A dynamic spatio-temporal graph neural network model for aviation delay prediction and application
By using a dynamic spatiotemporal graph neural network model, which combines graph convolutional networks and recurrent neural networks, the problem of airport heterogeneity in flight delay prediction is solved, and accurate prediction of the average airport delay in future time steps is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2023-11-07
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies fail to effectively account for airport heterogeneity in flight delay prediction, making it difficult to construct accurate ripple effect prediction models.
A dynamic spatiotemporal graph neural network model is adopted, which combines graph convolutional networks and recurrent neural networks. Using airport distance and flight traffic as the graph structure, an adaptive vector is designed to dynamically weight and fuse heterogeneous airports, so as to achieve accurate prediction of the average airport delay in future time steps.
It enables accurate prediction of flight delays, solves the problem of difficult integration of multi-dimensional correlations in airport networks, and improves the quantitative prediction capability of flight delays.
Smart Images
Figure CN117521719B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of civil aviation delay prediction technology, specifically to a dynamic spatiotemporal graph neural network model for aviation delay prediction and its application. Background Technology
[0002] In recent years, civil aviation has become an important component of building a modern comprehensive transportation system due to its convenient and efficient transportation capabilities. However, with the increasing demand for transportation and the significant increase in flight traffic, the contradiction between the rapid growth of air transport within limited airspace and the shortage of key resources and weak support capabilities has become increasingly prominent. Coupled with frequent unforeseen events such as severe weather and air traffic control disruptions, this directly leads to a decline in flight punctuality and frequent flight delays. These delays then spread within the network, triggering a ripple effect on the air traffic network and causing serious economic losses and social impacts. Currently, Chinese patents CN111738425B, CN111523641A, and CN115018073A have conducted research and exploration on predicting civil aviation flight delay information.
[0003] Flight delay prediction utilizes past airport delay data to learn prior knowledge, extract sequential trend information, and eliminate the random effects of disturbances to predict future airport delay trends. However, the complexity of air transportation systems, the diversity of prediction models, and the large amount of data in aviation systems make it difficult to construct accurate ripple effect prediction models. Domestic and international researchers categorize ripple effect prediction methods into four types: those based on statistical theory, those based on machine learning, those based on deep learning, and those based on neural networks. In 2017, Guo Yechenfeng et al. established a multiple linear regression model, achieving effective prediction of airport delays under adverse weather conditions. In 2009, Xu Tao et al., considering the possibility of insufficient airport capacity leading to flight delays, modeled continuously updated flight data using an improved SVM algorithm, achieving an accuracy rate of 80%. In 2016, Elaraby NM et al. predicted the average daily flight ripple delay based on the LSTM model. Experiments showed that using a prediction model based on recurrent neural networks could effectively obtain the temporal correlation of daily airport delays, thereby predicting the daily ripple delay. In 2022, Qu Jingyi et al. proposed a flight ripple delay prediction method based on convolutional neural networks. By extracting spatial feature information through convolutional networks, they were able to achieve relatively accurate ripple delay prediction. In 2021, Song Chunyue et al. proposed an aviation network ripple effect prediction method that combines attention mechanisms within a sequence-to-sequence framework. Also in 2021, Cai et al. proposed a dynamic structure based on the STGCN model to perform dynamic spatiotemporal feature learning, achieving effective prediction of flight delays.
[0004] In recent years, researchers have studied the problem of flight delay prediction based on graph neural network models. They use airports as nodes, airport delays as node features, and airport distance and flight traffic as edge weights to characterize the airport network structure. However, current research has not considered the issue of airport heterogeneity, neglecting the impact of heterogeneous airports on airport association extraction. Summary of the Invention
[0005] In view of the above problems, this invention provides a dynamic spatiotemporal graph neural network model and its application for predicting flight delays. Based on an attention mechanism-based encoder-decoder architecture, it couples graph convolutional networks with recurrent neural networks, uses airport distance and flight traffic as the graph structure, and designs adaptive vectors to dynamically weight and fuse heterogeneous airports, thereby achieving accurate prediction of the average airport delay in future time steps.
[0006] This invention provides a dynamic spatiotemporal graph neural network model for predicting flight delays and its application, including: Input layer, graph convolution-heterogeneous airport weighted fusion layer, graph convolution-recurrent neural network coupling layer, and output layer; The input layer is used to receive delay information of each airport in multiple historical time periods, flight information of each airport in each historical time period, and distance information between each airport. The graph convolution-heterogeneous airport weighted fusion layer is used to perform graph convolution calculation and dynamic weighted fusion of airport delay information for each historical time period to obtain the total hidden layer vector for each historical time period. The graph convolutional-recurrent neural network coupling layer couples the total hidden layer vectors of each historical time period with a graph convolutional-recurrent neural network to establish an encoder-decoder architecture and obtain the decoder output hidden layer vectors. The output layer obtains the prediction of future flight delay times by performing a nonlinear transformation on the hidden layer vector output by the decoder.
[0007] Preferably, the input layer further includes an airport delay map vector acquisition layer, which inputs the delay information of each airport in the multiple historical time periods into the airport delay map vector acquisition layer and outputs multiple historical time period airport delay map vectors. The static airport distance map acquisition layer takes the distance information between the airports as input to the static airport distance map acquisition layer to acquire the static airport distance map. The dynamic flight flow map acquisition layer takes the flight information of the airport in each historical time period as input to obtain the dynamic flight flow map for each historical time period. Furthermore, the airport delay map vectors for the multiple historical time periods, the dynamic flight flow maps for each historical time period, and the static airport distance maps are obtained. Specific steps include: Step a) Extract historical flight data and latitude / longitude coordinates of each airport in the aviation network; calculate the average flight delay data and the number of flights to and from each airport for each historical time period in the historical flight data to obtain the historical time period dataset. H data , H data = { h m}, m= 1,2,... M , M This represents the total number of segments in the historical time period. Furthermore, the historical time period described in step a is divided as follows: a day is divided into 24 hours. M Each historical period is divided into two parts, with each part consisting of half an hour.
[0008] Step b, use any of the steps in step a. m The average delay data of all airports over a historical time period is used to construct the first... m Airport delay vector map for a historical time period r m ; Use step a m Using the number of flights to and from each airport during a historical time period as the flight flow value, the first... m Dynamic flight flow chart for historical time periods A static airport distance map is obtained based on the latitude and longitude coordinates of each airport in the aviation network in step a. A Distance ; Furthermore, obtain the first step described in step b. m Airport delay vector map for a historical time period r m The specific steps include: Select the m The first historical period, with the first... m The airports in the first historical time period are used as node vectors in the graph, with the first... m The average airport delay time over the given historical time period is used as the delay feature of the node vector to obtain the th... m The update node vector for the nth historical time period will be... m Concatenate all update node vectors from the historical time period to construct the _th _ ... m Airport delay vector map for a historical time period r m ,like Figure 2 As shown; The first m Airport delay vector map for a historical time period r mIt has undergone standardization processing; the standardization processing is: obtaining M Airport delay graph vector for a total historical time period r m Given a dataset, the maximum and minimum values of the dataset are standardized, with the standardized values set to range [0,1]. The expression is: ； in, r m ′ For the first m Updated airport delay map vectors for each historical time period; For the first m The minimum value in the historical time period airport delay graph vector dataset. For the first m The maximum value in the airport delay graph vector dataset for each historical time period.
[0009] Furthermore, step b, which involves calculating the static airport distance based on the airport's latitude and longitude coordinates, is expressed as follows: ; Where r is the Earth's radius. D ij For the airport i With the airport j distance; x i 、y i Airports i Longitude and latitude x j 、 y j Airports j The longitude and latitude are represented by the following values: east longitude is positive, west longitude is negative, north latitude is positive, and south latitude is negative.
[0010] Furthermore, the static airport distance map matrix mentioned in step b... A Distance The edge weights are the airport distances, and it is a symmetric undirected graph, such as... Figure 3 As shown.
[0011] Furthermore, obtain the dynamic flight flow map structure described in step b. The specific steps include: The first m The number of flights between each pair of airports within a historical time period is taken as the flight traffic, with the airport as the node. m Using flight traffic over a historical time period as the edge weights, a dynamic flight traffic graph matrix is generated. , is a directed graph, such as Figure 4 As shown.
[0012] Preferably, the specific steps for obtaining the total hidden layer vector for each historical time period include: Step c: Obtain the static airport distance map described in step b. A Distance and the m Dynamic flight flow chart for historical time periods As input information for the graph structure, for the first m Airport delay vector map for a historical time period r m Perform graph convolution calculations to obtain the first... m Airport distance hidden layer vector for each historical time period Airport traffic hidden layer vector ; Furthermore, obtain the first step described in step c. m Airport distance to hidden layer vector during historical time periods Airport traffic hidden layer vector The specific steps are as follows: Step c1, the first step described in step b m Airport delay vector map for a historical time period r m Static airport distance map A Distance and the m Dynamic flight flow chart for historical time periods As input information for the graph structure, for the first m Airport delay vector map for a historical time period r m Perform graph convolution calculations and obtain the first... m The state vector of the convolutional hidden layer of a static airport map for a historical time period. Convolutional hidden layer state vectors of airport traffic graph ; Step c2, based on the first step described in step c1 m The state vector of the convolutional hidden layer of a static airport map for a historical time period. Convolutional hidden layer state vectors of airport traffic graph , for the m Delay graph vector for each historical time period r m Perform graph convolution calculation to obtain the first m Airport distance to hidden layer vector during historical time periods Airport traffic hidden layer vector .
[0013] Furthermore, the static airport graph convolutional hidden layer state vector described in step c1 Hmmm The expression is: ; The expression for the state vector of the convolutional hidden layer of the airport traffic map is: ; in, For the first m The convolutional hidden layer state vectors of static airport graphs for each historical time period For activation functions; D Static airport distance map matrix à D and dynamic flight flow map matrix à F The degree matrix, à D This is a static airport distance map matrix; H m For the first m The state vector of the convolutional hidden layer of a historical time period graph; W For parameter weights; A D This refers to the static airport distance; I N Let N be the identity matrix of the Nth airport dimension; For the first m Hidden state vectors of airport traffic over a historical time period; à F A dynamic flight flow map matrix; A F For dynamic flight flow m =1,2,..., M , M for M There are 1, 2, ..., n historical time periods; N = 1, 2, ..., n, where n is the number of airports.
[0014] Furthermore, the activation function for sigmoid function, tanh function or Relu function.
[0015] Furthermore, the first step described in step c2 m Airport distance to hidden layer vector during historical time periods Airport traffic hidden layer vector The expression is: ; in, For the first m Airport distance hidden layer vector for each historical time period, For activation functions; i D*G The convolution process for the airport static distance map; r m For the first m Airport delay graph vector for each historical time period; For the first m The airport distance hidden layer vector in the previous historical time period; For the first m Hidden vectors of airport traffic over a historical time period; i F*G The convolution process for the airport dynamic flight flow map; For the first m Hidden layer vector of airport traffic in the previous historical time period; m ≥1, m =1,2,..., M , M for M A total historical period.
[0016] Step d: Design the static adaptive vector W for heterogeneous airports. distance Dynamic adaptive vectors of heterogeneous airports The airport distance hidden layer vector described in step c is respectively... and the airport traffic hidden layer vector Perform dynamic weighted fusion of heterogeneous airports to obtain the first m Total hidden layer vector for each historical time period ; Furthermore, the expression for the dynamic weighted fusion of heterogeneous airports in step d is: ; in, For the first m Total hidden layer vector for each historical time period; W distance For heterogeneous airports, the static adaptive vector is denoted as . For the first m Airport distance hidden layer vector for each historical time period; For heterogeneous airports, the dynamic adaptive vector is denoted as . No. m Hidden vectors of airport traffic over a historical time period; m =1,2,..., M , M for M A total historical period.
[0017] Preferably, the specific steps for obtaining the decoder output hidden layer vector include: Step e, splicing step b as described in the first step mAirport delay vector map for a historical time period r m Compared with step d, the first m -1 total hidden layer vector for a historical time period As temporal correlation data for a recurrent neural network, the temporal correlation data is input into the matrix multiplication unit of a graph convolution, based on the static airport distance map described in step b and the first... m The graph convolution calculation is performed on the dynamic flight flow map of each historical time period, and then the gating vector is obtained by the first neural network function. The splicing step b described in the first step m Airport delay vector map for a historical time period r m Compared with step d, the first m -1 total hidden layer vector for a historical time period And perform the first neural network function calculation to obtain the first... m -1 historical time period update hidden layer vector Based on the gate vector, the first m -1 total hidden layer vector for a historical time period and updating hidden layer vectors Perform weighted extraction to obtain the first... m Hidden layer vectors for each historical time period The hidden layer vectors for each historical time period Iteratively obtain the final hidden layer vector for the total historical time period as the encoder output vector. ; Further, obtain the encoder output vector described in step e. The specific steps include: Step e1, splicing the first m Airport delay vector map for a historical time period r m With the m- Hidden layer vector for one historical time period Based on step b, the first m The dynamic flight flow map and static airport distance map for each historical time period are subjected to graph convolution, and nonlinear features are added after passing through the first activation function to obtain the th... m The historical time period gating vector: the m Total hidden feature gating vector for each historical time period , No. m Correct feature gating vectors for each historical time period , No. m Historical feature gating vectors for each historical time period The expression is as follows: ; in, For the first m The total hidden feature gating vector for each historical time period, θ O*G For the total hidden feature map convolution process, r m For the first m Airport delay graph vectors for a historical time period; For the first m The feature gating vector is modified for each historical time period, θ F*G To correct the feature map convolution process, For the first m Historical feature gating vectors for each historical time period sigmoid The function is the first activation function; θ I*G For the historical feature map convolution process, m≥1; Step e2, splicing the first m Airport delay map vector for each historical time period r m With the m -1 hidden layer vector for a historical time period Based on step b, the first m The dynamic flight flow map and static airport distance map for each historical time period are convolved, and then subjected to a nonlinear transformation using a second activation function to generate the first... m -1 historical time period update hidden layer vector The expression is: ; in, For the first m Update the hidden layer vector for each historical time period. r m For the first m Airport delay graph vectors for a historical time period; tanh The function is the second activation function. For the first m -1 hidden layer vector for a historical time period, m ≥1, θ U*G For graph convolution functions; Step e3, using the gating vector described in step e1 to... m -1 historical time period hidden layer vector Compared with step e2, the first m -1 historical time period update hidden layer vector Perform weighted extraction to obtain the first... m Hidden layer vectors for each historical time period The expression is: ; in, For the firstm Hidden layer vectors for each historical time period tanh For the second activation function, For the first m -1 hidden layer vector for a historical time period, For the first m -1 historical time interval update hidden layer vector; Step e4: Iterate through each historical time period according to the expression described in step e3 to generate the hidden layer state vector, and obtain... M The hidden layer vector for the total historical time period is used as the encoder output vector. .
[0018] Step f, Obtain M The first m Hidden layer vectors for each historical time period ,Will M The first m Hidden layer vectors for each historical time period Concatenate the data to obtain the total historical time period concatenated hidden layer vector. The encoder output vector described in step e Concatenate the hidden layer vector with the total historical time period mentioned above. Perform function enhancements to obtain the encoder output vector updated for each total historical time period. ; Update the encoder output vector with the total historical time period mentioned above. Copy and concatenate to obtain the transpose of the encoder output mapping vector. The hidden layer vector is concatenated with the total historical time period. Similarity calculations are performed to obtain the attention weights for the total historical time period; based on these attention weights, the decoder weighted hidden layer vector is obtained. ; Furthermore, step f describes obtaining the decoder weighted hidden layer vector. The specific steps include: Step f1: Obtain respectively M The first m Hidden layer vectors for each historical time period , will the M The first m Hidden layer vectors for each historical time period Concatenate the data to obtain the total historical time period concatenated hidden layer vector. ; Step f2: Concatenate the encoder output vector described in step e. The hidden layer vector of the total historical time period described in step f1 Based on step b, the first mThe encoder output mapping vector is obtained by performing graph convolution on the dynamic flight flow map and the static airport distance map for each historical time period, followed by a third activation function calculation. The total historical time period is concatenated into the hidden layer mapping vector. ; The encoder outputs a mapping vector The total historical time period is concatenated into the hidden layer mapping vector. The expression is: ; in, fullconnect It is a fully connected layer; Relu It is the third activation function; Step f3: Map the encoder output vector described in step f2. The transpose of the encoder output mapping vector is obtained by copying and concatenating the vectors based on their dimensions at each historical time period. The transpose of the encoder output mapping vector Concatenate the hidden layer mapping vector Calculate the similarity and assign attention weights to different historical time periods based on the similarity to obtain the attention weights for each historical time period; The expression for the similarity is: ; Where α is the similarity; softmax It is a normalized exponential function; Step f4: Based on the attention weights described in step f3, map the hidden layer vectors described in step f2. Weighting is performed to generate the weighted hidden layer vector of the decoder. The expression is: ; Step g: Convert the encoder output vector Weighted hidden layer vectors of decoder As input, the first m Using a dynamic flight flow map and a static airport distance map for a historical time period as a graph structure, graph convolution calculation is performed to obtain the graph convolution calculation result. Based on the graph convolution calculation result and the static adaptive vector W of the heterogeneous airport in step d, distance Dynamic adaptive vectors of heterogeneous airports Obtain the decoder output hidden layer vector ; Furthermore, in step g, the decoder outputs the hidden layer vector. The expression is: ; in, Output the distance vector; This is the flow output vector.
[0019] Preferably, the specific steps for obtaining the predicted airport network delay time for the future historical time period include: The decoder outputs the hidden layer vector as described in step g. Converted into a one-dimensional vector, this one-dimensional vector is used as a predicted value for airport network delay time in a future historical time period. r output .
[0020] Furthermore, obtain the predicted airport network delay time for the next future time period. r output ,include: The fully connected layer is designed to output the hidden layer vector from the decoder described in step g. Convert to a one-dimensional vector The single-dimensional vector The predicted airport network delay time for a future time period is obtained after calculation using the first activation function. ， The expression is: ; in, sigmoid The first activation function is a function; fullconnect It is a fully connected layer. r output This is the one-dimensional output vector of the decoder.
[0021] Furthermore, the decoder's one-dimensional output vector r output As a predicted value for airport network delays in a future historical time period, the predicted airport delay values for each historical time period are characterized as quantitative values of the ripple effect, thereby realizing the quantitative prediction of the ripple effect of the aviation airport network.
[0022] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention designs an adaptive vector matrix to perform weighted fusion of spatial correlation and traffic service correlation of heterogeneous airports, so as to comprehensively extract multidimensional correlation information of heterogeneous airports and solve the problem of difficult fusion of multidimensional correlation of airport network. (2) The present invention uses a coupled graph convolutional network and recurrent neural network model to extract the spatiotemporal features of delay vectors under the deencoder architecture. It also conducts research on the prediction method of aviation network spillover effect considering the heterogeneous association fusion of airports, and realizes the quantitative prediction of flight delays. (3) This invention considers the importance of airport association in each historical time period, adds an attention mechanism to the decoder part, weights the importance of the hidden layer in each historical time period in the encoder, and uses the weighted hidden layer vector as the input of the decoder to finally achieve the prediction of future time step delay. Attached Figure Description
[0023] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.
[0024] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a schematic diagram of the airport network delay graph vector for a single historical time period according to the present invention; Figure 3 This is a schematic diagram of the multi-timescale causal relationships in the aviation network of the present invention; Figure 4 This is a schematic diagram of the airport network flight traffic relationship of the present invention. Detailed Implementation
[0025] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.
[0026] A specific embodiment of the present invention, such as Figure 1 This invention discloses a dynamic spatiotemporal graph neural network model for predicting flight delays and its application. To illustrate the effectiveness of the proposed method, a specific embodiment is provided below for detailed explanation of the above technical solution. The specific implementation steps are as follows: Input layer, graph convolution-heterogeneous airport weighted fusion layer, graph convolution-recurrent neural network coupling layer, and output layer; The input layer is used to receive delay information of each airport in multiple historical time periods, flight information of each airport in each historical time period, and distance information between each airport. The graph convolution-heterogeneous airport weighted fusion layer is used to perform graph convolution calculation and dynamic weighted fusion of airport delay information for each historical time period to obtain the total hidden layer vector for each historical time period. The graph convolutional-recurrent neural network coupling layer couples the total hidden layer vectors of each historical time period with a graph convolutional-recurrent neural network to establish an encoder-decoder architecture and obtain the decoder output hidden layer vectors. The output layer obtains the prediction of future flight delay times by performing a nonlinear transformation on the hidden layer vector output by the decoder.
[0027] Preferably, the input layer further includes an airport delay map vector acquisition layer, which inputs the delay information of each airport in the multiple historical time periods into the airport delay map vector acquisition layer and outputs multiple historical time period airport delay map vectors. The static airport distance map acquisition layer takes the distance information between the airports as input to the static airport distance map acquisition layer to acquire the static airport distance map. The dynamic flight flow map acquisition layer takes flight information of airports in each historical time period as input to obtain dynamic flight flow maps for each historical time period. Further, it acquires airport delay map vectors for the multiple historical time periods, dynamic flight flow maps for each historical time period, and static airport distance maps. Specific steps include: Step a) Extract historical flight data and latitude / longitude coordinates of each airport in the aviation network; calculate the average flight delay data and the number of flights to and from each airport for each historical time period in the historical flight data to obtain the historical time period dataset. H data , H data = { h m}, m= 1,2,... M , M This represents the total number of segments in the historical time period. Furthermore, the historical time period described in step a is divided as follows: a day is divided into 24 hours. M Each historical period is divided into two parts, with each part consisting of half an hour.
[0028] Step b, use any of the steps in step a. m The average delay data of all airports over a historical time period is used to construct the first... m Airport delay vector map for a historical time period r m ; Use step a m Using the number of flights to and from each airport during a historical time period as the flight flow value, the first... m Dynamic flight flow chart for historical time periods A static airport distance map is obtained based on the latitude and longitude coordinates of each airport in the aviation network in step a. A Distance ; Furthermore, obtain the first step described in step b. m Airport delay vector map for a historical time period r m The specific steps include: Select the m The first historical period, with the first... mThe airports in the first historical time period are used as node vectors in the graph, with the first... m The average airport delay time over the given historical time period is used as the delay feature of the node vector to obtain the th... m The update node vector for the nth historical time period will be... m Concatenate all update node vectors from the historical time period to construct the _th _ ... m Airport delay vector map for a historical time period r m ,like Figure 2 As shown; The first m Airport delay vector map for a historical time period r m It has undergone standardization processing; the standardization processing is: obtaining M Airport delay graph vector for a total historical time period r m Given a dataset, the maximum and minimum values of the dataset are standardized, with the standardized values set to range [0,1]. The expression is: ； in, r m For the first m Airport delay graph vectors for historical time periods r m ′ For the first m Updated airport delay map vectors for each historical time period; For the first m The minimum value in the historical time period airport delay graph vector dataset. For the first m The maximum value in the airport delay graph vector dataset for each historical time period.
[0029] Furthermore, step b, which involves calculating the static airport distance based on the airport's latitude and longitude coordinates, is expressed as follows: ; Where r is the Earth's radius. D ij For the airport i With the airport j distance; x i 、y i Airports i Longitude and latitude x j 、 y j Airportsj The longitude and latitude are represented by the following values: east longitude is positive, west longitude is negative, north latitude is positive, and south latitude is negative.
[0030] Furthermore, the static airport distance map matrix mentioned in step b... A Distance The edge weights are the airport distances, and it is a symmetric undirected graph, such as... Figure 3 As shown.
[0031] Furthermore, obtain the dynamic flight flow map structure described in step b. The specific steps include: The first m The number of flights between each pair of airports within a historical time period is taken as the flight traffic, with the airport as the node. m Using flight traffic over a historical time period as the edge weights, a dynamic flight traffic graph matrix is generated. , is a directed graph, such as Figure 4 As shown.
[0032] Preferably, the specific steps for obtaining the total hidden layer vector for each historical time period include: Step c: Obtain the static airport distance map described in step b. A Distance and the m Dynamic flight flow chart for historical time periods As input information for the graph structure, for the first m Airport delay vector map for a historical time period r m Perform graph convolution calculations to obtain the first... m Airport distance hidden layer vector for each historical time period Airport traffic hidden layer vector ; Furthermore, obtain the first step described in step c. m Airport distance to hidden layer vector during historical time periods Airport traffic hidden layer vector The specific steps are as follows: Step c1, the first step described in step b m Airport delay vector map for a historical time period r m Static airport distance map A Distance and the m Dynamic flight flow chart for historical time periods As input information for the graph structure, for the first m Airport delay vector map for a historical time period r m Perform graph convolution calculations and obtain the first... mThe state vector of the convolutional hidden layer of a static airport map for a historical time period. Convolutional hidden layer state vectors of airport traffic graph ; Step c2, based on the first step described in step c1 m The state vector of the convolutional hidden layer of a static airport map for a historical time period. Convolutional hidden layer state vectors of airport traffic graph , for the m Delay graph vector for each historical time period r m Perform graph convolution calculation to obtain the first m Airport distance to hidden layer vector during historical time periods Airport traffic hidden layer vector .
[0033] Furthermore, the static airport graph convolutional hidden layer state vector described in step c1 Hmmm The expression is: ; The expression for the state vector of the convolutional hidden layer of the airport traffic map is: ; in, For the first m The convolutional hidden layer state vectors of static airport graphs for each historical time period For activation functions; D Static airport distance map matrix à D and dynamic flight flow map matrix à F The degree matrix, à D This is a static airport distance map matrix; H m For the first m The state vector of the convolutional hidden layer of a historical time period graph; W For parameter weights; A D This refers to the static airport distance; I N Let N be the identity matrix of the Nth airport dimension; For the first m Hidden state vectors of airport traffic over a historical time period; à F A dynamic flight flow map matrix; A F For dynamic flight flow m =1,2,..., M , M for MThere are 1, 2, ..., n historical time periods; N = 1, 2, ..., n, where n is the number of airports.
[0034] Furthermore, the activation function for sigmoid function, tanh function or Relu function.
[0035] Furthermore, the first step described in step c2 m Airport distance to hidden layer vector during historical time periods Airport traffic hidden layer vector The expression is: ; in, For the first m Airport distance hidden layer vector for each historical time period, For activation functions; i D*G The convolution process for the airport static distance map; r m For the first m Airport delay graph vector for each historical time period; For the first m The airport distance hidden layer vector in the previous historical time period; For the first m Hidden vectors of airport traffic over a historical time period; i F*G The convolution process for the airport dynamic flight flow map; For the first m Hidden layer vector of airport traffic in the previous historical time period; m ≥1, m =1,2,..., M , M for M A total historical period.
[0036] Step d: Design the static adaptive vector W for heterogeneous airports. distance Dynamic adaptive vectors of heterogeneous airports The airport distance hidden layer vector described in step c is respectively... and the airport traffic hidden layer vector Perform dynamic weighted fusion of heterogeneous airports to obtain the first m Total hidden layer vector for each historical time period ; Furthermore, the expression for the dynamic weighted fusion of heterogeneous airports in step d is: ; in, For the firstm Total hidden layer vector for each historical time period; W distance For heterogeneous airports, the static adaptive vector is denoted as . For the first m Airport distance hidden layer vector for each historical time period; For heterogeneous airports, the dynamic adaptive vector is denoted as . No. m Hidden vectors of airport traffic over a historical time period; m =1,2,..., M , M for M A total historical period.
[0037] Preferably, the specific steps for obtaining the decoder output hidden layer vector include: Step e, splicing step b as described in the first step m Airport delay vector map for a historical time period r m Compared with step d, the first m -1 total hidden layer vector for a historical time period As temporal correlation data for a recurrent neural network, the temporal correlation data is input into the matrix multiplication unit of a graph convolution, based on the static airport distance map described in step b and the first... m The graph convolution calculation is performed on the dynamic flight flow map of each historical time period, and then the gating vector is obtained by the first neural network function. The splicing step b described in the first step m Airport delay vector map for a historical time period r m Compared with step d, the first m -1 total hidden layer vector for a historical time period And perform the first neural network function calculation to obtain the first... m -1 historical time period update hidden layer vector Based on the gate vector, the first m -1 total hidden layer vector for a historical time period and updating hidden layer vectors Perform weighted extraction to obtain the first... m Hidden layer vectors for each historical time period The hidden layer vectors for each historical time period Iteratively obtain the final hidden layer vector for the total historical time period as the encoder output vector. ; Further, obtain the encoder output vector described in step e. The specific steps include: Step e1, splicing the first m Airport delay vector map for a historical time period r m With the m- Hidden layer vector for one historical time period Based on step b, the first m The dynamic flight flow map and static airport distance map for each historical time period are subjected to graph convolution, and nonlinear features are added after passing through the first activation function to obtain the th... m The historical time period gating vector: the m Total hidden feature gating vector for each historical time period , No. m Correct feature gating vectors for each historical time period , No. m Historical feature gating vectors for each historical time period The expression is as follows: ; in, For the first m The total hidden feature gating vector for each historical time period, θ O*G For the total hidden feature map convolution process, r m For the first m Airport delay graph vectors for a historical time period; For the first m The feature gating vector is modified for each historical time period, θ F*G To correct the feature map convolution process, For the first m Historical feature gating vectors for each historical time period sigmoid The function is the first activation function; θ I*G For the historical feature map convolution process, m≥1; Step e2, splicing the first m Airport delay map vector for each historical time period r m With the m -1 hidden layer vector for a historical time period Based on step b, the first m The dynamic flight flow map and static airport distance map for each historical time period are convolved, and then subjected to a nonlinear transformation using a second activation function to generate the first... m -1 historical time period update hidden layer vector The expression is: ; in, For the first m Update the hidden layer vector for each historical time period. r m For the first m Airport delay graph vectors for a historical time period; tanh The function is the second activation function. For the first m -1 hidden layer vector for a historical time period, m ≥1, θ U*G For graph convolution functions; Step e3, using the gating vector described in step e1 to... m -1 historical time period hidden layer vector Compared with step e2, the first m -1 historical time period update hidden layer vector Perform weighted extraction to obtain the first... m Hidden layer vectors for each historical time period The expression is: ; in, For the first m Hidden layer vectors for each historical time period tanh For the second activation function, For the first m -1 hidden layer vector for a historical time period, For the first m -1 historical time interval update hidden layer vector; Step e4: Iterate through each historical time period according to the expression described in step e3 to generate the hidden layer state vector, and obtain... M The hidden layer vector for the total historical time period is used as the encoder output vector. .
[0038] Step f, Obtain M The first m Hidden layer vectors for each historical time period ,Will M The first m Hidden layer vectors for each historical time period Concatenate the data to obtain the total historical time period concatenated hidden layer vector. The encoder output vector described in step e Concatenate the hidden layer vector with the total historical time period mentioned above. Perform function enhancements to obtain the encoder output vector updated for each total historical time period. ; Update the encoder output vector with the total historical time period mentioned above. Copy and concatenate to obtain the transpose of the encoder output mapping vector. The hidden layer vector is concatenated with the total historical time period. Similarity calculations are performed to obtain the attention weights for the total historical time period; based on these attention weights, the decoder weighted hidden layer vector is obtained. ; Furthermore, step f describes obtaining the decoder weighted hidden layer vector. The specific steps include: Step f1: Obtain respectively M The first m Hidden layer vectors for each historical time period , will the M The first m Hidden layer vectors for each historical time period Concatenate the data to obtain the total historical time period concatenated hidden layer vector. ; Step f2: Concatenate the encoder output vector described in step e. The hidden layer vector of the total historical time period described in step f1 Based on step b, the first m The encoder output mapping vector is obtained by performing graph convolution on the dynamic flight flow map and the static airport distance map for each historical time period, followed by a third activation function calculation. The total historical time period is concatenated into the hidden layer mapping vector. ; The encoder outputs a mapping vector The total historical time period is concatenated into the hidden layer mapping vector. The expression is: ; in, fullconnect It is a fully connected layer; Relu It is the third activation function; Step f3: Map the encoder output vector described in step f2. The transpose of the encoder output mapping vector is obtained by copying and concatenating the vectors based on their dimensions at each historical time period. The transpose of the encoder output mapping vector Concatenate the hidden layer mapping vector Calculate the similarity and assign attention weights to different historical time periods based on the similarity to obtain the attention weights for each historical time period; The expression for the similarity is: ; Where α is the similarity; softmax It is a normalized exponential function; Step f4: Based on the attention weights described in step f3, map the hidden layer vectors described in step f2. Weighting is performed to generate the weighted hidden layer vector of the decoder. The expression is: ; Step g: Convert the encoder output vector Weighted hidden layer vectors of decoder As input, the first m Using a dynamic flight flow map and a static airport distance map for a historical time period as a graph structure, graph convolution calculation is performed to obtain the graph convolution calculation result. Based on the graph convolution calculation result and the static adaptive vector W of the heterogeneous airport in step d, distance Dynamic adaptive vectors of heterogeneous airports Obtain the decoder output hidden layer vector ; Furthermore, in step g, the decoder outputs the hidden layer vector. The expression is: ; in, Output the distance vector; This is the flow output vector.
[0039] Preferably, the specific steps for obtaining the predicted airport network delay time for the future historical time period include: The decoder outputs the hidden layer vector as described in step g. Converted into a one-dimensional vector, this one-dimensional vector is used as a predicted value for airport network delay time in a future historical time period. r output .
[0040] Furthermore, obtain the predicted airport network delay time for the next future time period. r output ,include: The fully connected layer is designed to output the hidden layer vector from the decoder described in step g. Convert to a one-dimensional vector The single-dimensional vector The predicted airport network delay time for a future time period is obtained after calculation using the first activation function. ， The expression is: ; in, sigmoid The first activation function is a function; fullconnect It is a fully connected layer. r output This is the one-dimensional output vector of the decoder.
[0041] Furthermore, the decoder's one-dimensional output vector r output As a predicted value for airport network delays in a future historical time period, the predicted airport delay values for each historical time period are characterized as quantitative values of the ripple effect, thereby realizing the quantitative prediction of the ripple effect of the aviation airport network.
[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A dynamic spatiotemporal graph neural network model for predicting flight delays, characterized in that, It includes an input layer, a graph convolution-heterogeneous airport weighted fusion layer, a graph convolution-recurrent neural network coupling layer, and an output layer; The input layer is used to receive delay information of each airport in multiple historical time periods, flight information of each airport in each historical time period, and distance information between each airport. The graph convolution-heterogeneous airport weighted fusion layer is used to perform graph convolution calculation and dynamic weighted fusion of airport delay information in each historical time period to obtain the total hidden layer vector for each historical time period. The specific steps are as follows: The distance information between airports and the flight information of airports in each historical time period are used to perform graph convolution calculation on the airport delay graph vectors in each historical time period to obtain the airport distance hidden layer vector and airport traffic hidden layer vector in each historical time period. Heterogeneous airport dynamic weighted fusion is performed on the airport distance hidden layer vector and airport traffic hidden layer vector for each historical time period to obtain the total hidden layer vector for each historical time period. The expression for the dynamic weighted fusion of heterogeneous airports is: = + in, For the first m Total hidden layer vector for each historical time period; W distance is the static adaptive vector of the heterogeneous airport; For the first m Airport distance hidden layer vector for each historical time period; For heterogeneous airports, the dynamic adaptive vector is denoted as . The graph convolutional-recurrent neural network coupling layer couples the total hidden layer vectors of each historical time period with a graph convolutional-recurrent neural network to establish an encoder-decoder architecture and obtain the decoder output hidden layer vectors. The output layer obtains the airport network delay time for a future time period by performing a nonlinear transformation on the hidden layer vector output by the decoder.
2. The dynamic spatiotemporal graph neural network model according to claim 1, characterized in that, It also includes an airport delay map vector acquisition layer, which is used to input the delay information of each airport in the multiple historical time periods into the airport delay map vector acquisition layer and output multiple historical time period airport delay map vectors.
3. The dynamic spatiotemporal graph neural network model according to claim 1, characterized in that, The specific steps for obtaining the airport distance hidden layer vector and airport traffic hidden layer vector for each historical time period include: The distance information between airports and the flight information of airports in each historical time period are used to perform graph convolution calculation on the airport delay graph vectors in each historical time period, and the hidden state vectors of the airport distance graph convolution and the hidden state vectors of the airport traffic graph convolution are obtained respectively. Based on the airport distance graph convolutional hidden layer state vector and airport traffic graph convolutional hidden layer state vector for each historical time period, graph convolution is performed on the airport delay graph vector for each historical time period to obtain the airport distance hidden layer vector and airport traffic hidden layer vector for each historical time period.
4. The dynamic spatiotemporal graph neural network model according to claim 1, characterized in that, The expressions for the airport distance hidden layer vector and the airport traffic hidden layer vector for each historical time period are as follows: ； in, For the first m Airport distance hidden layer vector for each historical time period, For activation functions; θ D*G The convolution process for the airport static distance map; ρ m For the first m Airport delay graph vector for each historical time period; m =1,2,..., M , M This represents the total number of historical time periods. m When =1, it is the initial historical time period; For the first m -1 historical time period airport distance hidden layer vector; For the first m Hidden vectors of airport traffic over a historical time period θ F*G The convolution process for the airport dynamic flight flow map; For the first m -1 hidden layer vector of airport traffic for a historical time period.
5. The dynamic spatiotemporal graph neural network model according to claim 1, characterized in that, The specific steps for obtaining the decoder output hidden layer vector include: Based on the airport delay map vectors for each historical time period and the total hidden layer vectors for each historical time period, the temporal correlation data of the recurrent neural network is obtained, and the gating vectors for each historical time period are calculated from the temporal correlation data. By concatenating the gate vectors and the total hidden layer vectors of each historical time period and calculating similarity, the attention weights of the total historical time period are obtained, and then the decoder weighted hidden layer vectors are obtained. The decoder's weighted hidden layer vector is used as input. Graph convolution is performed based on the distance information between airports and the flight information of airports in each historical time period to obtain the calculation results. The decoder's output hidden layer vector is obtained based on the calculation results.
6. The dynamic spatiotemporal graph neural network model according to claim 5, characterized in that, The decoder outputs the hidden layer vector expression as follows: ； in, W is the hidden layer vector output by the decoder. distance Let be the static adaptive vector of the heterogeneous airport. This is the distance output vector. For heterogeneous airports, the dynamic adaptive vector is... This is the flow output vector.
7. The dynamic spatiotemporal graph neural network model according to claim 1, characterized in that, Obtaining the airport network delay time for the future time period includes: The fully connected layer is designed to convert the decoder output hidden layer vector into a one-dimensional vector. The one-dimensional vector is then processed by a first activation function to obtain the decoder's one-dimensional output vector. This one-dimensional output vector is then used to represent the airport network delay time in the future.
8. An application of a dynamic spatiotemporal graph neural network model, using the dynamic spatiotemporal graph neural network model according to any one of claims 1-7 to predict airport network delay time in the future.