Civil aviation passenger flow volume prediction method based on dynamic graph convolution

Abstract

The invention discloses a civil aviation passenger flow volume prediction method based on dynamic graph convolution, and relates to the technical field of navigation station management, and the method comprises the following steps: carrying out attribute information and structure information extraction in advance, extracting navigation station attribute information by using a multi-layer perceptron MLP, including a plurality of feature attributes of flight plan, regional visa policy and regional weather, and carrying out structure information extraction; the method comprises the following steps of: extracting navigation station structure information by applying a Steuc2Vec method, establishing edges with passenger flow volume exceeding a threshold value between navigation station buildings to construct a graph structure, splicing navigation station attribute information extracted by a multi-layer perceptron (MLP) and the navigation station structure information extracted by the Steuc2Vec method to represent feature information, and taking the feature information as initial input of GNN (Global Navigation Network); according to the method, the GRU and the GNN are introduced, and the navigation station feature information and the structure information are fused, so that the accuracy and the stability of the navigation station passenger flow volume prediction model are improved, the method has the capabilities of flexibly selecting the prediction model and effectively predicting the navigation station passenger flow volume, and a scientific decision basis is provided for navigation station management and planning.

Description

Technical Field

[0001] The present invention relates to the technical field of terminal management, and in particular to a civil aviation passenger flow prediction method based on dynamic graph convolution. Background Art

[0002] Simply put, in terminal management and planning, accurate passenger flow forecasting is of great significance for rationally arranging resources and providing efficient services. However, terminal passenger flow is affected by many factors, such as flight schedules, ticket prices, regional visa policies, etc. These factors are interrelated, and traditional forecasting methods are difficult to fully capture these complex relationships.

[0003] More specifically, the existing background technology has some defects:

[0004] 1. Statistical models based on linear assumptions: Classical time series models are based on the assumption that there is a linear relationship between past and future values. They cannot handle nonlinear relationships well and lack correlation considerations. Existing models do not fully consider the impact of changes, trends, and randomness of various factors on passenger flow, resulting in low prediction accuracy.

[0005] 2. Only use time series information modeling (such as RNN), mainly based on the historical data of current airport forecasts, to predict future passenger traffic, and regard this problem as a regression prediction based on neural networks.

[0006] On the one hand, the RNN model alone cannot model and capture the complex spatial dependencies between different airports. The traffic distribution of each airport station is closely related to its neighboring stations. Secondly, due to the existence of fixed flight routes, traffic flows between long-distance stations will also affect each other. For example, there are many scheduled flights between Beijing and Shanghai. Although they are far apart, there is still an important spatial dependency between them. Finally, the flight path of the aircraft has a fixed route (multi-hop edges in the graph structure, such as Beijing-Shanghai-Hong Kong-Hawaii), which leads to dependencies on the areas traversed by a specific route, which we call long path dependencies in this article.

[0007] On the other hand, for civil aviation travel data with variable characteristics, existing models are difficult to successfully model the help of different time series information for prediction, such as ticket price changes, passenger flow, and regional visa policies in the long term.

[0008] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0009] In view of the problems in the related technology, the present invention proposes a civil aviation passenger flow prediction method based on dynamic graph convolution to overcome the above-mentioned technical problems existing in the existing related technology.

[0010] The technical solution of the present invention is achieved in this way:

[0011] A civil aviation passenger flow prediction method based on dynamic graph convolution includes the following steps:

[0012] Attribute information and structural information are extracted in advance. The multi-layer perceptron MLP is used to extract terminal attribute information, including multiple characteristic attributes such as flight schedule, regional visa policy and regional weather. The Struc2Vec method is used to extract terminal structural information. The edges between terminals with passenger flow exceeding the threshold are established to construct a graph structure. The terminal attribute information extracted by the multi-layer perceptron MLP and the terminal structural information extracted by the Struc2Vec method are concatenated to represent the characteristic information as the initial input of the GNN.

[0013] GNN message passing modeling terminal neighborhood information, which constructs GNN to analyze the characteristic information of the terminal, takes the terminal as the node of the graph, and forms edges with the connection between the terminals where the passenger flow exceeds the threshold. A graph structure with the terminal as the node and the passenger flow as the edge weight is established. GNN is used to process the graph structure, and the neighbor node information of each node is aggregated to update the node representation to capture the correlation between the terminals.

[0014] GRU models the time series information of the terminal. It uses GRU to model the output of GNN in time series, takes the output of GNN as the input of GRU, and uses the memory unit and gating mechanism of GRU to model the characteristic information of the terminal in time series.

[0015] Carry out passenger flow forecasting, use the feature information and structural information extracted by GRU and GNN according to the prediction MLP model to predict passenger flow, and make corresponding terminal resource scheduling and decisions based on the prediction results.

[0016] The step of extracting terminal attribute information using a multi-layer perceptron MLP includes the following steps:

[0017] The feature attributes of different dimensions are concatenated to form a unified feature vector, including the following steps:

[0018] For each feature attribute, data preprocessing is performed to ensure that different features have the same value range or unified representation form;

[0019] The preprocessed feature attributes are upgraded by MLP according to the dimension to enhance the expression ability, which is expressed as:

[0020]

[0021] in, represents the attribute information vector of the terminal node v, x vRepresents node attribute information, and its dimension is set to 128.

[0022] The step of extracting terminal structure information using the Struc2Vec method includes the following steps:

[0023] The calibration S is the node embedding generated by Struc2Vec for the graph G, whose dimension is d, and the embedding representation is obtained as:

[0024] S = struc2vec(G, d);

[0025] Among them, S represents the set of structural information of all nodes in the terminal network. Therefore, the structural information vector h of node v is v (s) It can be expressed as:

[0026]

[0027] The step of aggregating the neighbor node information of each node to update the representation of the node includes: applying the message passing and aggregation process to the graph composed of the terminal nodes, including the following steps:

[0028] The message transmission between the calibration terminal nodes is expressed as:

[0029]

[0030] Among them, l represents the current level, and Respectively represent the feature representation of node v and its neighboring node u in the previous layer, represents the set of neighboring nodes of node i, w ij Represents the weight between nodes i and j, which is set according to the distance between terminals or past traffic data. Msg (l) is the message passing function at layer l, which is initialized to the form of degree normalized passing;

[0031] The terminal node aggregates the neighborhood information and updates its own representation:

[0032] Based on the message passing, each terminal node aggregates its own feature representation with the passed message to update the node's feature representation, which is expressed as:

[0033]

[0034] Among them, l represents the current level, represents the feature representation of node v in the previous layer, represents the delivery message received by node v at layer l, CoM (l)(*) is the aggregation function at layer l, which combines the node’s own features with the transmitted message to generate a new information representation.

[0035] The step of performing time series modeling on the characteristic information of the terminal includes the following steps:

[0036] Define the node representation matrix Represents the node representation of the lth GNN layer above, and at the same time, there is a representation matrix from the previous time step t-1 The goal is to update the representation of a node by combining these two, expressed as:

[0037]

[0038] GRU updates node timing information, including:

[0039] Calibrate the update gate: Calculate the update gate vector Z t , which is used to control whether to update the hidden state of the current time step. The calculation method is as follows:

[0040]

[0041] Among them, W z is a learnable weight matrix, [,] represents a concatenation operation;

[0042] Calibrate the reset gate: Calculate the reset gate vector R t , which is used to control whether to reset the hidden state of the current time step. The calculation method is as follows:

[0043]

[0044] Among them, W r is a learnable weight matrix;

[0045] Calibrate candidate hidden states: Calculate candidate hidden states by combining the reset gate and the hidden state of the previous time step. The calculation method is as follows:

[0046]

[0047] Among them, W h is a learnable weight matrix;

[0048] Calibrate the current hidden state: According to the update gate and the candidate hidden state, calculate the hidden state of the current time step. The calculation method is as follows:

[0049]

[0050] in, means keeping the hidden state of the previous time step, Indicates using the candidate hidden state to update the hidden state.

[0051] Among them, the prediction of passenger flow is performed using the feature information and structural information extracted by GRU and GNN according to the prediction model, including the following steps:

[0052] Use an MLP prediction function to predict the weighted passenger flow y of the edge from node v to node u vu , expressed as:

[0053]

[0054] The node representation and the edge weights between them are combined and predicted through the MLP model to infer the passenger flow between different nodes. Then, the known data is used for training to reduce the loss, which is expressed as:

[0055]

[0056] Beneficial effects of the present invention:

[0057] 1. Comprehensively capture the correlation between terminals: By introducing GRU and GNN, this method can more comprehensively capture the correlation between different features. GRU can model time series data, while GNN can process graph structure data. This fusion enables the model to comprehensively consider the interaction between feature information and structural information, thereby improving the accuracy and stability of the prediction model.

[0058] 2. Flexible prediction model selection: This method also has the ability to select different prediction models according to actual needs, such as regression models. This allows the prediction model to be adjusted according to the nature of the specific prediction problem to obtain better prediction results. This flexibility makes this method suitable for a variety of different types of prediction tasks.

[0059] 3. Effectively predict passenger flow at terminals in a larger area: The fusion model proposed by this invention can effectively predict passenger flow at terminals in various regions around the world. Accurate prediction of passenger flow is crucial for terminal management and planning. It can help decision makers arrange resources scientifically, provide efficient services, and provide reliable decision-making basis for the operation of different terminals.

[0060] In summary, by introducing GRU and GNN and integrating terminal feature information with structural information, this invention improves the accuracy and stability of the terminal passenger flow prediction model. At the same time, this method also has the ability to flexibly select prediction models and effectively predict terminal passenger flow, providing a scientific decision-making basis for terminal management and planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] Figure 1 It is a schematic diagram of a flow chart of a civil aviation passenger flow prediction method based on dynamic graph convolution according to an embodiment of the present invention;

[0063] Figure 2 is a schematic diagram of a scenario of a civil aviation passenger flow prediction method based on dynamic graph convolution according to an embodiment of the present invention;

[0064] Figure 3 It is a schematic diagram of GNN and GRU joint modeling of a civil aviation passenger flow prediction method based on dynamic graph convolution according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0066] According to an embodiment of the present invention, a civil aviation passenger flow prediction method based on dynamic graph convolution is provided.

[0067] like Figure 1-Figure 3 As shown, the civil aviation passenger flow prediction method based on dynamic graph convolution according to an embodiment of the present invention includes the following steps:

[0068] Step S1, extract attribute information and structural information in advance, and use the multi-layer perceptron MLP to extract terminal attribute information, including multiple characteristic attributes such as flight schedule, regional visa policy and regional weather. Use the Struc2Vec method to extract terminal structural information, and establish edges between terminals where passenger flow exceeds the threshold to construct a graph structure. The terminal attribute information extracted by the multi-layer perceptron MLP and the terminal structural information extracted by the Struc2Vec method are concatenated to represent the characteristic information as the initial input of the GNN.

[0069] Among them, the multi-layer perceptron MLP is used to extract terminal attribute information, which includes multiple characteristic attributes such as flight schedule, regional visa policy and regional weather. Each characteristic attribute can be represented as a vector or matrix, and the specific dimensions are as follows:

[0070] Among them, flight plan: flight plan information may include departure time, arrival time, flight route, etc. This information can be expressed in the form of timestamp, flight number, and route code. For example, the departure time can be expressed as a timestamp vector, the arrival time can be expressed as another timestamp vector, and the flight route can be expressed as a unique hot vector of the route code.

[0071] Among them, regional visa policy: Regional visa policy can usually be encoded as a binary vector, where each dimension represents a visa policy status, such as visa, visa-free, visa on arrival, etc. The value of each dimension can be 0 or 1, indicating the existence or non-existence of the corresponding policy status.

[0072] Among them, regional weather: Regional weather information usually includes temperature, precipitation, wind speed, etc. This information can be represented as a numerical vector, with each dimension representing a weather indicator, such as temperature, precipitation, and wind speed.

[0073] In addition, the feature attributes of different dimensions are concatenated to form a unified feature vector. The specific processing process includes the following steps:

[0074] For each feature attribute, appropriate data preprocessing is performed, such as standardization, normalization, or one-hot encoding, to ensure that different features have the same numerical range or unified representation.

[0075] The preprocessed feature attributes are upgraded by MLP according to the dimension to enhance the expression ability, which is expressed as:

[0076]

[0077] in, represents the attribute information vector of the terminal node v, x v Represents node attribute information (such as flight schedule), and its dimension is set to 128.

[0078] In addition, the edges between terminals with passenger flow exceeding the threshold are established to construct the terminal relationship dependency graph, including:

[0079] According to the past traffic dependency and geographical location between terminals, a terminal relationship dependency graph is established. The terminal relationship dependency graph can be divided into transfer relationship, one-way direct relationship and two-way relationship, and the sparsity of the terminal relationship dependency graph can be determined according to a certain threshold. The details are as follows:

[0080] Among them, the terminal relationship dependency graph is constructed:

[0081] First, based on the past traffic dependency and geographic location information between the terminals, an undirected or directed graph is constructed to represent the relationship between the terminals. A transfer relationship refers to a relationship of transfer through one or more transfer terminals, a one-way direct relationship refers to a direct flight relationship in only one direction, and a two-way round-trip relationship refers to a round-trip flight relationship between two terminals;

[0082] Among them, constructing a sparse graph: In order to build a sparse graph, the relationship between terminals is screened according to a certain threshold. This threshold can be determined based on past traffic data, distance information or other relevant factors. Only when the relationship between terminals exceeds the threshold, it is considered a valid edge, thus constructing a sparse graph;

[0083] Specifically, the Struc2Vec method is used to extract the terminal structure information. After establishing the terminal relationship dependency graph, the Struc2Vec and other graph embedding algorithms are used to extract the structure information.

[0084] This technical solution, Struc2Vec, is an unsupervised graph embedding algorithm for learning node representation. It converts the structural information of the graph into a low-dimensional continuous vector representation by taking into account the structural neighborhood of the node. When applied, the structural information is also set to 128 dimensions, including the following steps:

[0085] The calibration S is the node embedding generated by Struc2Vec for the graph G, with dimension d. The embedding is obtained as follows:

[0086] S = struc2vec(G, d);

[0087] Among them, S represents the set of structural information of all nodes in the terminal network. Therefore, the structural information vector of node v is It can be expressed as:

[0088]

[0089] Therefore, the initial vector h of the initial input of the GNN is finally left v (The initial vector representation of terminal v) is the concatenation of the terminal node attribute information and the terminal structure information, expressed as:

[0090]

[0091] Step S2, GNN (graph neural network) message passing modeling terminal neighborhood information, constructing GNN (graph neural network) to analyze the characteristic information of the terminal. The terminal is regarded as the node of the graph, and the connection between the terminals with passenger flow exceeding the threshold forms an edge, and a graph structure with the terminal as the node and passenger flow as the edge weight is established. GNN is used to process the graph structure, and the neighbor node information of each node is aggregated to update the node representation to capture the correlation between the terminals.

[0092] This technical solution exchanges information between terminal nodes and aggregates and updates the node's own information through message passing, and models the terminal as a relationship graph, so that in the task of terminal traffic prediction, message passing between nodes is very important for obtaining information such as travel dynamics of neighboring terminals and surrounding areas. Through the message passing process, each terminal node can collect and share information about its neighboring nodes, thereby obtaining more comprehensive contextual information and further improving the accuracy of traffic prediction.

[0093] The process of message passing and aggregation is applied to the graph composed of terminal nodes. The specific steps are as follows:

[0094] Message passing between terminal nodes (MessagePassing):

[0095] For each terminal node v i , it will collect its neighboring terminal node u j This process allows the node to obtain relevant data such as traffic, flight information, number of passengers, etc. at neighboring airports. Message passing is represented as:

[0096]

[0097] Among them, l represents the current level, and Respectively represent the feature representation of node v and its neighboring node u in the previous layer, represents the set of neighboring nodes of node i, w ij Represents the weight between nodes i and j, which can be set based on the distance between terminals or past traffic data. (l) is the message passing function at layer l, which is initialized as a degree-normalized passing form.

[0098] The terminal node aggregates the neighborhood information and updates its own representation (Aggregation):

[0099] Based on the message passing, each terminal node aggregates its own feature representation with the passed message to update the node's feature representation, which is expressed as:

[0100]

[0101] Among them, l represents the current level, represents the feature representation of node v in the previous layer, Represents the delivery message received by node v at layer l. (l) (*) is the aggregation function at layer l, which combines the node’s own features with the transmitted message to generate a new information representation.

[0102] By iteratively applying the above message passing and aggregation steps, each terminal node can gradually update its feature representation to capture richer contextual information, continuously expand its field of view, model the travel demand in the local area, and help improve the accuracy of the terminal traffic prediction task.

[0103] With the help of the above solution, for different terminals, according to the frequency of passenger flow interaction between them, they are abstracted into graph structures, and various feature information and structural information are extracted. Then GNN is used: by introducing the GNN feature extractor, the relationship between terminals in different regions can be modeled, and information in a larger field of view can be extracted through message passing, and the feature expression ability of the data can be improved to improve the accuracy.

[0104] Step S3, GRU (Gated Recurrent Neural Network) models the time series information of the terminal, which uses GRU to perform time series modeling on the output of GNN. The output of GNN is used as the input of GRU, and the characteristic information of the terminal is modeled through the memory unit and gating mechanism of GRU. GRU can learn the time series change law of the characteristic information of the terminal, providing an accurate basis for the prediction of passenger flow.

[0105] In this technical solution, the time scale factor is very important when it comes to terminal traffic prediction. Passenger traffic data is constrained and affected by time, so to accurately predict future traffic conditions, it is necessary to consider the temporal dependencies in the time series. In order to capture this temporal information and incorporate it into the model, GRU (Gated Recurrent Unit, gated recurrent neural network) is used.

[0106] In this technical solution, GRU is a variant of recurrent neural network (RNN), which controls the information to be updated and forgotten at each step in the sequence through a gating mechanism. GRU can help the model model long-term dependencies and short-term dependencies, thereby better capturing information on the time scale.

[0107] In the terminal traffic prediction task, GRU can be applied to the output of GNN at each time step. Specifically, for each time step t, the output of GNN can be used as the input of GRU. GRU will update the hidden state of the current time step and generate the output of the current time step based on the input of the current time step and the hidden state of the previous time step.

[0108] Specifically, we first define the node representation matrix Represents the node representation of the lth GNN layer above. At the same time, there is a representation matrix from the previous time step t-1 The goal is to update the representation of a node by combining these two, expressed as:

[0109]

[0110] GRU updates the node timing information. The details are as follows:

[0111] Calibrate the update gate (UpdateGate): Calculate the update gate vector Z t , which is used to control whether to update the hidden state of the current time step. The calculation method is as follows:

[0112]

[0113] Among them, W z is a learnable weight matrix, and [,] represents a concatenation operation.

[0114] Calibrate the reset gate (ResetGate): Calculate the reset gate vector R t , which is used to control whether to reset the hidden state of the current time step. The calculation method is as follows:

[0115]

[0116] Among them, W r is a learnable weight matrix.

[0117] Calibrate candidate hidden state (CandidateHiddenState): Calculate the candidate hidden state by combining the reset gate and the hidden state of the previous time step. The calculation method is as follows:

[0118]

[0119] Among them, W h is a learnable weight matrix.

[0120] Current Hidden State: Based on the update gate and the candidate hidden state, the hidden state of the current time step is calculated as follows:

[0121]

[0122] in, means keeping the hidden state of the previous time step, Indicates using the candidate hidden state to update the hidden state.

[0123] In addition, GNN is used to jointly model terminal information. GRU successfully combines the information of time step and space step, captures the temporal dependency of terminal flow information at different times, and updates the representation of nodes, providing more accurate feature representation for terminal flow prediction tasks.

[0124] In addition, the output of each layer of GNN is used for GRU timing modeling, and GNN and GRU are jointly modeled, such as Figure 3 shown.

[0125] With the help of the above scheme, GRU performs time series processing on data from different regions encoded by GNN, and finally splices their results for prediction. Its advantages are: introducing the GRU network to solve the long-term and short-term dependency problems of time series data, solving the data dependency of samples through logic gating units, and further improving the prediction performance.

[0126] Step S4, passenger flow prediction is performed, and the passenger flow is predicted based on the feature information and structural information extracted by GRU and GNN according to the prediction model, and corresponding terminal resource scheduling and decision-making are performed based on the prediction results.

[0127] This technical solution predicts passenger flow based on prediction models (such as regression models) using feature information and structural information extracted by GRU and GNN. Based on the prediction results, corresponding terminal resource scheduling and decision-making can be carried out to improve the terminal's operating efficiency and service quality. The specific flow chart is as follows:

[0128] Generate node representation vectors, concatenate and predict. Based on the terminal node representations at different time steps calculated in the previous step, concatenate the representation vectors of the two nodes, and then use an MLP prediction function to predict the weighted passenger flow y of the edge from node v to node u vu , expressed as:

[0129]

[0130] Combine the node representation and the edge weights between them, and use the MLP model to make predictions, thereby inferring the passenger flow between different nodes, and then use known data for training to reduce losses. It can be expressed as:

[0131]

[0132] In summary, with the help of the above technical solution of the present invention, the following effects can be achieved:

[0133] 1. Comprehensively capture the correlation between terminals: By introducing GRU and GNN, this method can more comprehensively capture the correlation between different features. GRU can model time series data, while GNN can process graph structure data. This fusion enables the model to comprehensively consider the interaction between feature information and structural information, thereby improving the accuracy and stability of the prediction model.

[0134] 2. Flexible prediction model selection: This method also has the ability to select different prediction models according to actual needs, such as regression models. This allows the prediction model to be adjusted according to the nature of the specific prediction problem to obtain better prediction results. This flexibility makes this method suitable for a variety of different types of prediction tasks.

[0135] 3. Effectively predict passenger flow at terminals in a larger area: The fusion model proposed by this invention can effectively predict passenger flow at terminals in various regions around the world. Accurate prediction of passenger flow is crucial for terminal management and planning. It can help decision makers arrange resources scientifically, provide efficient services, and provide reliable decision-making basis for the operation of different terminals.

[0136] In summary, by introducing GRU and GNN and integrating terminal feature information with structural information, this invention improves the accuracy and stability of the terminal passenger flow prediction model. At the same time, this method also has the ability to flexibly select prediction models and effectively predict terminal passenger flow, providing a scientific decision-making basis for terminal management and planning.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. After considering the disclosure of the specification and the examples, those skilled in the art will easily think of other embodiments of the present disclosure. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art that are not disclosed in the present disclosure. The specification and the examples are only regarded as exemplary, and the true scope and spirit of the present disclosure are indicated by the claims.

[0138] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A civil aviation passenger flow prediction method based on dynamic graph convolution, characterized in that: The following steps are involved: Attribute information and structural information are extracted in advance. The multi-layer perceptron MLP is used to extract terminal attribute information, including multiple characteristic attributes such as flight schedule, regional visa policy and regional weather. The Struc2Vec method is used to extract terminal structural information. The edges between terminals with passenger flow exceeding the threshold are established to construct a graph structure. The terminal attribute information extracted by the multi-layer perceptron MLP and the terminal structural information extracted by the Struc2Vec method are concatenated to represent the characteristic information as the initial input of the GNN. GNN message passing modeling terminal neighborhood information, which constructs GNN to analyze the characteristic information of the terminal, takes the terminal as the node of the graph, and forms edges with the connection between the terminals where the passenger flow exceeds the threshold. A graph structure with the terminal as the node and the passenger flow as the edge weight is established. GNN is used to process the graph structure, and the neighbor node information of each node is aggregated to update the node representation to capture the correlation between the terminals. GRU models the time series information of the terminal. It uses GRU to model the output of GNN in time series, takes the output of GNN as the input of GRU, and uses the memory unit and gating mechanism of GRU to model the characteristic information of the terminal in time series. Carry out passenger flow forecasting, use the feature information and structural information extracted by GRU and GNN according to the forecasting model to forecast passenger flow, and make corresponding terminal resource scheduling and decisions based on the forecasting results.

2. The civil aviation passenger flow prediction method based on dynamic graph convolution according to claim 1 is characterized in that: The method of extracting terminal attribute information by using a multi-layer perceptron MLP includes the following steps: The feature attributes of different dimensions are concatenated to form a unified feature vector, including the following steps: For each feature attribute, data preprocessing is performed to ensure that different features have the same value range or unified representation form; The preprocessed feature attributes are upgraded by MLP according to the dimension to enhance the expression ability, which is expressed as: in, represents the attribute information vector of the terminal node v, x v Represents node attribute information, and its dimension is set to 128.

3. The civil aviation passenger flow prediction method based on dynamic graph convolution according to claim 2 is characterized in that: The steps of using the Struc2Vec method to extract terminal structure information include the following steps: The calibration S is the node embedding generated by Struc2Vec for the graph G, whose dimension is d, and the embedding representation is obtained as: S = struc2vec(G, d); Among them, S represents the set of structural information of all nodes in the terminal network. Therefore, the structural information vector of node v is It can be expressed as:

4. The civil aviation passenger flow prediction method based on dynamic graph convolution according to claim 1, characterized in that: The step of aggregating neighbor node information of each node to update the representation of the node includes: applying the process of message passing and aggregation to the graph composed of terminal nodes, including the following steps: The message transmission between the calibration terminal nodes is expressed as: Among them, l represents the current level, and Respectively represent the feature representation of node v and its neighboring node u in the previous layer, represents the set of neighboring nodes of node i, w ij Represents the weight between nodes i and j, which is set according to the distance between terminals or past traffic data. Msg (l) is the message passing function at layer l, which is initialized to the form of degree normalized passing; The terminal node aggregates the neighborhood information and updates its own representation: Based on the message passing, each terminal node aggregates its own feature representation with the passed message to update the node's feature representation, which is expressed as: Among them, l represents the current level, represents the feature representation of node v in the previous layer, represents the delivery message received by node v at layer l, CoM (l) (*) is the aggregation function at layer l, which combines the node’s own features with the transmitted message to generate a new information representation.

5. The civil aviation passenger flow prediction method based on dynamic graph convolution according to claim 1, characterized in that: The characteristic information of the terminal is subjected to time series modeling, including the following steps: Define the node representation matrix Represents the node representation of the lth GNN layer above, and at the same time, there is a representation matrix from the previous time step t-1 The goal is to update the representation of a node by combining these two, expressed as: GRU updates node timing information, including: Calibrate the update gate: Calculate the update gate vector Z t , which is used to control whether to update the hidden state of the current time step. The calculation method is as follows: Among them, W z is a learnable weight matrix, [,] represents a concatenation operation; Calibrate the reset gate: Calculate the reset gate vector R t , which is used to control whether to reset the hidden state of the current time step. The calculation method is as follows: Among them, W r is a learnable weight matrix; Calibrate candidate hidden states: Calculate candidate hidden states by combining the reset gate and the hidden state of the previous time step. The calculation method is as follows: Among them, W h is a learnable weight matrix; Calibrate the current hidden state: According to the update gate and the candidate hidden state, calculate the hidden state of the current time step. The calculation method is as follows: in, means keeping the hidden state of the previous time step, Indicates using the candidate hidden state to update the hidden state.

6. The civil aviation passenger flow prediction method based on dynamic graph convolution according to claim 1, characterized in that: The prediction model uses the feature information and structural information extracted by GRU and GNN to predict passenger flow, including the following steps: Use an MLP prediction function to predict the weighted passenger flow y of the edge from node v to node u vu , expressed as: The node representation and the edge weights between them are combined and predicted through the MLP model to infer the passenger flow between different nodes. Then, the known data is used for training to reduce the loss, which is expressed as:

Images