Method, device and system for verifying airport network delay propagation causality

Abstract

This invention discloses a method, apparatus, and system for verifying the causal relationship of airport network delay propagation. The method includes: constructing delay time series for each airport based on acquired historical flight data to generate an original dataset; inputting the original dataset into a pre-constructed deep separable dilated temporal convolutional neural network based on an attention mechanism to obtain an attention vector; based on the attention vector and the original dataset, sequentially performing attention interpretation, causal relationship verification, and causal effect delay mining to obtain the causal relationship of airport network delay propagation; and drawing a corresponding airport delay propagation causal network diagram based on the airport network delay propagation causal relationship to identify key airports in the airport network. This invention solves the problem of difficulty in capturing the causal relationship of delay propagation between airports in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of air traffic management technology, specifically relating to a method, apparatus, and system for testing the causal relationship of airport network delay propagation, and particularly to a method, apparatus, and system for testing the causal relationship of airport network delay propagation based on DS-ADTCNs. Background Technology

[0002] With the booming development of the economy, my country is at a critical stage of transitioning from a major civil aviation country to a leading civil aviation power. However, the management capacity of the existing air transport system is insufficient to meet the ever-increasing demand for air transport, resulting in widespread and profound flight delays. Flight delays have long been a major concern due to their implications for changes in passenger travel plans, adjustments to airport flight schedules, and increased workload for air traffic control departments, and are considered one of the key bottlenecks hindering the high-quality development of civil aviation.

[0003] Delays caused by a single flight can propagate to subsequent flights between airport pairs, resulting in widespread flight delays. These numerous airport pairs constitute a complex airport network. Identifying causal relationships within this complex network is a crucial issue in current air traffic management research. It helps clarify the causal relationships between airports and ensures the safe and efficient operation of the air traffic system. The continuous development of big data technology has made it possible to test causal relationships in airport network delay propagation through massive amounts of flight operation data. Currently, existing research on the causal relationships of delay propagation between airports primarily employs Granger causality tests, transfer entropy and its improved algorithms from information theory, and convergent cross-mapping algorithms. However, it has not yet utilized the representational capabilities of deep learning to mine causal relationships between airports from delay time series, and most studies lack relevant causal verification steps, thus compromising their credibility. To further analyze the mechanism of airport network delay propagation, it is necessary to fully consider the characteristics of flight delay propagation. We will use deep separable attention-based dilated temporal convolutional networks (DS-ADTCNs) based on attention mechanisms to test the causal relationship of airport network delay propagation, in order to reduce delays, identify important airports, and improve the efficiency of civil aviation operations. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method, apparatus, and system for testing the causal relationship of delay propagation in airport networks, which can solve the problem of difficulty in capturing the causal relationship of delay propagation between airports in the prior art.

[0005] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0006] In a first aspect, the present invention provides a method for testing the causal relationship of airport network delay propagation, comprising:

[0007] Based on the acquired historical flight data, delay time series for each airport are constructed to generate the original dataset;

[0008] The original dataset is input into a pre-constructed depthwise separable dilated temporal convolutional neural network based on an attention mechanism to obtain an attention vector;

[0009] Based on the attention vector and the original dataset, attention interpretation, causal relationship verification, and causal effect delay mining are performed sequentially to obtain the causal relationship of airport network delay propagation.

[0010] Based on the causal relationship of airport network delay propagation, draw the corresponding causal network diagram of airport delay propagation and identify the key airports in the airport network.

[0011] Optionally, the methods for constructing delay time series for each airport include:

[0012] For airport i, the average departure delay time within each hourly time interval is used. To construct the delay time series D i The average departure delay time of airport i within the time interval (t, t+1). Defined as:

[0013]

[0014] in, and num i (t) represents the total delay time and the total number of delayed flights for airport i during the time interval (t, t+1), respectively.

[0015] Obtain the delay time series D i ,

[0016] Optionally, the method for generating the original dataset χ includes:

[0017] For target airport j, a deep separable dilated temporal convolutional neural network based on an attention mechanism is used to predict its delay time series in time slice t. First, the delay time series χ of target airport j. j Each data point is shifted forward by one time step and padded with zeros on the left to complete the delay time series of target airport j. Become The delay time series for other airports remains unchanged.

[0018] Optionally, the method for constructing the attention-based depth-separable dilated temporal convolutional neural network includes:

[0019] The first step is to use a one-dimensional convolutional neural network with dilated kernels, where the kernels have a stride of f = c. l Skipping part of the input, the receptive field R of a one-dimensional convolutional kernel dilation time convolutional neural network D-TCN Defined as:

[0020]

[0021] Where c is the dilation coefficient, l is the layer containing the convolution kernel, and K is the number of convolution kernels;

[0022] The second step is to extend the one-dimensional convolutional kernel dilation time convolutional neural network into a one-dimensional depthwise separable architecture, specifically including:

[0023] exist Each network has N channels, and each channel corresponds to a delay time series of a target airport and delay time series of other airports.

[0024] During depthwise convolution, each channel is independently convolved with a convolutional kernel to generate an intermediate feature map with the same number of output channels. A 1*1 convolutional kernel is applied to each channel of the intermediate feature map to generate an output feature map, thus completing pointwise convolution and keeping the input airport delay time series separate.

[0025] The third step involves incorporating an attention mechanism into the network to obtain the impact of delay time series from other airports on the target airport's delays. Specifically, this includes:

[0026] a. Obtaining network The query matrix Q, key matrix K, and value matrix V are defined in the matrix, where J = 1, 2, ..., N, and N is the total number of airport nodes. The calculation formula is as follows:

[0027] Q = χ⊙W Q

[0028] K = χ⊙W K

[0029] V=χ⊙W V

[0030] Among them, W Q W K W V These are learnable parameters;

[0031] b. Computational Networks The attention scores are used to determine the delay impact between airports, and the calculation formula is as follows:

[0032]

[0033] Where, d k Key matrix K i The dimension;

[0034] c. Obtaining network access attention vector a j =[a 1j ,a 2j ,...,a jj ,...,a Nj ], a ij Indicates on the network Medium target delay time series χ j For the input other airport delay time series χ i≠j The degree of attention paid to the delay, i.e., the degree of impact of the delay, is calculated using the following formula:

[0035] a j =Soft max(A j V

[0036] Where Soft max(·) is the activation function, A j =[A 1j A 2j ,...,A jj ,...,A Nj ] indicates network All attention scores;

[0037] d. Except for the output layer, residual connections are used after each convolutional layer. The calculation formula is as follows:

[0038]

[0039] Where x is the input of this layer, This is the output of this layer.

[0040] Optionally, the attention interpretation method includes:

[0041] a. Obtaining attention score h ij The attention score set h is obtained by organizing the data. j Based on the obtained network attention vector a j Then, select HardSoftmax(·) for further processing. The calculation formula is as follows:

[0042]

[0043] Where, τ j It is a threshold and τ j ≥1;

[0044] c. Obtain the target airport delay time series χ j The set of potential causes P ∈χ j Specifically, it includes:

[0045] For the target airport delay time series χ j If h ij ∈h j If χ² > 0, it indicates that other airport delay time series χ² i≠j The target airport delay time series χ j The potential causes are placed into the potential cause set P. j .

[0046] Optionally, the causal relationship verification method includes:

[0047] The original dataset is input into a depthwise separable dilated temporal convolutional neural network based on an attention mechanism, and the network's loss is obtained. G ;

[0048] Only the underlying cause χ in the original dataset was changed i ∈P j The values ​​of are randomly arranged to generate an intervention dataset;

[0049] The intervention dataset is input into a deep, separable, dilated temporal convolutional neural network based on an attention mechanism to obtain the intervention loss. I ;

[0050] Obtain the target delay time series χ j The true cause set T of ∈χ j The specific method is as follows:

[0051] By comparing intervention loss I loss G The value of several pre-loss values. I Significantly greater than loss G χ indicates the potential cause. i ∈P j It is χ j The real reason is placed into the set of real reasons T. j The definition of "significantly greater than" is as follows:

[0052] Obtain the original dataset in a deep, separable, dilated temporal convolutional neural network based on an attention mechanism. The difference in loss between epoch1 and epochε after training εepochs. And epoch 1 and intervention loss I The difference

[0053] Comparison of Δloss I With Δloss G The value of ·s, if Δloss I ≤Δloss G ·s, then the intervention loss I Significantly greater than loss G , where s is the significance measure.

[0054] Optionally, the causal effect time delay mining method includes:

[0055] Obtain the kernel weights ω within each layer of a depthwise separable dilated temporal convolutional neural network based on an attention mechanism. i ;

[0056] Find the maximum value ω of the convolutional kernel weights in each layer. i-max The constructed path yields the correct causal effect time delay path;

[0057] The delay value of this causal effect is defined as the delay index. delay The specific calculation formula is as follows:

[0058] Index delay =Tt In

[0059] Where T refers to the network Medium delay time series χ i Length, t In This refers to the input layer χ in the causal effect delay path. In The corresponding time slice.

[0060] Optionally, the method for drawing the causal network diagram of airport delay propagation includes:

[0061] Using airports as nodes, if there is a causal relationship of delay propagation between airport pairs, an edge is formed, with the delay exponent as the weight of the edge, resulting in the airport delay propagation causal network G = (v, ε, w), where v = {v1, v2, ..., v...} N} is a set of N airport nodes. Let w be the set of edges, and let w be the set of edge weights.

[0062] Secondly, the present invention provides a device for testing the causal relationship of airport network delay propagation, comprising:

[0063] The raw dataset generation module is used to construct the delay time series of each airport based on the acquired historical flight data and generate the raw dataset.

[0064] The model prediction module is used to input the original dataset into a pre-constructed deep separable dilated temporal convolutional neural network based on an attention mechanism to obtain an attention vector;

[0065] The airport network delay propagation causality calculation module is used to perform attention interpretation, causal relationship verification, and causal effect delay mining sequentially based on the attention vector and the original dataset to obtain the causal relationship of airport network delay propagation.

[0066] The identification module is used to draw a corresponding airport delay propagation causal network diagram based on the causal relationship of the airport network delay propagation, and to identify key airports in the airport network.

[0067] Thirdly, the present invention provides a system for testing the causal relationship of airport network delay propagation, including a storage medium and a processor;

[0068] The storage medium is used to store instructions;

[0069] The processor is configured to operate according to the instructions to perform the method according to any one of the first aspects.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] This invention proposes a method, device, and system for verifying the causal relationship of airport network delay propagation. Specifically, it includes: constructing a deep separable dilated temporal convolutional neural network based on an attention mechanism; designing an attention interpretation step to obtain a set of potential causes of delays affecting airports; designing a causal relationship verification step to obtain a set of true causes of delays affecting airports; obtaining a delay index in the causal relationship of airport network delay propagation through causal effect delay mining; drawing a causal network diagram of airport delay propagation; accurately capturing the causal relationship of delay propagation in the airport network; and further identifying key airports in the airport network. This is beneficial for civil aviation management departments to formulate corresponding measures to reduce delays and improve flight punctuality. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0073] Figure 1 This is a flowchart illustrating a method for testing the causal relationship of airport network delay propagation according to an embodiment of the present invention.

[0074] Figure 2 This is a schematic diagram of a framework for causal relationship verification based on the DS-ADTCNs model according to an embodiment of the present invention;

[0075] Figure 3 This is a partial causal network diagram of airport delay propagation according to an embodiment of the present invention. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0077] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0078] Example 1

[0079] like Figure 1 As shown, this embodiment of the invention provides a method for testing the causal relationship of airport network delay propagation, including the following steps:

[0080] (1) Obtain historical flight data, perform data preprocessing, construct delay time series for each airport, and construct the original dataset;

[0081] (2) Construct a deep-separable attention-based dilated temporal convolutional network (DS-ADTCNs) model based on the attention mechanism, and input the data from the original dataset into the network model to obtain the attention vector;

[0082] (3) Based on the attention vector and the original dataset, perform three steps: attention interpretation, causal relationship verification, and causal effect delay mining to obtain the causal relationship of airport network delay propagation;

[0083] (4) Based on the obtained causal relationship of delay propagation between airports, draw the corresponding airport delay propagation causal network diagram and identify the key airports in the airport network.

[0084] In one specific embodiment of the present invention, the data preprocessing method includes:

[0085] Collect and use national flight operation data for a certain summer and autumn flight season; perform data cleaning, missing value and outlier handling; retain domestic passenger flights that have completed their transportation tasks from historical flight data; retain data from historical flight data with delay values ​​greater than 15 minutes and less than 180 minutes.

[0086] In one specific embodiment of the present invention, the construction of the original dataset includes:

[0087] Obtain the airport delay time series. For airport i, use the average departure delay time within each hourly time interval. To construct the delay time series D i The average departure delay time of airport i within the time interval (t, t+1). Defined as:

[0088]

[0089] in, and num i (t) represents the total delay time and the total number of delayed flights for airport i within the time interval (t, t+1), respectively. Therefore, the daily delay time series for airport i is a series of data points that change over time, such as...

[0090] For target airport j, the DS-ADTCNs model is used to predict its delay time series in time slice t. To prevent future data leaks from affecting prediction accuracy, it is necessary to process the target airport delay time series (i.e., the delay time series χ of target airport j). j Each data point of the target airport delay time series is shifted forward by one time step and padded with zeros on the left (this ensures that the delay time series of each airport has the same length and can be obtained from a causal relationship). The target airport delay time series is then derived from... Become The delay time series for other airports remains unchanged.

[0091] In one specific embodiment of the present invention, the method for constructing the DS-ADTCNs model includes:

[0092] The first step is to use a one-dimensional dilated temporal convolutional neural network (D-TCN), where the convolutional kernel has a stride of f = c. l Skipping part of the input, increasing the receptive field, the receptive field R of the one-dimensional convolutional kernel dilation time convolutional neural network D-TCN D-TCN Defined as:

[0093]

[0094] Where c is the dilation coefficient, l is the layer containing the convolution kernel, and K is the number of convolution kernels;

[0095] The second step is to extend the one-dimensional convolutional kernel dilated temporal convolutional neural network (D-TCN) into a one-dimensional depthwise separable architecture, specifically including:

[0096] exist Each network has N channels, and each channel corresponds to a delay time series of a target airport and delay time series of other airports. For the network... The time series of delays at the target airport for channel j is as follows: Other airport delay time series are

[0097] During depthwise convolution, each channel is independently convolved with a convolutional kernel to generate an intermediate feature map with the same number of output channels. A 1*1 convolutional kernel is applied to each channel of the intermediate feature map to generate an output feature map, thus completing pointwise convolution and keeping the input airport delay time series separate.

[0098] The third step involves incorporating an attention mechanism into the network to obtain the impact of delay time series from other airports on the target airport's delays. Specifically, this includes:

[0099] a. Obtaining network The query matrix Q, key matrix K, and value matrix V are defined in the matrix, where J = 1, 2, ..., N, and N is the total number of airport nodes. The calculation formula is as follows:

[0100] Q = χ⊙W Q

[0101] K = χ⊙W K

[0102] V=χ⊙W V

[0103] Among them, W Q W K W V These are learnable parameters;

[0104] b. Computational Networks The attention scores are used to determine the delay impact between airports, and the calculation formula is as follows:

[0105]

[0106] Where, d k Key matrix K i The dimension;

[0107] c. Obtaining network access attention vector a j =[a 1j ,a 2j ,...,a jj ,...,a Nj ], a ij Indicates on the network Medium target delay time series χ j For the input other airport delay time series χ i≠j The degree of attention (i.e., the degree of impact of the delay) is calculated using the following formula:

[0108] a j =Soft max(A j V

[0109] Where Soft max(·) is the activation function, A j =[A 1j A 2j ,...,A jj ,...,A Nj ] indicates network All attention scores;

[0110] d. Except for the output layer, residual connections are used after each convolutional layer. The calculation formula is as follows:

[0111]

[0112] Where x is the input of this layer, This is the output of this layer; the PReLU function is chosen as the activation function, and the calculation formula is as follows:

[0113]

[0114] Among them, y p Let a be the input of the nonlinear activation function in the p-th channel. p The slope is the negative slope of the activation function.

[0115] A schematic diagram of the constructed DS-ADTCNs model structure is shown below. Figure 2 As shown.

[0116] In one specific embodiment of the present invention, the attention interpretation method based on the DS-ADTCNs model includes:

[0117] a. Obtaining attention score h ij The attention score set h is obtained by organizing the data. j Based on the acquired network attention vector a j Then, select HardSoftmax(·) for further processing. The calculation formula is as follows:

[0118]

[0119] Where, τ j It is a threshold and τ j ≥1, the calculation method is as follows:

[0120] All attention scores A ij Sort the attention scores from highest to lowest and find two adjacent attention scores A. ij and A lm≠ij The maximum difference g between them max ∈G=[g0,...,g N-1 ], threshold τ j Then it equals the maximum difference g max The attention score on the left, and τ j The selection needs to satisfy 1) the maximum difference g max It cannot be between the highest and second-highest attention scores; 2) the maximum gap g max It needs to be in G = [g0,...,g N-1 The first half of the text;

[0121] b. Obtain the delay time series χ for each target airport j The set of potential causes P ∈χ j Specifically, it includes:

[0122] For the target airport delay time series χ j If h ij ∈h j If χ² > 0, it indicates that other airport delay time series χ² i≠j The target airport delay time series χ j The potential causes are placed into the potential cause set P. j ;

[0123] In one specific embodiment of the present invention, the causal relationship verification method based on the DS-ADTCNs model includes:

[0124] a. Input the original dataset into the DS-ADTCNs model and obtain the network loss. G

[0125] b. For each potential cause χi ∈P j Construct an intervention dataset, i.e., change only the underlying cause χ in the original dataset. i ∈P j The value is set to be randomly arranged;

[0126] c. Input the intervention dataset into the DS-ADTCNs model to obtain the intervention loss. I ;

[0127] d. Obtain the delay time series χ for each target airport j The true cause set T of ∈χ j The specific method is as follows:

[0128] By comparing intervention loss I loss G The value of several pre-loss values. I Significantly greater than loss G χ indicates the potential cause. i ∈P j It is χ j The real reason is placed into the set of real reasons T. j The definition of "significantly greater than" is as follows:

[0129] Obtain the original dataset in the DS-ADTCNs model network The difference in loss between epoch1 and epochε after training εepochs. And epoch 1 and intervention loss I The difference

[0130] Comparison of Δloss I With Δloss G The value of ·s, if Δloss I ≤Δloss G ·s, then the intervention loss I Significantly greater than loss G s is a measure of significance;

[0131] In one specific embodiment of the present invention, the causal effect time delay mining method based on the DS-ADTCNs model includes:

[0132] Obtain the kernel weights ω in each layer of the DS-ADTCNs model. i Find the maximum weight ω of each convolutional kernel. i-max The constructed path, i.e., the correct causal effect delay path, is then defined as the delay index. delayThe specific calculation formula is as follows:

[0133] Index delay =Tt In

[0134] Where T refers to the network Delay time series of other airports in China χ i Time series length, t In This refers to the input layer χ in the causal effect delay path. In The corresponding time slice.

[0135] In one specific embodiment of the present invention, the method for drawing the airport delay propagation causal network diagram includes:

[0136] Using airports as nodes, if there is a causal relationship of delay propagation between airport pairs, an edge is formed, with the delay exponent as the weight of the edge, resulting in the airport delay propagation causal network G = (v, ε, w), where v = {v1, v2, ..., v...} N} is a set of N airport nodes. Let w be the set of edges, and w be the set of edge weights, such as... Figure 3 As shown.

[0137] Example 2

[0138] Based on the same inventive concept as Embodiment 1, this embodiment of the invention provides a device for testing the causal relationship of airport network delay propagation, comprising:

[0139] The raw dataset generation module is used to construct the delay time series of each airport based on the acquired historical flight data and generate the raw dataset.

[0140] The model prediction module is used to input the original dataset into a pre-constructed depthwise separable dilated temporal convolutional neural network based on an attention mechanism, and to input the attention vector;

[0141] The airport network delay propagation causality calculation module is used to perform attention interpretation, causal relationship verification, and causal effect delay mining sequentially based on the attention vector and the original dataset to obtain the causal relationship of airport network delay propagation.

[0142] The identification module is used to draw a corresponding airport delay propagation causal network diagram based on the causal relationship of the airport network delay propagation, and to identify key airports in the airport network.

[0143] The rest are the same as in Example 1.

[0144] Example 3

[0145] Based on the same inventive concept as in Embodiment 1, this embodiment of the invention provides a system for testing the causal relationship of airport network delay propagation, including a storage medium and a processor;

[0146] The storage medium is used to store instructions;

[0147] The processor is configured to operate according to the instructions to execute the method according to any one of Embodiments 1.

[0148] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0153] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for testing the causal relationship of airport network delay propagation, characterized in that, include: Based on the acquired historical flight data, delay time series for each airport are constructed to generate the original dataset; The original dataset is input into a pre-constructed depthwise separable dilated temporal convolutional neural network based on an attention mechanism to obtain an attention vector; Based on the attention vector and the original dataset, attention interpretation, causal relationship verification, and causal effect delay mining are performed sequentially to obtain the causal relationship of airport network delay propagation. Based on the causal relationship of airport network delay propagation, draw the corresponding airport delay propagation causal network diagram and identify the key airports in the airport network; The method for constructing the attention-based depthwise separable dilated temporal convolutional neural network includes: The first step is to use a one-dimensional convolutional neural network with dilated kernels, where the kernels are dilated with a stride of [missing value]. Skipping part of the input, one-dimensional convolutional kernel dilation time convolutional neural network receptive field Defined as: ， in, The expansion coefficient, The layer containing the convolution kernel; The number of convolution kernels; The second step is to extend the one-dimensional convolutional kernel dilation time convolutional neural network into a one-dimensional depthwise separable architecture, specifically including: exist Each network is set up Each channel corresponds to a delay time series of a target airport and delay time series of other airports; During depthwise convolution, each channel is independently convolved with a convolutional kernel to generate an intermediate feature map with the same number of output channels. A 1*1 convolutional kernel is applied to each channel of the intermediate feature map to generate an output feature map, thus completing pointwise convolution and keeping the input airport delay time series separate. The third step involves incorporating an attention mechanism into the network to obtain the impact of delay time series from other airports on the target airport's delays. Specifically, this includes: a. Obtaining network Query matrix in Key matrix Value matrix , , The total number of airport nodes is calculated using the following formula: ， ， ， in, These are learnable parameters; b. Computational Networks The attention scores are used to determine the delay impact between airports, and the calculation formula is as follows: ， in, Key matrix The dimension; c. Obtaining network access attention vector , Indicates on the network Medium target delay time series Other airport delay time series input The degree of attention paid to the delay, i.e., the degree of impact of the delay, is calculated using the following formula: ， in, For activation function, Represents network All attention scores; d. Except for the output layer, residual connections are used after each convolutional layer. The calculation formula is as follows: ， in, This is the input to this layer. This is the output of this layer; The attention interpretation method includes: a. Obtaining attention scores The attention score set was obtained by organizing the data. Based on the obtained network attention vector ,choose Further processing is required, and the calculation formula is as follows: ， in, It is a threshold and ; b. Obtain the delay time series of the target airport potential cause set Specifically, it includes: For the target airport delay time series ,like This indicates the delay time series of other airports. It is the time series of delays at the target airport. Potential causes, place them into the potential cause set. ; The causal relationship verification method includes: The original dataset is input into a deep, separable, dilated temporal convolutional neural network based on an attention mechanism, and the network's loss is obtained. ; Only change the underlying cause in the original dataset The values ​​of are randomly arranged to generate an intervention dataset; The intervention dataset is input into a deep, separable, dilated temporal convolutional neural network based on an attention mechanism to obtain the intervention loss. ; Obtain the target delay time series The real reasons set The specific method is as follows: By comparing intervention losses With loss The value of several pre-losses Significantly greater than the loss This indicates a potential cause. yes The real reason, put it into the set of real reasons. The definition of "significantly greater than" is as follows: Obtain the original dataset in a deep, separable, dilated temporal convolutional neural network based on an attention mechanism. train back, and loss difference as well as and intervention losses The difference ; Compare and The value, if Intervention loss Significantly greater than the loss , For significance measurement; The causal effect time delay mining method includes: Obtain the kernel weights within each layer of a depthwise separable dilated temporal convolutional neural network based on an attention mechanism. ; Find the maximum value of the kernel weights in each layer of convolution. The constructed path yields the correct causal effect time delay path; The delay value of the causal effect is defined as the delay index. The specific calculation formula is as follows: ， in, It refers to the network Delay time series of other airports in China Length, This refers to the input layer in the causal effect delay path. The corresponding time slice.

2. The method for testing the causal relationship of airport network delay propagation according to claim 1, characterized in that: The methods for constructing delay time series for each airport include: For airports The average departure delay time within each hourly time interval is used. To construct delay time series ,Airport exist Average departure delay time within the time interval Defined as: ， in, and Airports exist Total delay time and total number of delayed flights within the time interval; Obtain the delay time series , .

3. The method for testing the causal relationship of airport network delay propagation according to claim 1, characterized in that: The original dataset The generation methods include: Target airport In this regard, a deep, separable, dilated temporal convolutional neural network based on an attention mechanism is used to predict its time slice. Delay time series Before, the target airport Delay time series Each data point is shifted forward by one time step and filled with zeros on the left to complete the target airport. The delay time series is composed of Become Other airports The delay time series remains unchanged, for .

4. The method for testing the causal relationship of airport network delay propagation according to claim 1, characterized in that, The method for drawing the causal network diagram of airport delay propagation includes: Using airports as nodes, if there is a causal relationship of delay propagation between airports, an edge is formed, with the delay index as the weight of the edge, thus obtaining the airport delay propagation causal network. ,in yes A set of airport nodes, It is a set of edges. It is the set of edge weights.

5. A device for testing the causal relationship of airport network delay propagation, characterized in that, include: The raw dataset generation module is used to construct the delay time series of each airport based on the acquired historical flight data and generate the raw dataset. The model prediction module is used to input the original dataset into a pre-constructed deep separable dilated temporal convolutional neural network based on an attention mechanism to obtain an attention vector; The airport network delay propagation causality calculation module is used to perform attention interpretation, causal relationship verification, and causal effect delay mining sequentially based on the attention vector and the original dataset to obtain the causal relationship of airport network delay propagation. The identification module is used to draw a corresponding airport delay propagation causal network diagram based on the causal relationship of the airport network delay propagation, and to identify key airports in the airport network. The method for constructing the attention-based depthwise separable dilated temporal convolutional neural network includes: The first step is to use a one-dimensional convolutional neural network with dilated kernels, where the kernels are dilated with a stride of [missing value]. Skipping part of the input, one-dimensional convolutional kernel dilation time convolutional neural network receptive field Defined as: in, The expansion coefficient, The layer containing the convolution kernel; The number of convolution kernels; The second step is to extend the one-dimensional convolutional kernel dilation time convolutional neural network into a one-dimensional depthwise separable architecture, specifically including: exist Each network is set up Each channel corresponds to a delay time series of a target airport and delay time series of other airports; During depthwise convolution, each channel is independently convolved with a convolutional kernel to generate an intermediate feature map with the same number of output channels. A 1*1 convolutional kernel is applied to each channel of the intermediate feature map to generate an output feature map, thus completing pointwise convolution and keeping the input airport delay time series separate. The third step involves incorporating an attention mechanism into the network to obtain the impact of delay time series from other airports on the target airport's delays. Specifically, this includes: a. Obtaining network Query matrix in Key matrix Value matrix , , The total number of airport nodes is calculated using the following formula: ， ， ， in, These are learnable parameters; b. Computational Networks The attention scores are used to determine the delay impact between airports, and the calculation formula is as follows: ， in, Key matrix The dimension; c. Obtaining network access attention vector , Indicates on the network Medium target delay time series Other airport delay time series input The degree of attention paid to the delay, i.e., the degree of impact of the delay, is calculated using the following formula: ， in, For activation function, Represents network All attention scores; d. Except for the output layer, residual connections are used after each convolutional layer. The calculation formula is as follows: ， in, This is the input to this layer. This is the output of this layer; The attention interpretation method includes: a. Obtaining attention scores The attention score set was obtained by organizing the data. Based on the obtained network attention vector ,choose Further processing is required, and the calculation formula is as follows: ， in, It is a threshold and ; c. Obtain the delay time series of the target airport potential cause set Specifically, it includes: For the target airport delay time series ,like This indicates the delay time series of other airports. It is the time series of delays at the target airport. Potential causes, place them into the potential cause set. ; The causal relationship verification method includes: The original dataset is input into a deep, separable, dilated temporal convolutional neural network based on an attention mechanism, and the network's loss is obtained. ; Only change the underlying cause in the original dataset The values ​​of are randomly arranged to generate an intervention dataset; The intervention dataset is input into a deep, separable, dilated temporal convolutional neural network based on an attention mechanism to obtain the intervention loss. ; Obtain the target delay time series The real reasons set The specific method is as follows: By comparing intervention losses With loss The value of several pre-losses Significantly greater than the loss This indicates a potential cause. yes The real reason, put it into the set of real reasons. The definition of "significantly greater than" is as follows: Obtain the original dataset in a deep, separable, dilated temporal convolutional neural network based on an attention mechanism. train back, and loss difference as well as and intervention losses The difference ; Compare and The value, if Intervention loss Significantly greater than the loss , For significance measurement; The causal effect time delay mining method includes: Obtain the kernel weights within each layer of a depthwise separable dilated temporal convolutional neural network based on an attention mechanism. ; Find the maximum value of the kernel weights in each layer of convolution. The constructed path yields the correct causal effect time delay path; The delay value of the causal effect is defined as the delay index. The specific calculation formula is as follows: ， in, It refers to the network Delay time series of other airports in China Length, This refers to the input layer in the causal effect delay path. The corresponding time slice.

6. A system for testing the causal relationship of airport network delay propagation, characterized in that, Including storage media and processor; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the method according to any one of claims 1-4.

Images