi , Ss ) ≤ 2 NM the mention region and has no missing positions during
or dis(wpi , St ) ≤ 7 NM} (7) a time window of 1000 seconds.

17888 VOLUME 10, 2022

P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

current position. We then combine the aircraft current posi-

tion with k most recent positions to form the input sequence of
length (k + 1). The sequence of next n points after the aircraft
current position is regarded as the true value of the output
sequence. Thus, each sample contains an input sequence and
a true output sequence.

2) COORDINATES TRANSFORMATION
Initial experiments showed that the training is more robust
when all positional coordinates in each training sample are
expressed in a local Cartesian coordinate system instead of
using the global geodetic coordinates (i.e., latitude, longitude,
altitude). Therefore, for each sample, we transform all geo-
graphic positions from the global geodetic coordinate to a
FIGURE 8. The geographical region (shaded area) of the ADS-B data local tangent plane, i.e., east-north-up (ENU), coordinate.
considered in this work.
This transformation also applies to the computation of the
intent vector. The origin of the local ENU frame is always
placed at the aircraft current position of each data sample.
• The computed ground speed from aircraft’s positions
The output of the decoder are also coordinates in the local
should be stable. Therefore, we discard trajectories that
ENU frame.
give rise to speed outliers. Concretely, the top 1 percent
and the bottom 1 percent in distribution of computed
TABLE 3. Hyper-parameters settings for model training.
speeds is omitted.
• Only parts of the trajectories that are between flight
levels 290 and 410 are maintained since we focus on the
en-route phase.
After this filtering process, 2882 trajectories remained for
training and evaluation. Since the proposed model requires
the input and output sequences to have equal time interval
between any two consecutive data points, we employed cubic
spline interpolation for trajectory re-sampling with a fixed
time interval of 15 seconds. We chose this re-sampling inter- IV. EXPERIMENTS & RESULTS
val because it helps to reduce the amount of training data A. MODEL TRAINING AND HYPER-PARAMETERS
being fed to the model and therefore speed up model training The proposed model architecture and learning algorithm are
without compromising the prediction accuracy. implemented in Python language using PyTorch deep learn-
ing library. In particular, we utilize a mini-batch stochastic
2) TRAIN & TEST DATASETS SPLITTING gradient descent method to train our network model with
We construct the train dataset using the trajectory data automatic differentiation. The Adam optimizer, a variant of
dated from March 15th to March 25th, and the remaining this method implemented in PyTorch, is selected to use in our
data is used for testing and evaluation. The train dataset experiment with learning rate set to 0.0001.
includes 690 flight tracks with 358,294 data points and the RNNs in general and GRUs specifically are slow to train
test dataset includes 2192 flight tracks with 1,130,913 data due to the fact that elements in the output are computed
points. We use data from the first 11 days, about 25% of sequentially, one after another. Using large mini-batch size
the total data, to train the model and use the remaining data and taking advantage of GPU computing can reduce the
to test the performance of the trained model. Experiments, training time significantly. To further save the training time,
which is discussed in detail in Section IV-A, showed that the we start with large mini-batch size (i.e., 256 samples per
proposed model converges with this amount of training data batch) at the beginning of the training, and gradually reduce
and there is little benefit of using more training data. We also the batch size to as small as 32 samples. In such manner, the
observed that allocating more data for testing could help to final set of model’s parameters are obtained after 100 training
better evaluate the prediction robustness and stability of the epochs.
model. In the final settings, our model predicts the next 40 future
positions of the aircraft given the most recent 10 positions,
C. DATASET PREPARATION including the current position. In the training phase, however,
1) TRAINING SAMPLES GENERATION the output length of the decoder is not necessarily fixed.
From the pre-processed trajectory data, we generate a sample In fact, we observed that the model converges faster and
by choosing at random a point in a trajectory as the aircraft more robust if we allow the output length to alternatively take

VOLUME 10, 2022 17889

 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

values from the range [5, 40]. More specifically, small values
of n (i.e., 5 or 10) help the model to converge faster in the
‘‘right’’ direction at the beginning. After that, larger values
of n allow the model to improve its performance on longer
prediction horizon. Thus, in training time, n takes controlled
random values within this range, as indicated in Table 3. Once
the model has been trained, we set n = 40 for evaluation.
Such flexibility in the output length is offered by the GRU
network of the decoder.
We also empirically determine the dynamics stability loss
coefficient (α) to be 0.25. Hyper-parameters regarding the
model architecture are shown in Table 2. Hyper-parameters
for model training are indicated in Table 3.

B. INDEX OF PERFORMANCE
Euclidean distance is a common choice for measuring simi- FIGURE 9. Along-track error (ATE) and cross-track error (CTE) of predicted
larities in trajectory prediction evaluation. We employ Mean aircraft position.

Absolute Error (MAE) and Root Mean Square Error (RMSE)

to evaluate the error in the predicted trajectory.
v The analysis in Table 4 also indicates that providing the
u n
u1 X model with enriched intent significantly reduces both errors,
d2 p̂i , pi


RMSE = t (9) comparing to the case where only ordinary intent is con-
n
i=1 sidered. In specific, when enriched intent is used, RMSE
n
1X improves approximately 14% to 30% depending on the look-
d p̂i , pi


MAE = (10) ahead time, and MAE improves approximately 11% to 20%.
n
i=1 It is very interesting to observe that the improvement in two
where n is the length of output sequence, p̂i and pi are the true types of error is more significant when the prediction horizon
and predicted positions, respectively. is farther into the future. This suggests that enriched intent
We also consider along-track error (ATE), cross-track of the aircraft plays a more important role in maintaining
error (CTE) for performance evaluation [5]: accurate prediction when one performs prediction with longer
look-ahead time. For short-term prediction, the recent state
ATE = 1p(1) sin ϕ + 1p(2) cos ϕ (11) of the aircraft influences the prediction performance more
CTE = 1p(1) cos ϕ − 1p(2) sin ϕ (12) than the intent does. In fact, finding the dynamic balance
between the influence of recent behavior and that of future
where 1p(1) = p̂(1) − p(1) , 1p(2) = p̂(2) − p(2) , ϕ is intent when making prediction at different look-ahead times
the angle between the aircraft track and the north direc- is an important exercise in model design.
tion. Note that p(1) and p(2) are coordinates in the local
ENU reference frame. The ATE and CTE are illustrated in TABLE 4. Model prediction performance in terms of RMSE and MAE (in
nautical miles) at different look-ahead times. The results are shown for
Figure 9. two different levels of intent reconstruction.

C. RESULTS ANALYSIS
Table 4 compares model’s performance on inputs with ordi-
nary intent and that with enriched intent, in terms of RMSE
and MAE of the Euclidean distances between true positions
and the predicted ones, at different look-ahead times. In gen-
eral, the prediction errors grow linearly with the prediction
horizon. When ordinary intent is provided, from 1 to 10 min-
utes look-ahead time, RMSE grows from 0.49 NM to 4.9 NM,
and MAE from 0.3 NM to 3.3 NM. When enriched intent is
available, RMSE ranges from 0.4 NM to 3.4 NM and MAE
from 0.3 to 2.6 NM, from 1 to 10 minutes of prediction
horizon. One can also observe that the RMSE/MAE ratio in We now examine further the model performance by inves-
the case with ordinary intent is higher than that when using tigating cross-track error (CTE) and along-track error (ATE).
enriched intent. This suggests that large errors arise more fre- As shown in Figure 9, CTE indicates how much the aircraft
quently when the provided intent information is insufficient slide off the actual course, while ATE measures the difference
or incorrect, which happens when the intent is not enriched. between the projected position along the actual course and its

17890 VOLUME 10, 2022

P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

true value. Figure 10 presents boxplots of these two errors Now, we offer a comparison between performance of our
at different prediction horizons. For CTE (Figure 10a), the model and that of the model recently developed in [46].
model achieves a good prediction accuracy, where 50% of We choose the work [46] for benchmark for it also attempted
the errors are below 1 NM and 75 % of them below 1.8 NM, to embed intent of the aircraft in the prediction model in the
at 10 minutes prediction horizon. For ATE (Figure 10b), form of physical constraints. The model in [46] also made use
the median value of the error goes from about 0.15 NM of LSTM network, which can be considered as a variant of
at 1 minute look-ahead time to 1.5 NM at 10 minutes. The recurrent neural network like GRU. One difference between
ATE is not as low as the CTE due to the fact that uncertainty the two is that the model in [46] makes use of the constraint
grows faster during the longitudinal progress. One attribute of as a part of the loss function, while we input the aircraft intent
the ATE is that it can be used to estimate the error in arrival to the decoder. In this study, the proposed model takes as
time of the aircraft at certain point in the future, given infor- input the last 10 positions and can predict future trajectory
mation about the aircraft speed. A common property of CTE as a positions sequence of flexible length, where the time
and ATE observed from Figure 10 is that they both linearly interval between two consecutive points is 15 seconds. The
increase with the look-ahead time. The accumulation of errors model in [46] takes as input the last 10 positions and predicts
over many sequential prediction steps can not be completely a position at 15 seconds into the future. Thus, for comparison,
avoided, unfortunately. Nevertheless, the model is successful we allow our model to predict one step of 15 seconds and
in avoiding large explosion of the errors over time, and the compare the prediction with that of [46]. The prediction error
consistent linear relationships between errors and prediction resulted from our model is accumulated after every prediction
horizon makes the errors’ behavior more predictable. step; thus, a small error in a single-step prediction is desirable.
Figure 11 shows the distributions of CTE and ATE at three
different prediction horizons of 2, 5, and 10 minutes, per- TABLE 5. Performance benchmark with state-of-the-art model.
formed on the test dataset. These distributions also indicate
that a majority of test cases produce low errors. For example,
it can be observed that at 2 minutes look-ahead time, about
75% of the cases have both CTE and ATE below 0.5 NM.
When we look farther into the future, e.g., at 5 minutes, about
75% of the test cases have CTE below 1 NM and have ATE The comparison is presented in Table 5. The benchmark
below 1.3 NM. Longer look-ahead time amplifies the gap shows that our model performs significantly better and more
between CTE and ATE because of the effect of increasing stable than cLSTM does, in terms of MAE and RMSE of
uncertainty during the longitudinal progress, as we mentioned latitude and longitude predictions. The fact that the ratios
earlier. RMSE/MAE of our model (1.38 for latitude and 1.31 for
The model predicts the aircraft future trajectory without longitude) are significant lower than that of cLSTM (4.06
information about aircraft performance. In air traffic man- for latitude and 4.59 for longitude) suggests that our model
agement, aircraft performance (or aircraft type) is usually produces very low variance compared to cLSTM, and that
associated with their wake turbulence categories (WTC), large errors happen less frequently in our case than they do in
i.e., light, medium, and heavy aircraft. To verify the model cLSTM.
performance for different aircraft categories, a further inves-
tigation on the prediction errors is performed on different TABLE 6. Model prediction increment in mean and standard deviation of
errors by minute.
WTCs. Figure 12a shows the statistics of different air-
craft types presented in the test dataset, and how they are
grouped into medium and heavy WTCs (en-route airspace
have medium and heavy aircraft only). We allow the model
to perform prediction for the medium and the heavy groups
separately, and the results are shown in Figure 12b in terms
of mean Euclidean distance errors. It is observable from
Figure 12b that the difference in prediction errors between
the two groups of WTCs is insignificant. The insignificant
difference between these errors suggests that the model does
not have any bias toward any specific WTC, and the model’s
performance is consistent regardless of aircraft WTC. This
also indicates that our design of the encoder is able to char- To further assess the model performance, we compute
acterize the dynamics of different aircraft types; thus, the the error growth rates at different prediction horizons. The
explicit information of the aircraft type can be safely removed error growth rate measures how much the prediction error
from the input. This is an important feature offered by our is worsened after every minute of look-ahead time. A good
model as it helps to limit the amount of information needed predictor should be able to keep these error growth rates
while maintaining stable and high performance. within bounds. Table 6 presents the growth rates of ATE and

VOLUME 10, 2022 17891

 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

FIGURE 10. Boxplots of (a) Cross-Track error and (b) Along-Track error at different prediction horizons.

FIGURE 11. Distribution of Cross-Track Error (CTE) and Along-Track Error (ATE) at different prediction horizons. The first row shows CTE and the second
row shows ATE. The three columns indicate three prediction horizons at 2, 5, 10 minutes, from left to right.

CTE in terms of mean and standard deviation (std) of the 0.1 NM/min and standard deviation of 0.3 NM/min. In terms
errors, in the unit of NM per minute. It can be observed that of mean growth rate, the model performance is quite close
the error growth rates produced by our model are very stable, to the specifications. The model demonstrates that it can
in terms of mean and std of the error, for both along-track maintain a very low variances in the error growth rates despite
and cross-track. This indicates that our model can effectively increasing look-ahead time. In specific, for 10 minutes look-
manage the accumulated prediction errors when prediction ahead, the proposed model offers ATE growth rate std rang-
horizon becomes farther into the future. ing from 0.132 NM/min to 0.167 NM/min, which is lower
The EUROCONTROL Specification for Trajectory Predic- than 0.2 NM/min. For CTE, in 10 minutes prediction hori-
tion [2] specifies the required growth rates for ATE of mean zon, the proposed model produces growth rates std ranging
0.1 NM/min and std 0.2 NM/min, and for CTE of mean from 0.121 NM/min to 0.171 NM/min, which is significantly

17892 VOLUME 10, 2022

P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

FIGURE 12. Prediction errors for different aircraft wake categories. (a) Number of different aircraft types in the test set. (b) Mean of prediction errors for
medium and heavy aircraft.

FIGURE 13. Four scenarios of aircraft trajectories being simulated in the BlueSky ATC Simulator and their corresponding predicted trajectories by the
model. In each scenario, the left figure demonstrates the trajectory segment on the BlueSky’s screen, and the right figure shows comparison between
simulated (red) and predicted (green) trajectories, given the input to the model (i.e., recent locations of the aircraft) in blue.

low comparing to 0.3 NM/min. Although there is room for the constraints (i.e., air speed, altitude, heading etc.) at each
improvement in the mean of error growth rates, the results waypoint.
are promising if one takes into consideration that the model To facilitate the validation, we prepare a validation tra-
requires only recent trajectory and simple intent information jectories dataset in which each trajectory is simulated by
as input. the BlueSky’s Flight Management System (FMS) given
the respective pre-defined flight-plan. During flight simula-
V. SIMULATION VALIDATION tion, locations of aircraft are recorded at every 15 seconds.
To further evaluate the performance of the prediction model, The trained model takes the 10 most recent locations of
we perform model validation in an air traffic simulator. an aircraft as inputs and predicts the aircraft’s location in
The BlueSky ATC Simulator [49] is used for this pur- the next 10 minutes. Model performance is evaluated by
pose. BlueSky is an open-source and research-oriented air the errors when comparing the predicted trajectories with the
traffic simulator that can integrate highly realistic aircraft BlueSky simulated ones, in terms of along-track and cross-
dynamics via the EUROCONTROL’s Base of Aircraft Data track errors. The validation dataset consists of 1,138,343 data
(BADA) [50]. To simulate an aircraft trajectory, the BlueSky points in total.
simulator takes the flight-plan and aircraft type as inputs and Figure 13 depicts four scenarios of aircraft trajectories
generates a 4D trajectory. A flight-plan generally includes being simulated in the BlueSky simulator and the compar-
all the waypoints that the aircraft is going to traverse and ison between the simulated trajectories and the predicted

VOLUME 10, 2022 17893

 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

the cross-track and the along-track error growth rates during

simulation validation, as shown in Table 7, also approach the
test results performed on ADS-B data earlier (Table 6). This
consistency in model’s performance suggests that the model,
which is trained on ADS-B dataset, can generalize well in the
simulation environment without the need of further training.

VI. CONCLUSION
In this work, a deep learning model, based on encoder-
decoder architecture, is proposed for aircraft trajectory
prediction by modeling and incorporating aircraft intent.
Specifically, an encoder network that effectively learns pat-
terns in the behavior of aircraft without the need of explicit
information about the aircraft performance and a decoder
network that fuses information about aircraft future intent into
the knowledge about its recent behavior, and effectively pre-
dicts future trajectory of the aircraft is designed. We demon-
strate that an effective modeling and incorporation of intent,
could improve the prediction performance up to 30% in terms
of Root Mean Squared Error of Euclidean distance between
predicted and true positions, at 10 minutes prediction hori-
zon (look ahead time for Conflict Detection models in Air
Traffic Control systems). Our approach also produce very low
variance in the prediction, compared to the standards required
by EUROCONTROL. The accumulated prediction error over
time is also well managed by the proposed model. The model
is also able to discriminate different aircraft dynamics during
prediction without the need of explicit information about
aircraft wake categories. With these features, the proposed
FIGURE 14. Boxplots of (a) cross-track and (b) along-track errors at model outperforms the existing state-of-the-art model in air-
different prediction horizons between the predicted trajectories by the craft trajectory prediction. Another benefit from the proposed
model and the simulated trajectories by the BlueSky Simulator.
approach is that the model requires minimal amount of infor-
mation to perform the prediction.
TABLE 7. Validation results in BlueSky simulator: along-track and
cross-track errors growth rate at different prediction horizons. A key factor that contributes to the safe operation of an
airspace sector is the situational awareness of air traffic con-
trollers, and this in turn depends on how well the controllers
anticipate the traffic movement, particularly all aircraft future
locations. Thus, an accurate and reliable aircraft trajectory
prediction would be valuable for the controllers to probe
any potential incidents in the sector and resolve them timely
and efficiently. Future extensions of this work may include
(1) extended design of model architecture for incorpora-
tion of multi-aircraft intent, (2) modeling of uncertainty in
the implementation of ATCO instructions by flight crews,
(3) analysis of effects of traffic flow management strategies
used by ATCOs on aircraft intent, (4) further investigation of
ones. These four scenarios are chosen to demonstrate the model performance using actual ATC clearance data instead
model’s performance at different kinds of aircraft intent of enriched intent, etc.
(i.e., waypoints in the flight-plan) and at different segments
(i.e., straight and turn segments) of the trajectory. It is observ- ACKNOWLEDGMENT
able that the model predicts the aircraft future locations Any opinions, findings, and conclusions or recommenda-
equally well during the straight and turn segments of the tions expressed in this material are those of the authors and
trajectory. ATE and CTE of the validation are presented in do not reflect the views of National Research Foundation,
Figure 14, in which the errors resulted from the validation Singapore, and the Civil Aviation Authority of Singapore.
in BlueSky Simulator are comparable with the errors per- (Phu N. Tran and Hoang Q. V. Nguyen contributed equally
formed on the ADS-B test dataset (Figure 10). Furthermore, to this work.)

17894 VOLUME 10, 2022

P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

REFERENCES [26] Y. Pang and Y. Liu, ‘‘Probabilistic aircraft trajectory prediction considering
[1] (2004). FAA/EUROCONTROL. Action Plan 16: Common Trajectory weather uncertainties using dropout as Bayesian approximate variational
Prediction Capability. Accessed: Sep. 18, 2020. [Online]. Available: inference,’’ in Proc. AIAA Scitech Forum, Jan. 2020, p. 1413.
https://sites.google.com/site/trajectorymgt/ap16—white-paper [27] Y. Pang and Y. Liu, ‘‘Conditional generative adversarial networks (CGAN)
[2] (2017). EUROCONTROL. EUROCONTROL Specification of for aircraft trajectory prediction considering weather effects,’’ in Proc.
Trajectory Prediction. Accessed: Sep. 18, 2020. [Online]. Available: AIAA Scitech Forum, Jan. 2020, p. 1853.
https://www.eurocontrol.int/publication/eurocontrol-specification- [28] Z. Chen, D. Guo, and Y. Lin, ‘‘A deep Gaussian process-based flight
trajectory-prediction trajectory prediction approach and its application on conflict detection,’’
[3] ICAO. ICAO Global TBO Concept. Accessed: Dec. 15, 2021. [Online]. Algorithms, vol. 13, no. 11, p. 293, Nov. 2020.
Available: https://www.icao.int/airnavigation/tbo/PublishingImages [29] H.-C. Choi, C. Deng, and I. Hwang, ‘‘Hybrid machine learning and
/Pages/Why-Global-TBO-Concept/ estimation-based flight trajectory prediction in terminal airspace,’’ IEEE
[4] S. Swierstra and S. Green, ‘‘Common trajectory prediction capability for Access, vol. 9, pp. 151186–151197, 2021.
decision support tools,’’ in Proc. 5th USA/Eurocontrol ATM R&D Seminar, [30] S. Hochreiter and J. Schmidhuber, ‘‘Long short-term memory,’’ Neural
Budapest, Hungary, 2003, pp. 23–27. Comput., vol. 9, no. 8, pp. 1735–1780, 1997.
[5] S. Mondoloni, S. Swierstra, and M. Paglione, ‘‘Assessing trajectory pre- [31] K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares,
diction performance-metrics definition,’’ in Proc. 24th Digit. Avionics Syst. H. Schwenk, and Y. Bengio, ‘‘Learning phrase representations using
Conf., vol. 1, Oct. 2005, p. 3. RNN encoder-decoder for statistical machine translation,’’ 2014,
[6] M. Paglione and R. Oaks, ‘‘Implementation and metrics for a trajectory arXiv:1406.1078.
prediction validation methodology,’’ in Proc. AIAA Guid., Navigat. Control [32] I. Sutskever, O. Vinyals, and Q. V. Le, ‘‘Sequence to sequence learning
Conf. Exhib., Aug. 2007, p. 6517. with neural networks,’’ in Proc. Adv. Neural Inf. Process. Syst., 2014,
[7] S. Torres, ‘‘Determination and ranking of trajectory accuracy factors,’’ in pp. 3104–3112.
Proc. 29th Digit. Avionics Syst. Conf., Oct. 2010, p. 1. [33] C. Wang, L. Ma, R. Li, T. S. Durrani, and H. Zhang, ‘‘Exploring trajec-
[8] B. Sridhar, ‘‘Applications of machine learning techniques to aviation oper- tory prediction through machine learning methods,’’ IEEE Access, vol. 7,
ations: Promises and challenges,’’ in Proc. Int. Conf. Artif. Intell. Data pp. 101441–101452, 2019.
Analytics Air Transp. (AIDA-AT), Feb. 2020, pp. 1–12. [34] S. H. Park, B. Kim, C. M. Kang, C. C. Chung, and J. W. Choi,
[9] R. Wu, G. Luo, J. Shao, L. Tian, and C. Peng, ‘‘Location prediction on tra- ‘‘Sequence-to-sequence prediction of vehicle trajectory via lstm encoder-
jectory data: A review,’’ Big Data Mining Anal., vol. 1, no. 2, pp. 108–127, decoder architecture,’’ in Proc. IEEE Intell. Vehicles Symp. (IV), Jun. 2018,
2018. pp. 1672–1678.
[10] H. Georgiou, S. Karagiorgou, Y. Kontoulis, N. Pelekis, P. Petrou, [35] Y. Ma, X. Zhu, S. Zhang, R. Yang, W. Wang, and D. Manocha, ‘‘TrafficPre-
D. Scarlatti, and Y. Theodoridis, ‘‘Moving objects analytics: Survey on dict: Trajectory prediction for heterogeneous traffic-agents,’’ in Proc. AAAI
future location & trajectory prediction methods,’’ 2018, arXiv:1807.04639. Conf. Artif. Intell., vol. 33, 2019, pp. 6120–6127.
[11] J. K. Kuchar and L. C. Yang, ‘‘A review of conflict detection and resolu- [36] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones,
tion modeling methods,’’ IEEE Trans. Intell. Transp. Syst., vol. 1, no. 4, A. N. Gomez, L. Kaiser, and I. Polosukhin, ‘‘Attention is all you
pp. 179–189, Dec. 2000. need,’’ 2017, arXiv:1706.03762.
[12] J. Tang, ‘‘Conflict detection and resolution for civil aviation: A literature [37] Y. Liang, S. Ke, J. Zhang, X. Yi, and Y. Zheng, ‘‘Geoman: Multi-level
survey,’’ IEEE Aerosp. Electron. Syst. Mag., vol. 34, no. 10, pp. 20–35, attention networks for geo-sensory time series prediction,’’ in Proc. IJCAI,
Oct. 2019. 2018, pp. 3428–3434.
[13] M. Prandini, J. Hu, J. Lygeros, and S. Sastry, ‘‘A probabilistic approach to [38] D. Kong and F. Wu, ‘‘HST-LSTM: A hierarchical spatial-temporal long-
aircraft conflict detection,’’ IEEE Trans. Intell. Transp. Syst., vol. 1, no. 4, short term memory network for location prediction,’’ in Proc. IJCAI,
pp. 199–220, Dec. 2000. vol. 18, no. 7, 2018, pp. 2341–2347.
[14] W. Liu and I. Hwang, ‘‘Probabilistic trajectory prediction and conflict [39] A. Karatzoglou, A. Jablonski, and M. Beigl, ‘‘A Seq2Seq learning
detection for air traffic control,’’ J. Guid., Control, Dyn., vol. 34, no. 6, approach for modeling semantic trajectories and predicting the next loca-
pp. 1779–1789, 2011. tion,’’ in Proc. 26th ACM SIGSPATIAL Int. Conf. Adv. Geographic Inf.
[15] Z. Yu, ‘‘Trajectory data mining: An overview,’’ ACM Trans. Intell. Syst. Syst., Nov. 2018, pp. 528–531.
Technol., vol. 6, no. 3, pp. 1–41, 2015. [40] H. Wu, Z. Chen, W. Sun, B. Zheng, and W. Wang, ‘‘Modeling trajectories
[16] F. Giannotti, M. Nanni, F. Pinelli, and D. Pedreschi, ‘‘Trajectory pattern with recurrent neural networks,’’ in Proc. 26th Int. Joint Conf. Artif. Intell.,
mining,’’ in Proc. 13th ACM SIGKDD Int. Conf. Knowl. Discovery Data Aug. 2017, pp. 1–9.
Mining (KDD), 2007, pp. 330–339. [41] Z. Shi, M. Xu, Q. Pan, B. Yan, and H. Zhang, ‘‘LSTM-based flight
[17] M. G. Hamed, D. Gianazza, M. Serrurier, and N. Durand, ‘‘Statistical pre- trajectory prediction,’’ in Proc. Int. Joint Conf. Neural Netw. (IJCNN),
diction of aircraft trajectory: Regression methods vs point-mass model,’’ Jul. 2018, pp. 1–8.
ENAC, Toulouse, France, Tech. Rep. hal-00911709, 2013. [42] P. Han, W. Wang, Q. Shi, and J. Yang, ‘‘Real-time short-term trajectory
[18] A. de Leege, M. van Paassen, and M. Mulder, ‘‘A machine learning prediction based on gru neural network,’’ in Proc. IEEE/AIAA 38th Digit.
approach to trajectory prediction,’’ in Proc. AIAA Guid., Navigat., Control Avionics Syst. Conf. (DASC), Sep. 2019, pp. 1–8.
(GNC) Conf., Aug. 2013, p. 4782. [43] Y. Pang, N. Xu, and Y. Liu, ‘‘Aircraft trajectory prediction using LSTM
[19] S. Hong and K. Lee, ‘‘Trajectory prediction for vectored area navigation neural network with embedded convolutional layer,’’ in Annu. Conf. PHM
arrivals,’’ J. Aerosp. Inf. Syst., vol. 12, no. 7, pp. 490–502, 2015. Soc., vol. 11, no. 1, pp. 1–10, 2019.
[20] T.-H. Le, P. N. Tran, D.-T. Pham, M. Schultz, and S. Alam, ‘‘Short-term [44] L. Ma and S. Tian, ‘‘A hybrid CNN-LSTM model for aircraft 4D trajectory
trajectory prediction using generative machine learning methods,’’ in Proc. prediction,’’ IEEE Access, vol. 8, pp. 134668–134680, 2020.
Conf. (ICRAT), Tampa, FL, USA, Sep. 2020, pp. 1–8. [45] P. Han, J. Yue, C. Fang, Q. Shi, and J. Yang, ‘‘Short-term 4D trajectory
[21] Z. Wang, M. Liang, and D. Delahaye, ‘‘Short-term 4D trajectory pre- prediction based on LSTM neural network,’’ Proc. SPIE, vol. 11427,
diction using machine learning methods,’’ ENAC, Toulouse, France, Jan. 2020, Art. no. 114270M.
Tech. Rep. hal-01652041, Tech. Rep., 2017. [46] Z. Shi, M. Xu, and Q. Pan, ‘‘4-D flight trajectory prediction with con-
[22] G. Wang, H. Chen, K. Liu, R. Guo, and Y. Wei, ‘‘A flight trajectory strained LSTM network,’’ IEEE Trans. Intell. Transp. Syst., vol. 22, no. 11,
prediction method based on trajectory clustering,’’ in Proc. IEEE 1st Int. pp. 7242–7255, Nov. 2021.
Conf. Civil Aviation Saf. Inf. Technol. (ICCASIT), Oct. 2019, pp. 654–660. [47] H. Kim and K. Lee, ‘‘Air traffic prediction as a video prediction problem
[23] H. Georgiou, N. Pelekis, S. Sideridis, D. Scarlatti, and Y. Theodoridis, using convolutional LSTM and autoencoder,’’ Aerospace, vol. 8, no. 10,
‘‘Semantic-aware aircraft trajectory prediction using flight plans,’’ Int. p. 301, Oct. 2021.
J. Data Sci. Anal., vol. 9, no. 2, pp. 215–228, Mar. 2020. [48] I. Goodfellow, Y. Bengio, A. Courville, and Y. Bengio, Deep Learning,
[24] X. Zhang and S. Mahadevan, ‘‘Bayesian neural networks for flight tra- vol. 1, no. 2. Cambridge, MA, USA: MIT Press, 2016.
jectory prediction and safety assessment,’’ Decis. Support Syst., vol. 131, [49] J. M. Hoekstra and J. Ellerbroek, ‘‘Bluesky atc simulator project: An open
Apr. 2020, Art. no. 113246. data and open source approach,’’ in Proc. 7th Int. Conf. Res. Air Transp.,
[25] Y. Pang, X. Zhao, H. Yan, and Y. Liu, ‘‘Data-driven trajectory prediction vol. 131, 2016, p. 132.
with weather uncertainties: A Bayesian deep learning approach,’’ Transp. [50] A. Nuic, ‘‘User manual for the base of aircraft data (BADA) revision 3.10,’’
Res. C, Emerg. Technol., vol. 130, Sep. 2021, Art. no. 103326. Atmosphere, vol 2010, p. 1, Feb. 2010.

VOLUME 10, 2022 17895

 P. N. Tran et al.: Aircraft Trajectory Prediction With Enriched Intent Using Encoder-Decoder Architecture

PHU N. TRAN received the Ph.D. degree in DUC-THINH PHAM received the M.Sc. degree
mechanical engineering from Nanyang Techno- in computer science from Télécom ParisTech,
logical University, Singapore, in 2017. He is cur- France, in 2013, and the Ph.D. degree in computer
rently a Postdoctoral Research Fellow at the Air science from Paris Sciences and Letters (PSL),
Traffic Management Research Institute, Nanyang France, in 2019. He is currently a Research Follow
Technological University. His research interests at the Air Traffic Management Research Institute,
include artificial intelligence, machine learning, Nanyang Technological University, where he is
and the applications of those techniques in air also the Assistant Program Director of Artificial
traffic control. Intelligence and Data Analytics for Air Traffic
Management Program and the Leader of data-
driven surface movement management project. His research interests include
on deep learning, deep reinforcement learning, and application of machine
learning in air traffic management challenges, such as surface movement
management and conflict detection and resolution.

SAMEER ALAM received the Ph.D. degree in

computer science with specialization in artificial
intelligence from the University of New South
Wales (UNSW), Australia, in 2008. He is currently
HOANG Q. V. NGUYEN received the M.Eng. an Associate Professor at the School of Mechan-
degree in computer science from Télécom Paris- ical and Aerospace Engineering, Nanyang Tech-
Tech, France, in 2018. He is currently a Research nological University (NTU), Singapore, where
Associate at the Air Traffic Management Research he is also the Deputy Director of the Air Traf-
Institute, Nanyang Technological University. His fic Management Research Institute (ATMRI). His
research interests include machine learning and research interests include machine learning, com-
data drive approaches and theirs application on puter vision, multi-agent systems, applied to air traffic, and airport opera-
trajectory prediction, conflict detection, and reso- tions. He is an Editorial Board Member of Transportation Research Part C:
lution for air traffic management. Emerging Technologies.

17896 VOLUME 10, 2022