An Airport Traffic Flow Prediction Method Based on Graph Neural Network

Abstract

The present invention relates to the field of traffic flow prediction, and particularly to an airport traffic flow prediction method based on a graph neural network. First, obtain airport traffic flow observation data, airport aviation meteorological data, airway network data, flight schedule data, and flight duration data between airport pairs; construct a multi-relational airport network through the airway network data, flight schedule data, and flight duration data between airport pairs, and construct an airport traffic flow dataset according to the airport traffic flow observation data and aviation meteorological data; apply an attention mechanism to fuse the multi-relational airport network into a directed single-relational airport network; use a bidirectional GCN model to extract the spatial structure features of multiple airport traffic networks; embed the model into a GRU model to extract the spatio-temporal correlation of airport traffic flow data, and construct an airport traffic flow prediction model. The present invention realizes airport traffic flow prediction in a large range and multi-airport scenario, and improves the prediction accuracy of airport traffic flow.

Description

Technical Field

[0001] The present invention relates to traffic flow prediction, in particular to an airport traffic flow prediction method based on a graph neural network. Background Art

[0002] With the rapid development of the air transportation industry, the number of flights has increased significantly. Due to the limited airport capacity, airspace congestion and flight delays often occur at busy airports, bringing heavy workloads to air traffic controllers and airport management personnel. Compared with expanding airport infrastructure, establishing more refined and effective traffic management measures can improve control efficiency in the short term and significantly reduce the workload of air traffic controllers. Air traffic flow management includes strategic air traffic flow management, pre-tactical air traffic flow management, and tactical air traffic flow management (real-time air traffic flow management). In principle, strategic air traffic flow management and pre-tactical air traffic flow management are the mainstays, with real-time traffic management as a supplement. Short-term traffic flow prediction is one of the leading tasks in pre-tactical and tactical traffic management. Accurate prediction results can enable air traffic controllers to understand the airport airspace traffic situation and trends in the future period in advance, and then reasonably arrange and coordinate airspace resources, prevent and alleviate various flight conflicts, and ensure flight safety. However, due to the complexity of the flight process, there are many factors affecting the change of airport traffic flow, including airport meteorological conditions, flight schedules, route network structures, and real-time changing traffic states and other air traffic element information.

[0003] Commonly used airport traffic flow prediction methods include the airport traffic flow prediction method based on track deduction and the airport traffic flow prediction method based on data-driven. Among them, the airport traffic flow prediction method based on track deduction requires accurate mastery of various parameter indicators of aircraft and meteorological conditions, and needs to calculate the trajectory information of each flight. When predicting the traffic flow in a large-scale airspace, the computational amount of the prediction model will increase significantly. The airport traffic flow prediction method based on data-driven avoids the uncertainties and non-linear factors affecting flight navigation, and can achieve the airport traffic flow prediction task by analyzing airport traffic flow observation data, meteorological information, and daily flight plan information.

[0004] Studies have shown that when the prediction time range exceeds 20 minutes, the accuracy of trajectory deduction will decrease significantly, which in turn affects the prediction accuracy of traffic flow prediction methods. Commonly used data-driven methods include linear regression, integrated moving average autoregressive model, K-nearest neighbor method, XGBoost, and other methods. These methods are computationally simple, but do not fully consider the spatiotemporal dynamics and randomness of airport traffic flow. Such methods are mostly used for single-airport traffic flow prediction. However, when predicting large-scale, multi-step airport traffic flow, it is necessary to establish a prediction model for each airport in the airspace of interest, which significantly increases the amount of calculation. At present, there is no effective method that simultaneously considers multiple air traffic factors and the spatiotemporal correlation characteristics of traffic flow to quickly implement a large-scale airport traffic flow prediction method. Summary of the Invention

[0005] The purpose of the present invention is to provide an airport traffic flow prediction method based on graph neural network for large-scale, multi-time-step airport traffic flow prediction tasks, which require separate modeling and complex calculations, and do not fully consider the various influencing factors and spatiotemporal correlation characteristics of air traffic. It uses the characteristics of the attention mechanism to adaptively obtain effective information relationships, integrate multiple air traffic network elements to form a unified representation of multiple elements, and then apply the GCN model to model large-scale, multi-airport network spatial relationships. Based on the advantage of the GRU model with good time series capture ability, it explores the evolution pattern of large-scale airport traffic flow.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for predicting airport traffic flow based on graph convolutional neural networks, including:

[0008] Step S1, obtain air traffic element data, including airport route network data, flight schedule data, flight duration data between airport pairs, aviation meteorological data, and airport traffic flow observation data; construct a multi-relationship airport network based on airport route network data, flight schedule data, and flight duration data between airport pairs; construct an airport traffic flow dataset based on airport traffic flow observation data and aviation meteorological data.

[0009] In step S2, based on the multi-relational airport network, an attention mechanism is used to construct a multi-relational airport network fusion module, and the multi-relational airport network is fused into a single-relational airport network through a data-driven method.

[0010] Step S3: Use a bidirectional GCN to capture the spatial structure features of the single-relationship airport network, and replace the linear calculation function in the GRU with a bidirectional GCN to form a spatio-temporal modeling unit. According to the input traffic flow length, adaptively expand this spatio-temporal modeling unit to extract the spatio-temporal features of the airport traffic flow dataset, construct an airport traffic flow prediction model, and optimize the traffic flow prediction model using a loss function.

[0011] As a preferred solution of the present invention, for an airport traffic flow prediction method based on a graph neural network, in step S1, an interpolation method is used to complete the airport traffic flow observation data and aviation meteorological data, and an airport traffic flow dataset is constructed.

[0012] As a preferred solution of the present invention, for an airport traffic flow prediction method based on a graph neural network, label annotation is performed on the airport traffic flow dataset, and the label features include temperature, dew point, reflectivity, wind direction, wind speed, visibility, gust, takeoff flow, and landing flow.

[0013] As a preferred solution of the present invention, for an airport traffic flow prediction method based on a graph network, in step S1, an airport route network layer is constructed according to flight schedule data to describe the planned semantic relationship between airports; a flight duration network layer is constructed according to the flight duration data between airport pairs to describe the flight operation process between airports; a flight schedule network layer is constructed according to the airport route network data to describe the geographical information relationship between airports; a multi-relationship airport network is obtained, and the topological structure of the multi-relationship airport network is uniformly described in the form of an adjacency matrix.

[0014] As a preferred solution of the present invention, for an airport traffic flow prediction method based on a graph neural network, in step S2, an attention mechanism is used to construct a multi-relationship graph data fusion method, and the spatial topological information between airports is automatically extracted in a data-driven manner. The formula for the multi-relationship graph data fusion method is as follows:

[0015]

[0016] where, A i is the adjacency matrix of the airport network topological structure, W i is a learnable attention coefficient matrix, and A learned is the adjacent matrix of the finally learned airport network topological structure.

[0017] As a preferred solution of the present invention, for an airport traffic flow prediction method based on a graph neural network, a bidirectional GCN is used to capture the spatial structure features of the single-relationship airport network, and the formula for the bidirectional GCN is as follows:

[0018]

[0019] where, A is the adjacency matrix; is the degree matrix, where is the transpose matrix of; W0 and W1 are the weight coefficient matrices of the adjacency matrix respectively.

[0020] As a preferred solution of the present invention, a method for predicting airport traffic flow based on a graph neural network, the GRU formula is as follows:

[0021] r t = σ(W r [x t , h t-1 + b r )

[0022] z t = σ(W z [x t , h t-1 + b z )

[0023] c t = tanh(W c [x t , r t * h t-1 + b c )

[0024] h t = (1 - z t ) * h t-1 + z t * c t

[0025] where, r t represents the reset gate, z t represents the update gate, c t represents the candidate hidden layer, h t is the hidden state at time t, h t-1 is the hidden state at time t - 1, x t is the feature data at time t, A is the adjacency matrix; σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function; W r is the coefficient matrix of the reset gate r t , W z is the coefficient matrix of the update gate z t , W c is the coefficient matrix of the candidate hidden layer, c t ; b z , b r , b c represent the bias value of the update gate, the reset gate value, and the bias value of the candidate layer respectively.

[0026] As a preferred solution of the present invention, a method for predicting airport traffic flow based on a graph neural network, the formula of the airport traffic flow prediction model is as follows:

[0027] r t = σ(W r [f(x t , A), h t-1 + b r )

[0028] z t = σ(W z [f(x t , A), h t-1 + b z )

[0029] c t = tanh(W c [f(x t , A), r t * h t-1 + b c )

[0030] h t = (1 - z t ) * h t-1 + z t * c t

[0031] Among them, f(·) is a bidirectional GCN calculation unit, r t represents the reset gate, z t represents the update gate, c t represents the candidate hidden layer, h t is the hidden state at time t, h t-1 is the hidden state at time t - 1, x t is the feature data at time t, A is the airport network adjacency matrix; σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function; W r is the coefficient matrix of the reset gate r t , W z is the coefficient matrix of the update gate z t , W c is the coefficient matrix of the candidate hidden layer, c t ; b z , b r , b c respectively represent the bias value of the update gate, the reset gate value, and the bias value of the candidate layer.

[0032] As a preferred embodiment of the present invention, a method for predicting airport traffic flow based on a graph neural network uses the MSE function as the loss function, and the formula is as follows:

[0033]

[0034] where y is the label information, is the output vector of the airport traffic flow prediction model.

[0035] As a preferred embodiment of the present invention, a method for predicting airport traffic flow based on a graph neural network, the evaluation index of the traffic flow prediction model is:

[0036]

[0037] where RMSE is the root mean square error; MAE is the mean absolute error; y is the label information, is the output vector of the airport traffic flow prediction model.

[0038] As a preferred embodiment of the present invention,

[0039] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are:

[0040] The present invention models multiple air traffic elements such as flight duration, flight schedule, and route network structure to form a multi-relational airport network topology relationship network, applies the attention mechanism to fuse multiple spatial relationships to form a unified representation of the airport traffic network, extracts useful topological information on multiple airport network layers based on the bidirectional GCN model, combines GRU to effectively extract the change trend of network-level traffic flow in the time dimension, establishes a large-scale and multi-step airport traffic flow prediction model, realizes the prediction of airport traffic flow in large-scale and multi-airport scenarios, can quickly generate network-level airport traffic flow prediction results, does not need to establish a prediction model for each airport separately, improves the model training efficiency, and helps air traffic managers timely perceive the traffic situation in the airport airspace, provides data support for formulating flow control management measures, and improves the airspace utilization rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a schematic flowchart of the technical solution of the present invention.

[0042] Figure 2 Unified representation diagram of the airport traffic network structure.

[0043] Figure 3 Schematic diagram of the annotation of the airport traffic flow dataset.

[0044] Figure 4 Airport traffic flow prediction model diagram.

[0045] Figure 5 Landing flow prediction chart of Beijing Capital International Airport.

[0046] Figure 6 Landing flow prediction chart of Guangzhou Baiyun Airport.

[0047] Figure 7 Landing flow prediction chart of Shanghai Pudong Airport.

[0048] Figure 8 Landing flow prediction chart of Chengdu Shuangliu International Airport.

[0049] Reference numerals: 1 - Adjacency matrix. Detailed implementation manners

[0050] The present invention will be described in detail below with reference to the accompanying drawings.

[0051] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0052] Embodiment 1

[0053] The airport traffic flow prediction models mentioned in the present invention all refer to large-scale, multi-airport air traffic flow prediction models, and neural network methods are used to mine the airport network traffic status. Commonly used neural network models include artificial neural network method (ANN), convolutional neural network (CNN), recurrent neural network (RNN), graph convolutional neural network (GCN), gated recurrent unit (GRU), long short-term memory network (LSTM), etc.

[0054] As Figure 1 shown, a method for predicting airport traffic flow based on a graph neural network, based on airport meteorological factors and multi-relational airport network spatial topology, constructs an effective spatio-temporal graph modeling method to mine the spatio-temporal correlation and evolution pattern of airport traffic flow data, and constructs an airport traffic flow prediction model for large-scale, multi-airport scenarios, mainly including the following steps:

[0055] Step S1, obtaining air traffic element data, including airport route network data, flight schedule data, flight duration data between airport pairs, aviation meteorological data, and airport traffic flow observation data; constructing a multi-relational airport network through airport route network data, flight schedule data, and flight duration data between airport pairs; constructing an airport traffic flow data set through airport traffic flow observation data and aviation meteorological data.

[0056] Specifically, the airport traffic flow observation data is collected from the system of a certain air traffic control center of the Civil Aviation Administration of China; by manually screening out the abnormal points in the airport traffic flow observation data, the three - spline interpolation algorithm is used to fill in the missing data, and an airport traffic flow data set is constructed. Finally, according to the flight status, the airport traffic flow is classified to obtain two types of data: the take - off flow and the landing flow of the airport;

[0057] The aviation meteorological data uses the Aviation Routine Weather Report (METAR) as the data source; METAR is a common format for transmitting observed meteorological data. After it was standardized by the International Civil Aviation Organization, the World Meteorological Organization standardized its codes, making its compilation specifications unified in countries around the world. METAR is usually generated every half hour or hour and is issued by airports or some meteorological observation stations; the three - spline interpolation algorithm is used to fill in the missing data.

[0058] The national airway network structure is statistically analyzed. According to the airway information between airport pairs in historical flight plans, the number of airway convergence points passed by flights between city pairs is calculated to obtain the airport airway network data; the flight schedule in 2017 is selected as the data source for flight schedule data in Example 1;

[0059] The actual flight duration data between airport pairs is collected from the system of a certain air traffic control center of the Civil Aviation Administration of China. By tracking the implementation of flight plans, the actual take - off time and actual landing time are obtained, and the average flight duration of flights between each airport pair is statistically calculated.

[0060] Specifically, based on the flight schedule, an airport navigation matrix is constructed with airports as nodes. If there is a navigation flight between airport pairs, it is marked as 1; if there is no navigation flight between airport pairs, it is marked as 0. Considering that air traffic has strong planning and flight operations need to follow the flight schedule, the total flight volume information between city pairs within a certain period is statistically calculated based on the flight schedule to construct an edge weight matrix, and an airport airway network layer is constructed to describe the planned semantic relationship between airports; considering the impact of the flight process, using the flight duration data between airport pairs, the flight duration data between city pairs is statistically calculated, and with the flight duration as the edge weight, a flight duration network layer is modeled to describe the flight operation process between airports; considering the impact of the airway structure, using the airport airway network data, the number of airway convergence points passed by between city pairs is statistically calculated, and with the number of passed convergence points as the edge weight, a flight schedule network layer is modeled to describe the geographical information relationship between airports. Combining the airport airway network layer, the flight duration network layer and the flight schedule network layer, a multi - relationship airport network is obtained, and the adjacency matrix is used to uniformly describe the topological structure of the multi - relationship airport network, and the representation form is as Figure 2 shown.

[0061] By analyzing the flight statistics between city pairs, it can be found that the inbound and outbound flight volumes between cities are different. For example, for Beijing Capital International Airport ZBAA and Shanghai Hongqiao International Airport ZSSS in June 2017, the flight volume from ZBAA to ZSSS is 947, and the flight volume from ZSSS to ZBAA is 932. Therefore, the multi-relationship airport network constructed in this Embodiment 1 is a directed graph.

[0062] Specifically, determine the number of airports in the concerned airspace, wait for N moments, complete the accumulation of traffic flow data and the collection of aviation meteorological data for multiple airports. After the data accumulation is completed, the input feature data of the airport traffic flow prediction model is formed through normalization. The feature data is labeled in a sliding window manner to construct an airport traffic flow dataset, which specifically includes:

[0063] Input feature: net_ap = {ap_1, ap_2, ap_3, ap_4, … ap_i}, (i = 1, 2, 3 … w)

[0064] Label feature: net_ap_flow = {ap_flow_1, ap_flow_2, ap_flow_3, ap_flow_4, … ap_flow_i}, (i = 1, 2, 3 … w)

[0065] The set of traffic flow and aviation meteorological data of multiple airports within a certain time and space range is called the airport traffic feature set; ap_i is the airport information, and i represents the airport number; the airport information structure is ap_i_t, ap_i_t = {tmpf_i_t, dwpf_i_t, relh_i_t, drct_i_t, sknt_i_t, vsby_i_t, gust_i_t, dep_i_t, arr_i_t}, where tmpf_i_t represents the temperature of airport i at time t, dwpf_i_t represents the dew point of airport i at time t, relh_i_t represents the reflectivity of airport i at time t, drct_i_t represents the wind direction of airport i at time t, sknt_i_t represents the wind speed of airport i at time t, vsby_i_t represents the visibility of airport i at time t, gust_i_t represents the gust of airport i at time t, dep_i_t represents the takeoff flow of airport i at time t, and arr_i_t represents the landing flow of airport i at time t. The label feature is two types of data, the takeoff flow and landing flow of the airport, marked as ap_flow_i_t, where ap_flow_i_t = {dep_i_t, arr_i_t}, dep_i_t represents the takeoff flow of airport i at time t, and arr_i_t represents the landing flow of airport i at time t.

[0066] After window sliding, the construction of the airport traffic flow dataset is completed, and the formula is as follows:

[0067] dt_lb = (feature_ds{net_ap}, label_ds{net_ap_flow})

[0068] Among them, feature_ds{net_ap} represents the feature data set; label_ds{net_ap_flow} represents the label set corresponding to the training data, and dt_lb represents the association structure between the feature data set and the label set.

[0069] As Figure 3 shown in the label annotation process, the window width is n, the sliding step is 1, and the prediction step is m. Assuming that the starting time is 1, the input feature data data_feature = <net_ap_1,..., net_ap_n>, and the corresponding label feature data_label = <net_ap_flow_n+1,..., net_ap_flow_n+m>.

[0070] In this Embodiment 1, the time step length of the airport traffic flow dataset is 30 minutes.

[0071] Step S2, based on the multi-relationship airport network, use the attention mechanism to construct a multi-relationship airport network fusion module, and fuse the multi-relationship airport network into a single-relationship airport network through a data-driven method.

[0072] Specifically, use the attention mechanism to construct a multi-relationship graph data fusion method, and automatically extract effective spatial topological information between airports based on the data-driven method. The formula of the multi-relationship graph data fusion method is as follows:

[0073]

[0074] Among them, A i is the adjacency matrix of the airport network topological structure, W i is the learnable attention coefficient matrix, and A learned is the adjacency matrix of the finally learned airport network topological structure.

[0075] Step S3, use the bidirectional GCN to capture the spatial structure features of the single-relationship airport network, use the bidirectional GCN to replace the linear calculation function in GRU to form a spatio-temporal modeling unit; according to the input traffic flow length, adaptively expand the spatio-temporal modeling unit to extract the spatio-temporal features of the airport traffic flow dataset, construct an airport traffic flow prediction model, and optimize the traffic flow prediction model using the loss function.

[0076] Specifically, since the airport traffic network is a graph-structured data, for the spatial relationships on the airport traffic network, a GCN is used to capture the spatial structural features of the single-relationship airport network; in addition, the airport traffic flow has strong temporal characteristics, and both LSTM and GRU are suitable for extracting the characteristics of temporal data in terms of structure, and their prediction capabilities are also close. However, the GRU structure has a faster convergence speed than LSTM. Therefore, in this Embodiment 1, the airport traffic network structure is uniformly represented by a graph structure, and at the same time, considering the spatio-temporal correlation and randomness between airport traffic flow data, a bidirectional GCN is used to replace the linear calculation function in GRU to form a spatio-temporal modeling unit, and according to the input traffic flow length, this spatio-temporal modeling unit is adaptively expanded to extract the spatio-temporal features of the airport traffic flow dataset, and an airport traffic flow prediction model as shown in Figure 4 is constructed to achieve traffic flow prediction for a large range and multiple airports.

[0077] Furthermore, the core calculation formula of the GCN model is as follows:

[0078]

[0079] where A is the adjacency matrix; is the degree matrix, where is the transpose matrix of; W0 and W1 are the weight coefficient matrices of the adjacency matrix respectively.

[0080] Furthermore, the core calculation formula of the GRU model is as follows:

[0081] r t =σ(W r [x t ,h t-1 +b r )

[0082] z t =σ(W z [x t ,h t-1 +b z )

[0083] c t =tanh(W c [x t ,r t *h t-1 +b c )

[0084] h t =(1 - z t )*h t-1 +z t *ct

[0085] Among them, r t Represents the reset gate, z t represents the update gate, c t represents the candidate hidden layer, h t is the hidden state at time t, h t-1 is the hidden state at time t-1, x t is the feature data at time t, A is the adjacency matrix; σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function; W r Reset gate r t The coefficient matrix, W z To update the gate z t The coefficient matrix, W c is the candidate hidden layer, c t The coefficient matrix of b z 、b r 、b c They represent the update gate bias value, reset gate value, and bias value of the candidate layer respectively.

[0086] Furthermore, the bidirectional GCN and GRU are combined to construct an airport traffic flow prediction model. The core calculation formula is as follows:

[0087] r t =σ(W r [f(x t ,A),h t-1 ]+b r )

[0088] z t =σ(W z [f(x t ,A),h t-1 ]+b z )

[0089] c t =tanh(W c [f(x t ,A),r t *h t-1 ]+b c )

[0090] h t =(1-z t )*h t-1 +z t *c t

[0091] Among them, f(·) is the bidirectional GCN computing unit, r t Represents the reset gate, z t represents the update gate, ct Denotes the candidate hidden layer, h t Is the hidden state at time t, h t-1 Is the hidden state at time t-1, x t Is the feature data at time t, A is the airport network adjacency matrix; σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function; W r Is the reset gate r t Coefficient matrix of z Is the update gate z t Coefficient matrix of c [[ID=I7]]Is the candidate hidden layer, c t Coefficient matrix of; b z b r b c Respectively represent the update gate bias value, reset gate value, and candidate layer bias value.

[0092] Furthermore, the MSE function is used as the loss function of the airport traffic flow prediction model to measure the quality of the model prediction, which is manifested as the gap between the model prediction result and the actual data. The formula is as follows:

[0093]

[0094] Among them, y is the label information, Is the output vector of the airport traffic flow prediction model.

[0095] The root mean square error (RMSE) and mean absolute error (MAE) functions are used as the evaluation indicators of the airport traffic flow prediction model. The formulas include:

[0096]

[0097] Among them, y represents the label information, Is the model prediction output vector.

[0098] To sum up, for multiple air traffic elements such as flight duration, flight schedule, and route network structure, a model is built and uniformly represented. The attention mechanism is applied to form a unified representation of various air traffic elements. Combining the bidirectional GCN model and the GRU model, useful information on the airport traffic network structure is extracted, and a multi-time step prediction method for large-scale airport traffic flow is established. There is no need to build a prediction model for each airport separately and calculate them separately, and the network-level airport traffic flow prediction result can be calculated quickly.

[0099] Example 2

[0100] Obtain the airport traffic flow observation data and aviation meteorological data of a certain control center in 2017, and input this set of network-level airport traffic flow data and aviation meteorological data into the airport traffic flow prediction model. In the implementation of this model, the function network_flow_predict is used as the interface of the airport traffic flow prediction model. After calculation, the network_flow_predict function will return the predicted set of network-level airport traffic flow data.

[0101] Apply the GRU, historical average (HA) method, and the airport traffic flow prediction model (A-ATFPNet) proposed in this invention respectively to predict the network-level airport traffic flow data. Among them, the input length of the A-ATFPNet and GRU models is 12 time steps, and the interval of each time step is 30 minutes. The output length of the model is 4 time steps. The inputs and outputs in the model training process are as follows:

[0102] (1) Input: The aviation meteorological data and traffic flow data of multiple airports at the network level with a time span of 6 hours.

[0103] (2) Output: The landing flow data of the network-level airport with a time span of 2 hours.

[0104] The summary of the prediction results is shown in Table 1-1. It can be found from this table that the A-ATFPNet model has better prediction accuracy in terms of RMSE and MAE indicators.

[0105] Table 1-1 Comparison table of prediction results

[0106]

[0107] To facilitate the inspection of the actual prediction performance of the method described in this invention, visualize the actual landing flow and predicted landing flow of Beijing Capital International Airport (ZBAA), Guangzhou Baiyun Airport (ZGGG), Shanghai Pudong Airport (ZSPD), and Chengdu Shuangliu International Airport (ZUUU) respectively. The prediction results are as Figures 5 - 8 shown, and the prediction results obtained by this method are basically consistent with the actual landing flow of the airport.

[0108] In summary, as can be seen from the specific implementation manner of this invention, the large-scale airport traffic flow prediction model based on graph neural proposed in this invention can effectively capture the useful information of the traffic network structure diagrams of multiple airports; based on the GRU model, it can effectively obtain the changing trend of time series, thereby effectively improving the accuracy of airport traffic flow estimation, realizing more scientific formulation of effective traffic management measures, effectively improving air traffic management efficiency, and providing a strong guarantee for aviation safety.

[0109] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting airport traffic flow based on graph convolutional neural network, characterized by: include: Step S1, obtaining air traffic element data, including airport route network data, flight schedule data, flight duration data between airport pairs, aviation meteorological data, and airport traffic flow observation data; constructing a multi-relational airport network based on the airport route network data, flight schedule data, and flight duration data between airport pairs; and constructing an airport traffic flow dataset based on the airport traffic flow observation data and aviation meteorological data; Step S2: Based on the multi-relational airport network, an attention mechanism is used to construct a multi-relational airport network fusion module, and the multi-relational airport network is fused into a single-relational airport network through a data-driven method; Among them, the fusion adopts the attention mechanism to construct a multi-relationship graph data fusion method, and automatically extracts the spatial topology information between airports based on a data-driven approach. The formula of the multi-relationship graph data fusion method is as follows: Among them, A i is the adjacency matrix of the airport network topology, W i is the learnable attention coefficient matrix, A learned is the adjacency matrix of the final learned airport network topology; In step S3, bidirectional GCN is used to capture the spatial structural characteristics of the single-relation airport network, and bidirectional GCN is used to replace the linear calculation function in GRU to form a spatiotemporal modeling unit; according to the input traffic flow length, the spatiotemporal modeling unit is adaptively expanded to extract the spatiotemporal characteristics of the airport traffic flow dataset, an airport traffic flow prediction model is constructed, and the traffic flow prediction model is optimized using a loss function.

2. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 1 is characterized in that: In step S1, the airport traffic flow observation data and aviation meteorological data are supplemented by interpolation method to construct an airport traffic flow dataset.

3. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 2 is characterized in that: The airport traffic flow dataset is labeled, and the label features include temperature, dew point, reflectivity, wind direction, wind speed, visibility, gust, takeoff flow and landing flow.

4. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 2 is characterized in that: In step S1, an airport route network layer is constructed based on the flight schedule data to describe the planned semantic relationship between airports; a flight duration network layer is constructed based on the flight duration data between the airport pairs to describe the flight operation process between airports; Constructing a flight schedule network layer based on the airport route network data to describe the geographic information relationship between airports; A multi-relational airport network is obtained, and the topological structure of the multi-relational airport network is uniformly described in the form of an adjacency matrix.

5. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 1, characterized in that: Bidirectional GCN is used to capture the spatial structural characteristics of the single-relation airport network. The bidirectional GCN formula is as follows: Where A is the adjacency matrix; is the degree matrix, where for The transposed matrix; σ is the sigmoid activation function, W0 and W1 are the adjacency matrices The weight coefficient matrix of .

6. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 1, characterized in that: The GRU formula is as follows: r t =σ(W r [x t ,h t-1 ]+b r ) z t =σ(W z [x t ,h t-1 ]+b z ) c t =tanh(W c [x t ,r t *h t-1 ]+b c ) h t =(1-z t )*h t-1 +z t *c t Among them, r t Represents the reset gate, z t represents the update gate, c t represents the candidate hidden layer, h t is the hidden state at time t, h t-1 is the hidden state at time t-1, x t is the feature data at time t, A is the adjacency matrix; σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function; W r Reset gate r t The coefficient matrix, W z To update the gate z t The coefficient matrix, W c is the candidate hidden layer c t The coefficient matrix of b z 、b r 、b c They represent the bias value of the update gate, the reset gate value, and the bias value of the candidate layer respectively.

7. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 1, characterized in that: The formula of the airport traffic flow prediction model is as follows: r t =σ(W r [f(x t ,A),h t-1 ]+b r ) z t =σ(W z [f(x t ,A),h t-1 ]+b z ) c t =tanh(W c [f(x t ,A),r t *h t-1 ]+b c ) h t =(1-z t )*h t-1 +z t *c t Among them, f(·) is the bidirectional GCN computing unit, r t Represents the reset gate, z t represents the update gate, c t represents the candidate hidden layer, h t is the hidden state at time t, h t-1 is the hidden state at time t-1, x t is the characteristic data at time t, A is the airport network adjacency matrix; σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function; W r Reset gate r t The coefficient matrix, W z To update the gate z t The coefficient matrix, W c is the candidate hidden layer c t The coefficient matrix of b z 、b r 、b c They represent the bias value of the update gate, the reset gate value, and the bias value of the candidate layer respectively.

8. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 1, characterized in that: Using the MSE function as the loss function, the formula is as follows: Among them, y is the label information, is the output vector of the airport traffic flow prediction model.

9. The method for predicting airport traffic flow based on graph convolutional neural network according to claim 4, characterized in that: The evaluation index of the traffic flow prediction model is: Among them, RMSE is the root mean square error; MAE is the mean absolute error.

Images