1. Introduction
With the spectacular increase of flights in air transportation, a large number of congestions and flight delays occur due to the limited airport capacity. Compared with the high cost of extending the airport infrastructure, enhancing the efficiency of the air traffic flow management (ATFM) at airports is a preferred option for concerned departments in short-term. As an essential technique in air traffic control (ATC), airport arrival flow prediction (AAFP) can detect the arrival demand timely, which is the foundation to provide decision-making on the ATFM. On the one hand, it helps air traffic controllers (ATCOs) to foresee real-time airspace situations at airports, relieving the workload of the controllers. On the other hand, it can provide more time to make an efficient flow plan for alleviating airport traffic congestions in time.
The short-term arrival flow prediction in an ATC system is typically based on the historical and current airport traffic flow information, including departure flow, arrival flow, and so forth, to estimate the arrival flow in the near future. With regard to the air traffic prediction task [
1,
2], it is commonly agreed that it is subject to the complex spatial–temporal dependencies of traffic data. Therefore, when forecasting the airport arrival flow, it is necessary to consider the spatial and temporal dependencies integrally:
- (1)
Spatial dependencies: the evolution of the airport arrival flow relates to the topological structure of a given airport network. In air transportation, the flights commute between airports conforming to a flight schedule overall, which contains the departure time, flight routes, and arrival time at airports. Therefore, a semantic airport network can be constructed with this scheduling, which takes airports as nodes, the city-pairs as edges and the number of scheduled flights between airports as the weight values. Specifically, as illustrated in
Figure 1, the red link represents that there are scheduled flights that can commute between airports. The wider the link is, the more scheduled flights it has between both ends. Focusing on the city-pair between Beijing and Shanghai (ZBAA-ZSSS), both airports have a large number of actual arrival flights, which demonstrates that more scheduled flights may bring the larger arrival flow, and validate the effect of the semantic airport network.
- (2)
Temporal dependencies: the airport arrival flow changes in periodicity, trend and closeness. As the flight operation is usually arranged by airlines weekly, the airport arrival flow presents periodical patterns each week. Within one day, the peak hours of the airport arrival flow usually surge during 12:00–14:00 and 17:00–19:00, and the bottom mostly appears between 00:00 and 07:00. The closeness means that the arrival flow on the adjacent time slice often changes smoothly.
For years, great efforts have been made to improve the performance of traffic flow prediction. In terms of prediction methods, the previously published works can generally be classified into three categories, that is, flight plan-based algorithms, traffic flow model-driven algorithms and data-driven algorithms. The flight plan-based algorithms achieve air traffic flow of the airspaces concerned by evaluating the 4-D aircraft trajectories [
3] based on flight plans [
4,
5]. This method relies heavily on the trajectory prediction (TP) technique, which is susceptible to uncertainties of the real-time traffic state, such as the weather condition and the airport facility state. As a result, the flight plan-based approaches failed to offer sufficient insights into the dynamics of the traffic flow [
6]. The flow model-driven algorithms generally learn the evolution pattern of the traffic flow by some handcrafted traffic models. Prior knowledge is required to design the traffic model to forecast future traffic flow, including the cell transmission model [
7], the queuing theory model [
8], and the aggregate flow model [
9,
10,
11], and so forth. However, the traffic state is influenced by many factors so that it may be difficult to fully illustrate unsteady air traffic flow by a specific model. The core idea for data-driven algorithms is to extract informative knowledge by mining the input dataset. The history average model (HA) [
12] is an early representative method. Recently, there is an increasing interest in applying methods based on Machine Learning Techniques (MLT) to problems in Air Traffic Management (ATM) [
13]. Some neural network models were proposed to achieve the traffic prediction, including the artificial neural network (ANN) [
14], the long short-term memory (LSTM) [
15,
16,
17], the gated recurrent unit (GRU) [
18] and the convolutional neural network (CNN) [
19,
20]. In general, the LSTM and the GRU models are able to provide a higher performance for time-series prediction tasks. However, as it fails to consider the spatial dependencies, the structure of the airport network is not captured in the related AAFP task. To capture the spatial dependencies of air traffic flow, the CNN model was used to mine the grid-based traffic flow by dividing the space into grided regions [
2]. However, due to the graph nature of the airport network, it may not be an optimal solution to directly extract the underlying patterns of the topological structure of the airport network by the CNN.
As deep learning models have been particularly successful in dealing with speech [
21,
22,
23], images [
24,
25,
26,
27], or videos [
28], the increasing architectures based on the graph neural network were proposed to extract informative graph representations for subsequent tasks. The graph convolutional network (GCN) [
29] is a generalized CNN, which can mine the high-level information in a non-Euclidean space directly. In addition, some integrated neural networks were also developed to mine the spatial-temporal data efficiently. The spectral graph Markov network (SGMN) [
30] was proposed to approximately characterize the dynamic change of traffic data. However, the Markov assumption of the SGMN may limit its performance on the multiple-step prediction tasks.
To solve the aforementioned problems, we first represent the airport network as a weighted graph based on a flight schedule, generally describing the spatial interactions of flights between airports. In succession, an airport traffic flow prediction network (ATFPNet), constructed by the graph convolution operator and the gated recurrent unit, is proposed to capture the evolution patterns of the airport traffic flow. Specifically, the overall traffic flow in the airport network can be encoded into a single structure by the specific graph representation. Thanks to the recurrent mechanism of the ATFPNet cell, the proposed approach has the ability to predict the multiple-step airport arrival flow in a situational manner. In addition, a real-world airport traffic flow dataset is constructed to validate the proposed approach, and the results demonstrate that the proposal outperforms other baselines, achieving up to 17% MAE improvement. All in all, the main contributions of this paper are summarized as follows:
Considering the air transportation context, a semantic airport network is built up by the flight schedule, which generally models the flights interactions between airports;
In light of the semantic airport network, a deep learning-based ATFPNet framework is proposed to predict the airport arrival flow in a multiple-step and situational manner, which is able to consider the spatial-temporal dependencies of airport traffic flow integrally;
The graph convolutional network and gated recurrent unit are combined to construct the ATFPNet model, which is the key to extracting the high-level transition patterns of airport traffic flow. Specifically, the spatial dependencies of inter-airports and the time-varying airport traffic flow sequence can be modeled by the two blocks, respectively;
A real-world dataset from the Civil Aviation Administration of China (CAAC) is applied to evaluate the performance of the proposed approach. Compared to the other baselines, the experimental results demonstrate that the proposed approach yields performance superiority for the short-term situational airport arrival flow forecasting.
The rest of this paper is organized as follows. Implementation details of the proposed model are introduced in
Section 2. In
Section 3, we list experimental configurations and evaluate the experimental results by a real-world airport traffic flow dataset. The discussion of the experiment is reported in
Section 4. We conclude the paper and introduce the future work in
Section 5.
5. Conclusions
In this paper, we construct a deep-learning-based model, called the airport traffic flow prediction network (ATFPNet), to achieve the airport arrival flow prediction task. The spatial graph convolution operator and gated recurrent unit are combined to capture the transition patterns of airport traffic flow (departure and arrival). With respect to the air transportation context, a specific graph representation is built based on the flight schedule to illustrate the airport network. By further applying the GRU in the ATFPNet cell, the situational (network-level) multiple-step arrival flow can be achieved on the airport network. A real-world airport traffic dataset is applied to validate the proposed approach, and the experimental results show the performance superiority over other comparative baselines, concerning several data-driven models. Compared to the GRU model, the RMSE of ATFPNet is relatively improved from 0.7% to 2.1%. As for the GCN, the proposed approach obtains a better performance, reducing the RMSE from 3.2% to 5%. In summary, the airport network representation built on the flight schedule makes a great contribution to the situational airport arrival flow prediction task, and the proposed ATFPNet has the ability to capture the spatial and temporal features of airport traffic data.
In the future, except for the airport departure and arrival factors used in this paper, we plan to explicitly consider other factors to improve prediction accuracy, for example, weather information (visibility or thunderstorm), air traffic control information, the influence of international flights, and dynamic traffic movements on the network.