Large-area flight delay propagation prediction method and system based on space-time attention deep learning network

Abstract

The invention discloses a large-area flight delay propagation prediction method and system based on a space-time attention deep learning network. The method comprises the following steps: acquiring flight data and meteorological message data of a plurality of airports, analyzing time sequence autocorrelation of the flight data and the meteorological message data, and dividing the flight data by time slices; analyzing and converting the meteorological message data to obtain meteorological message data with indexes; preprocessing the flight data divided into time slices and the meteorological message data with indexes by analogy of the input of the space-time attention network model to obtain a space-time data sequence; taking the spatio-temporal data sequence as modeling data, and obtaining a new weight model according to a spatio-temporal attention network model and a propagation dynamics method; according to the prediction result of the new weight model, improving the new weight model to obtain a multi-airport flight delay prediction model; according to the method, the space-time network model combined with propagation dynamics is constructed, and the evaluation quality of airline companies and airports on the next-day flight delay condition can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data analysis, big data and traffic delay propagation technology, and in particular to a large-area flight delay propagation prediction method and system based on a spatiotemporal attention deep learning network. Background Art

[0002] In order to adapt to the rapidly growing demand for flight transportation, airlines usually formulate compact flight plans so that the same aircraft can perform multiple flight missions. At the same time, since flights will fly between various airports, the impact of delays and congestion at one airport will be brought to other airports as flights operate between airports. This will cause the impact of previous flights to propagate to subsequent airports through the flight chain, causing the scope of flight delays to continue to expand in the air transportation network, resulting in a series of delay propagation problems, which has also become an issue that cannot be avoided in the study of air traffic delays.

[0003] Domestic and foreign scholars have adopted different methods for analyzing flight delay propagation. Some researchers have used the method of constructing a propagation dynamics model to predict the propagation of flight delays. This method using a simple dynamics model cannot fully display the behavior of delay propagation, and the accuracy of the prediction needs to be improved. There is also a combination of complex network perspectives using propagation dynamics methods to analyze delay propagation, but the networks they construct are often generated through simulation, which is still different from the actual scenario.

[0004] After research, the inventors found that the flight delay propagation prediction method in the prior art still has at least the following defects:

[0005] The prediction accuracy of delayed flights at multiple airports is low, the spatiotemporal characteristics of the constructed network are not fully utilized, and there is no method of combining physical information learning to learn the characteristics of the delay propagation process.

[0006] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0007] The purpose of the present invention is to improve the prediction accuracy of delayed flights at multiple airports.

[0008] The present invention provides a large-area flight delay propagation prediction method based on a spatiotemporal attention deep learning network, comprising the steps of:

[0009] S11, obtaining flight data and weather message data of multiple airports;

[0010] S12, analyzing the time series autocorrelation of the flight data and the weather message data, dividing the flight data into time slices; and parsing and converting the weather message data to obtain weather message data with indexes;

[0011] S13, analogizing the input of the spatiotemporal attention network model, preprocessing the flight data divided into time slices and the indexed meteorological message data to obtain a spatiotemporal data sequence; the spatiotemporal data sequence includes a three-dimensional data set of airports, time and features;

[0012] S14, using the spatiotemporal data sequence as modeling data, obtaining a new weight model according to the spatiotemporal attention network model and the propagation dynamics method; the new weight model is used for single airport flight delay prediction;

[0013] S15. According to the prediction result of the new weight model, the new weight model is improved to obtain a flight delay prediction model; the flight delay prediction model is used for multi-airport flight delay prediction.

[0014] Preferably, in an embodiment of the present invention, the preprocessing of the flight data and the indexed meteorological message data divided into time slices includes:

[0015] By analogy with the spatiotemporal attention network model and input, the flight propagation dynamics model and flight status are defined; the number of flights in each flight status at each airport and the number of each aircraft type at each airport are counted; the flight status includes: normal flight, delayed flight and resumed flight;

[0016] The missing values ​​in the flight data divided into time slices are cleaned by deletion; the indexed meteorological message data is cleaned by adjacent value filling;

[0017] Merging the cleaned flight data divided into time slices and the indexed meteorological message data;

[0018] Perform differential calculation on each flight status in the fused data to obtain the change in the number of flights in each time slice in each flight status;

[0019] The quantity change of each time slice of each flight status is used as the regression task label; and the static flight data of each airport is integrated into each dynamic flight data.

[0020] Preferably, in an embodiment of the present invention, the preprocessing of the flight data and the indexed meteorological message data divided into time slices includes:

[0021] Encode all flight data that does not require the extraction of intermediate data features, including:

[0022] A standardized method was used to encode numerical data; a word frequency-inverse document frequency method was used to process non-numerical data and then standardize it;

[0023] Extract the uncoded data and optimize and organize the data according to the historical window, prediction window and sliding step size.

[0024] Preferably, in an embodiment of the present invention, each data without coding is extracted; and each data is optimized and sorted according to the historical window, the prediction window and the sliding step size, including:

[0025] According to the step length of each learning slide of the gated recurrent unit, the time dimension of each uncoded data is sorted into the total number of steps;

[0026] The feature dimension is changed from the original dimension of each time slice to a feature set that contains all time slices at each step;

[0027] The number of flight states in each step is extracted as the starting value for prediction by the flight propagation dynamics model, the number of changes in the number of delayed flights and the number of changes in the number of restored flights in each historical window are used as added features of the prediction results of the fully connected layer, and the number of changes in the number of delayed flights and the number of changes in the number of restored flights in each prediction window are used as the true value for calculating the loss function with the predicted value.

[0028] Preferably, in an embodiment of the present invention, the new weight model includes:

[0029] Forward propagation training model and back propagation training model, where

[0030] The total loss of the spatiotemporal attention network model is calculated in the forward propagation training model using the following formula:

[0031] L = MSE(L pred )+α phy MSE(L phy )1

[0032] In the formula, L is the total loss of the spatiotemporal attention network model, MSE represents the loss function of the mean square error, and L pred is the result predicted by the fully connected layer, α phy is the weight of the prediction result of the dynamics model participating in the mean square error loss function, L phy Predict results for kinetic models;

[0033] In the back propagation training model, the loss function obtained in the forward propagation training model is chain-derived by the following formula:

[0034]

[0035] In the formula, θ k is the network parameter of the kth layer, η is the learning rate, L is the total loss function of the network model, is the gradient operator.

[0036] Preferably, in an embodiment of the present invention, the new weight model includes:

[0037] The input layer is used to input the preprocessed encoded data into the spatiotemporal attention network model;

[0038] The feature extraction layer is used to extract features through the forward propagation training model and the back propagation training model, and perform weighted optimization using the same loss function with different weights according to the prediction values ​​of different methods;

[0039] The method separation layer is used to output the results of different prediction sequences according to different methods.

[0040] Preferably, in the embodiment of the present invention, the prediction result of the new weight model includes:

[0041] Predict the number of delayed flights at the airport in the future through the fully connected layer of the spatiotemporal attention network model output;

[0042] The method separates the layer and uses another fully connected layer to predict the parameters in the flight propagation dynamics model.

[0043] Preferably, in an embodiment of the present invention, the new weight model is improved, including:

[0044] In the graph attention network of the spatiotemporal attention network model, the parameters of each node are learned by the following formula:

[0045]

[0046] In the formula, e ij is the attention score of node i to node j; a is a learnable parameter; E i and E j are different embedding vectors for different nodes; W p is the weight pool; is the feature vector representing node i; is the feature vector of node j;

[0047] Preferably, in an embodiment of the present invention, the new weight model is improved, including:

[0048] Combining the propagation dynamics to increase the dynamic state transition, including the following formula:

[0049]

[0050] In the formula, and is the change of the three states from the current moment to the next moment; α t is the propagation rate of delayed flights at time t; β t is the flight recovery rate at time t, S(t) is the number of normal flights at the previous moment; I(t) is the number of state S and state I at the previous moment; γR(t) is the flight state conversion process, γ represents the normal rate, which is the conversion rate from the restored flight state to the normal flight state.

[0051] In another aspect of the present invention, a large-area flight delay propagation prediction system based on a spatiotemporal attention deep learning network is provided, comprising:

[0052] Historical data acquisition module, used to obtain flight data and weather message data of multiple airports;

[0053] A historical data analysis and processing module, used to analyze the time series autocorrelation of the flight data and the meteorological message data, divide the flight data into time slices; and parse and convert the meteorological message data to obtain meteorological message data with indexes;

[0054] A spatiotemporal data sequence generation module is used to pre-process the flight data and the indexed meteorological message data divided into time slices by analogy with the input of the spatiotemporal attention network model to obtain a spatiotemporal data sequence; the spatiotemporal data sequence includes a three-dimensional data set of airports, time and features;

[0055] A new weight model construction module is used to obtain a new weight model based on the spatiotemporal attention network model and the propagation dynamics method using the spatiotemporal data sequence as modeling data; the new weight model is used for single airport flight delay prediction;

[0056] The flight delay prediction model construction module is used to improve the new weight model according to the prediction results of the new weight model to obtain a flight delay prediction model; the flight delay prediction model is used for multi-airport flight delay prediction.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] In the present invention, the flight data is divided into time slices by analyzing the time series autocorrelation of the acquired flight data and meteorological message data; and the meteorological message data is parsed and converted to obtain indexed meteorological message data; analogously to the input of the spatiotemporal attention network model, the flight data divided into time slices and the meteorological message data with indexes are preprocessed to obtain a spatiotemporal data sequence; the spatiotemporal data sequence obtained in this way is more standardized and easier to converge for model training; the spatiotemporal data sequence is used as modeling data, and a new weight model is obtained according to the spatiotemporal attention network model and the propagation dynamics method; according to the prediction results of the new weight model, the new weight model is improved to obtain a multi-airport flight delay prediction model; the present invention combines the spatiotemporal attention network model with the propagation dynamics model, and constructs a spatiotemporal network model combined with propagation dynamics through the advantages of deep learning and physical information learning; the delay propagation results are analyzed using quantitative indicators, which helps airlines and airports improve the quality of evaluation of next-day flight delays, and assists decision makers in making wise decisions in air traffic management, thereby reducing air traffic delays and reducing economic losses.

[0059] Furthermore, the multi-airport flight delay prediction model of the present invention combines adaptive node learning with the spatiotemporal attention network model of propagation dynamics, so that each node can learn different characteristics, avoiding the homogenization of prediction results due to shared weight parameters, and making the network have stronger nonlinear capabilities. In addition, the combined propagation dynamics state transition equation is improved to better fit the characteristics of the multi-airport delay propagation network, and a better prediction effect is achieved by combining data learning and physical information learning, thereby realizing a high-accuracy prediction of flight delays at multiple airports.

[0060] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, and to make the above and other purposes, technical features and advantages of the present invention easier to understand, one or more preferred embodiments are listed below and described in detail with the accompanying drawings as follows. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0062] Figure 1 It is a step diagram of the large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network described in the present invention;

[0063] Figure 2It is a flow chart of the large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network described in the present invention;

[0064] Figure 3 It is a structural schematic diagram of the large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network described in the present invention;

[0065] Figure 4 It is a schematic diagram of the analogy of input data between the flight delay propagation model based on the spatiotemporal attention network model described in the present invention and the spatiotemporal attention network model for disease propagation prediction;

[0066] Figure 5 It is a schematic diagram of the autocorrelation function test of the time series of the change number of delayed flights at an airport for more than 1 hour and the change number of restored flights when the time step is 4 in the present invention;

[0067] Figure 6 It is a schematic diagram of the autocorrelation function test of the time series of the change number of delayed flights and the change number of restored flights at multiple airports within 6 hours in the present invention when the time step is 4;

[0068] Figure 7 It is a schematic diagram of the autocorrelation function test of the change number of delayed flights and the change number of resumed flights at Guangzhou Baiyun Airport in the 1h and 6h time series of the present invention;

[0069] Figure 8 It is a schematic diagram comparing the disease transmission dynamics model and the flight delay transmission dynamics model described in the present invention;

[0070] Fig. 9 a is a schematic diagram of the graph attention network described in the present invention forming attention between nodes;

[0071] Fig. 9 b is a schematic diagram of the graph attention network described in the present invention forming multi-head attention between each node;

[0072] Fig.10 It is a schematic diagram of the gated neural unit in the present invention sliding on the spatial feature sequence formed by the graph attention network to perform feature learning;

[0073] Fig.11 It is a schematic diagram of the prediction result of the dynamic state transition equation of the output position of the new weight model described in the present invention;

[0074] Fig.12 is a schematic diagram of the improved FSIRS propagation dynamics model described in the present invention;

[0075] Fig.13 It is a schematic diagram of the number of delayed flights at 21 airports in different time periods within one day in the present invention;

[0076] Fig.14 It is a schematic diagram comparing the predicted values ​​and true values ​​of the original model and the improved model at Beijing Capital Airport on September 30 in the present invention;

[0077] Fig.15 It is a schematic diagram comparing the predicted values ​​and true values ​​of the original model and the improved model at Xi'an Xianyang Airport in the present invention on September 30;

[0078] Fig.16 It is a structural schematic diagram of the large-area flight delay propagation prediction system based on the spatiotemporal attention deep learning network described in the present invention. DETAILED DESCRIPTION

[0079] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation modes.

[0080] Unless expressly stated otherwise, throughout the specification and claims, the term "comprising"

[0081] The term “comprising” or “including” or its variations such as “comprise” or “comprises” will be understood as including the stated elements or components but does not exclude other elements or components.

[0082] In this document, the terms "first", "second", etc. are used to distinguish two different elements or parts, and are not used to limit a specific position or relative relationship. In other words, in some embodiments, the terms "first", "second", etc. can also be interchangeable.

[0083] Embodiment 1

[0084] In order to improve the prediction accuracy of airport flight delays with multiple nodes, such as Figure 1 and Figure 2 As shown, in an embodiment of the present invention, a large-area flight delay propagation prediction method based on a spatiotemporal attention deep learning network is provided, comprising the steps of:

[0085] S11, obtaining flight data and weather message data of multiple airports;

[0086] In the embodiment of the present invention, all the flight operation information between the airports that can be collected is collected, and the routine aviation weather report (Meteorological Terminal Aviation Routine Weather Report, METAR) and terminal airport weather forecast (Terminal Aerodrom-e Forecasts, TAF) messages of the corresponding time period are obtained as the original data of the meteorological message.

[0087] In the implementation of the present invention, it is necessary to analyze the structure of the spatio-temporal attention network model, including: the method of the spatio-temporal attention network (STAN) model is a model for disease transmission prediction. After the data is processed, it is input into the spatio-temporal network composed of a graph attention network and a gated recurrent unit for learning, and the prediction result is obtained. The output part is restricted by the propagation dynamics model, and the loss function is constructed to update the model parameters. Since the process is similar, the model can be used by analogy in the large-scale flight delay propagation prediction task.

[0088] The input part of the model mainly consists of three parts: the construction of an undirected graph, static data, and dynamic data. The construction of the undirected graph of the disease spread prediction model is determined by factors such as the distance between cities and population density. That is, the adjacency matrix is ​​constructed according to the strength of the correlation between cities. Among multiple airports, the air traffic conditions between airports can be clearly known, so the adjacency matrix is ​​constructed according to the air traffic conditions. For static data, the disease spread prediction model obtains the population number and longitude and latitude of each city. By analogy, the number of parking spaces and longitude and latitude of each airport can be taken as static data for flight delay propagation prediction. The structure of the large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network, such as Figure 3 As shown in Figure 2; the flight delay propagation model based on the spatiotemporal attention network model is analogous to the input data of the spatiotemporal attention network model used for disease spread prediction, as shown in Figure 2. Figure 4 As shown, according to Figure 3 and Figure 4 It can be seen that in the adjacency matrix part, the disease data calculates a correlation weight W through the population of the two cities and the distance between them. ij , and then for each city, select places with a correlation greater than 18 and not in the same city to establish a connection, that is, in the adjacency matrix S ij Indicates whether city i is associated with city j. If so, it is marked as 1. If not, it is marked as 0.

[0089] The flight data is directly counted based on the historical flight data. If there is a flight that has flights between two of the 21 airports, it is marked as 1 in the adjacency matrix, otherwise it is marked as 0. Because the flight's departure delay will also affect its departure airport, the case where the two airports are the same airport is not discarded in the flight data. In this case, a more comprehensive adjacency matrix M can be constructed. 21×21 To represent the impact between airports; static data is the number of parking spaces, longitude and latitude of each airport by analogy with the city population, longitude and latitude in the disease data.

[0090] The disease transmission dynamics model used in the disease transmission prediction model is the Susceptible-Infective-Recovered (SIR) model, which includes three states: normal population S, infected population I, and recovered population R. Its transmission process can be expressed by the following state transition equation:

[0091]

[0092] in, and represents the change of the three states from the current moment to the next moment, β represents the infection rate, γ represents the recovery rate, S t and I t Indicates the number of states S and I at the previous moment.

[0093] S12, analyzing the time series autocorrelation of the flight data and the weather message data, dividing the flight data into time slices; and parsing and converting the weather message data to obtain weather message data with indexes;

[0094] In an embodiment of the present invention, the acquired meteorological message data is analyzed, wherein METAR is a routine observation message of the current weather conditions, which is usually issued at a fixed frequency (such as continuous hourly or half-hourly), including temperature, dew temperature, wind direction, cloud information, visibility and air pressure readings for correcting the altimeter for pilots. In addition, METAR messages may also include information on precipitation, thunder and other special weather conditions. The content of TAF messages is similar to that of METAR messages, but is generally updated 4 times a day and issued every 6 hours. The above messages are parsed by R language methods and converted into feature data indexed by time and various attributes.

[0095] Since the time interval between each flight data is very short and cannot form a one-to-one correspondence with the weather data, the 6-hour interval of TAF weather data and the 1-hour interval of METAR weather data are analyzed. The analysis method mainly uses the Auto Correlation Function (ACF) to observe the autocorrelation of time series. The autocorrelation function describes the correlation between a set of time series and a set of time series separated by n time slices. This method can be used to analyze whether the flight data has a certain direct connection with the data after several time slices when it is counted into 1-hour or 6-hour time slices. When this connection is strong, the prediction model can more accurately capture the characteristics and rules of the data before and after in the time dimension, which is very important for more accurate prediction results.

[0096] like Figure 5 and Figure 6As shown in the figure, they are the autocorrelation coefficients of each airport at 1h and 6h when there are four time steps different from the first moment; in addition to multiple airports, Guangzhou Baiyun Airport is selected as an example to show the change of the autocorrelation coefficients of 1h and 6h at a single airport over time, as shown in the figure Figure 7 As shown in the figure, it shows the change of the autocorrelation function value over time. The gray shadow represents the confidence level. The numerical points outside the gray shadow represent the credibility. It can be observed that when 6 hours is used as the time slice, every four time steps have a strong temporal autocorrelation with the start time, that is, the change of the number of delayed flights every day has a certain periodicity. Figure 7 It can be clearly seen from the four figures that when 6h is selected as the time division, the time series autocorrelation of the data is stronger, which is also conducive to the subsequent learning of the time characteristics of the data.

[0097] Through the above verification, it is found that when the flight data is divided into 6-hour time slices, its autocorrelation is stronger and has a periodic pattern over time. When the flight data is divided into 1-hour time slices, its autocorrelation is weaker and has no periodic pattern over time, showing a random state. Therefore, the data processed by TAF messages is used as the weather data spliced ​​with the airport data in this method.

[0098] S13, analogizing the input of the spatiotemporal attention network model, preprocessing the flight data divided into time slices and the indexed meteorological message data to obtain a spatiotemporal data sequence; the spatiotemporal data sequence includes a three-dimensional data set of airports, time and features;

[0099] In the embodiment of the present invention, the structure of the large-area flight delay propagation prediction method based on the spatiotemporal attention deep learning network and the spatiotemporal attention network model; firstly, the flight data divided into time slices and the indexed meteorological message data are preprocessed, specifically as follows Figure 3 As shown, the spatiotemporal network data preprocessing part.

[0100] By analogy with the spatiotemporal attention network model, the flight propagation dynamics model is defined; and the flight status is defined by analogy with the input of the spatiotemporal attention network model; among them, the dynamic data needs to label the dynamic status of the flight, such as Figure 8 As shown in the figure, in the disease transmission dynamics model SIR, there are normal population S, infected population I and recovered population R, which can be divided into three states: normal flight S, delayed flight I and recovered flight R in the flight delay transmission dynamics model FSIR. The take-off and landing delay time of the aircraft is used as the division standard. If the take-off delay is less than or equal to 15 minutes, it is state S. If the take-off delay time is greater than 15 minutes and the landing delay is greater than 0 minutes, it is state I, and the remaining flights are in state R.

[0101] The number of flights in each flight status at each airport and the number of each aircraft type at each airport are counted; after dividing the historical flight data by airport, the number of flights in the three dynamic states and the number of different aircraft types at each airport are counted using the same time interval as the weather data as the time slice.

[0102] Then, the missing values ​​in the flight data divided into time slices are cleaned by deleting them; the indexed meteorological message data are cleaned by filling in adjacent values; the cleaned flight data divided into time slices and the indexed meteorological message data are fused; since the time slices used for statistics are the same as the weather data, each flight data has the meteorological data at the corresponding time after fusion. According to the data used in the disease transmission prediction model, the three flight statuses of each airport in the fused multi-airport data are differentially calculated to obtain the quantity changes dS, dI and dR of each time slice of the three flight statuses as regression task labels, and the static data of each airport is fused into each dynamic data by copying. All flight data that do not need to extract intermediate data features are encoded, and the standardized method is used to encode the numerical data. The term frequency-inverse document frequency (TF-IDF) method is used to process the non-numerical data and then standardize them to obtain the encoded three-dimensional data set consisting of airports, times and features.

[0103] Furthermore, in an embodiment of the present invention, the uncoded data in the original airport flight delay data is screened out for extraction; and each data is optimized and sorted according to the historical window, the prediction window and the sliding step size.

[0104] Specifically, from the original airport flight delay data, the time dimension of the data is reorganized into the total number of steps according to the step size of each learning slide of the gated recurrent unit, the airport dimension remains the number of airports, and the feature dimension is changed from the original dimension of each time slice to a feature set of all time slices in each step; the number of flight status (i.e., S, I, R) of each step is extracted as the starting value for the prediction of the flight propagation dynamics model, the number of changes in the number of delayed flights (dI) and the number of changes in the number of restored flights (dR) in each historical window are used as added features of the prediction results of the fully connected layer, and the number of dI and dR in each prediction window are used as the true value of the loss function calculated with the predicted value; excluding the above data extracted from the original data, the original airport data is encoded as a whole and then the encoded data set is divided into training set, validation set and test set.

[0105] S14, using the spatiotemporal data sequence as modeling data, obtaining a new weight model according to the spatiotemporal attention network model and the propagation dynamics method; the new weight model is used for single airport flight delay prediction;

[0106] In an embodiment of the present invention, the processed data is sent to the STAN model for training to obtain a new weight model, and the change amount of each time slice of the three dynamic states of the processed multi-airport data set is differentially calculated to obtain the change number of normal flights dS, the change number of delayed flights dI and the change number of restored flights dR as labels, and the data set structure is reorganized according to the historical window, prediction window and time step set by the hyperparameters, and the variables required for the intermediate calculation are extracted. After encoding all the remaining unencoded data, the organized data is sent to the STAN model, and the airport is selected to predict the number of delayed flights at a single airport in the future.

[0107] Among them, the new weight model includes:

[0108] Forward propagation training model and back propagation training model, where

[0109] The total loss of the spatiotemporal attention network model is calculated in the forward propagation training model using the following formula:

[0110] L = MSE(L pred )+α phy MSE(L phy )1

[0111] In the formula, L is the total loss of the spatiotemporal attention network model, MSE represents the loss function of the mean square error, and L pred is the result predicted by the fully connected layer, α phy is the weight of the prediction result of the dynamics model participating in the mean square error loss function, L phy Predict results for kinetic models;

[0112] In the back propagation training model, the loss function obtained in the forward propagation training model is chain-derived by the following formula:

[0113]

[0114] In the formula, θ k is the network parameter of the kth layer, η is the learning rate, L is the total loss function of the network model, is the gradient operator.

[0115] New weight model, graph spatiotemporal network model such as Figure 3 As shown in the figure, it is mainly divided into three parts: the input layer is used to input the preprocessed encoded data into the spatiotemporal attention network model; the feature extraction layer is used to extract features through the forward propagation training model and the back-propagation training model, and perform weighted optimization according to the prediction values ​​of different methods using the same loss function but different weights; the method separation layer is used to output the results of different prediction sequences according to different methods.

[0116] The input layer includes a location map, static data and dynamic data. The location map is constructed by an adjacency matrix and is an undirected graph without edge weights. If the median value of the adjacency matrix is ​​1, the corresponding two airports in the graph are connected by an edge. According to the method used in this patent, the influence between airports is propagated through flights between airports. The undirected graph constructed in this way can intuitively represent the association of the airport aviation network diagram. The static data includes three attributes: the number of parking spaces, longitude and latitude, which are integrated into each of the time series of each airport by replication. The dynamic data is mainly based on the number of propagation dynamic states, including the number of take-off and landing aircraft of each type, the total number of take-off and landing flights, the average delay time for take-off and landing, and TAF weather message data.

[0117] The following defines the input data: N represents the total number of network nodes, that is, the number of airports, and X represents the shape of the input, which is N×(T×(F S +F D +F T ))'s dynamic feature matrix, where F S represents the number of static features corresponding to the airport, F D and F T represents the number of flight characteristics and weather characteristics of each airport, T represents the number of dynamic characteristics of each airport, that is, the total number of time steps from the beginning to the end. The task wants to predict the value of the number of delayed flights I(t)+R(t) at the airport in the tth time step in the future. The input data is constructed into a spatiotemporal network That is to say, t1-t0=6h, is time point t i The instantaneous network contains all the characteristics of all airports at that moment.

[0118] The feature extraction layer consists of two parts: the graph attention network (GAT) and the gated recurrent unit (GRU), which are responsible for learning the spatial and temporal features of the input data respectively; in the graph attention network, such as Fig. 9 As shown in a, the input of the layer is a set of airport node features. Where F is the number of node features of each airport, and this layer generates a new set of node features. This process consists of a learnable weight matrix W∈R F×F′ This is achieved through linear transformation. The weight matrix is ​​shared among all nodes, and then the shared attention mechanism is performed on the result of the concatenation of two nodes after linear transformation. The above process can be expressed as follows:

[0119]

[0120] In the formula, e ijis the importance of airport node j to airport node i. In order to make the coefficients easier to compare between different nodes, a softmax function is used to calculate the attention scores e of all i nodes and their adjacent nodes. ij Normalize:

[0121]

[0122] In the formula, N i is the set of neighboring points of node i. In the experiment, the attention mechanism a is a single-layer feedforward neural network composed of a weight vector After parameterization and application of the LeakyReLU nonlinear activation function, the coefficients calculated by the attention mechanism can be expressed as follows after full expansion:

[0123]

[0124] In the formula, is the transpose of the weight vector, || is the concatenation operation, and the normalized attention coefficient is used to calculate the linear combination of the features corresponding to them as the final output feature of each node; the above process can also use multi-head attention to stabilize the learning process:

[0125]

[0126] In the formula, σ is the activation function, Because the k-th attention mechanism Calculate the normalized attention coefficient, W k is the weight matrix of the corresponding input linear transformation, that is, k different linear transformations are used to calculate the attention scores between node i and the adjacent point set, and then all these attention scores are concatenated, so the final output is It is composed of KF' features of each node. This is the output method of the first multi-head attention learning. This model adopts the second method, that is, after executing multi-head attention, the average is used to replace the first method, and the use of the activation function is delayed, and finally the node update features of comprehensive multi-angle attention learning are obtained:

[0127]

[0128] like Fig. 9 As shown in Figure b, there is an attention mechanism with three attention heads. After a node learns the attention scores of itself and surrounding nodes through three sets of different parameter matrices, these attention scores are concatenated or averaged as the new feature value of the node to continue subsequent training. In this way, each airport in the airport network can learn all the airports connected to it and its own influence in the process of flight delay propagation.

[0129] After the graph attention network learns the spatial characteristics of the airport network, the gated recurrent unit learns its temporal characteristics based on the output spatial feature sequence of the airport network, such as Fig.10 As shown in the figure, the graph neural network starts from the head of the sequence data composed of the airport network, slides a certain step each time, extracts spatial features from the data in the fixed historical window, and then inputs the data with extracted spatial features into the gated recurrent unit. The process demonstrated in the figure sets the historical window to 4 and the prediction window to 5, that is, the gated recurrent unit predicts the output result through 4 time slices of historical data each time.

[0130] Taking a sliding step length of one time slice as an example, in the figure, the sequence fragment {X2, X3, X4, X5} in the time series is currently learned, and the spatial and temporal features of the sequence data are extracted through the gated recurrent units of the graph attention network. The subsequent prediction of the change in the number of flight delays in the next five time slices {X6, X7, X8, X9, X10} is 10}, then slide back one time step, learn the data features of the window {X3, X4, X5, X6}, and predict the window {X7, X8, X9, X 10 , X 11}’s flight delays change value.

[0131] By analogy, due to the network structure characteristics of the gated recurrent unit, after each time it learns the time characteristics of the time series, it can pass the learned information to the next time step, and then selectively retain historical information and learn new information based on the characteristics of the next window. This feature makes the gated recurrent unit network structure very effective for learning time series data with certain temporal autocorrelation.

[0132] After the two modules of the feature extraction layer complete the extraction and learning of the spatial and temporal features of the data, the method separation layer receives its output results, and uses three fully connected layers to predict the changes in the number of delayed flights dI, the changes in the number of restored flights dR, and the parameters of the propagation dynamics state transition equation. The changes in the number of delayed flights dI and the changes in the number of restored flights dR of the last time slice of each historical window extracted by the data processing part are taken as additional information and spliced ​​with the output results of the feature extraction layer, and are input together into the fully connected layer for prediction. This processing can combine the explicit historical information provided by humans with the implicit historical information processed by the gated recurrent unit, so as to better predict future values ​​based on time series.

[0133] The output of the first and second fully connected layers is the direct prediction value dI of the change in the number of delayed flights dI and the change in the number of restored flights dR in the future prediction window (four time slices) p and dR p, the prediction result of the third fully connected layer is the parameter propagation rate α and recovery rate β in the propagation dynamics state transition equation. Combined with the number of normal flights S, delayed flights I and restored flights R extracted in each time slice, the FSIR state transition equation is used to calculate the value of the future prediction window. Since the time slices have been divided into 6h in advance, the differential equation format of the original state transition equation can be simplified to:

[0134]

[0135] In the formula, S(t) is the number of normal flights in the previous time slice S; I(t) is the number of delayed flights I, α t The propagation rate obtained in the current prediction window, β t is the recovery rate parameter, dS(t), dI(t) and dR(t) are the number of changes in normal flights, delayed flights and restored flights in the next time slice, respectively. According to the task requirements, only the last two equations in the equation group need to be used to predict dI(t) and dR(t), as shown in the following example: Fig.11 As shown, assuming that the prediction window has four time slices of values ​​to be predicted, the S, I and R values ​​of the previous time slice are S0, I0 and R0, and the values ​​dI1 and dR1 of the first time slice in the prediction window are obtained according to the state transition equation. Adding them to the I0 and R0 values ​​to obtain I1 and R1 is then input into the state transition equation to obtain dI2 and dR2.

[0136] By analogy, the iterative calculation is continued to obtain the values ​​of the change in the number of delayed flights and the change in the number of restored flights in the four time slices in the prediction window; the method separation layer obtains the direct prediction value L by two methods: direct prediction through the deep learning fully connected layer and iterative calculation of the propagation dynamic state transition equation. pred and the kinetic prediction value L phy , back propagation is performed through the loss function to update the model learning parameters:

[0137] L = MSE(L pred )+α phy MSE(L phy )

[0138] In the formula, MSE(·) is the mean square error loss function, α phy is the weight of the propagation dynamics prediction result in the loss function, and L is the total value of the loss function.

[0139] In the embodiment of the present invention, the prediction result of the new weight model includes:

[0140] The fully connected layer Linear of the output part of the spatiotemporal attention network model predicts the number of delayed flights at the airport in the future. There are two fully connected layers. The input dimension of these two fully connected layers is +2. The dI and dR of the historical window are added as the additional features of the prediction results of the fully connected layer. The output dimension is 1. The dI and dR values ​​of each time slice in the window are predicted respectively. These two values ​​are the changes in the number of delayed flights and resumed flights in the future.

[0141] The method separates the layers and uses another fully connected layer to predict the parameters in the dynamic model. The input dimension is the same as the above two fully connected layers, and the output dimension is 2, which is the two parameters α and β in the dynamic model. The number of delayed flights at the airport is:

[0142] N d =N0+(dI+dR)

[0143] In the formula, N d is the number of delayed flights in the future prediction window, N0 is the number of delayed flights at the airport in the last time slice of the historical window, dI and dR are the predicted changes in the number of delayed flights and restored flights in the prediction window, and the sum of the two is the change in the number of delayed flights at the airport in the future.

[0144] S15. According to the prediction result of the new weight model, the new weight model is improved to obtain a flight delay prediction model; the flight delay prediction model is used for multi-airport flight delay prediction.

[0145] In an embodiment of the present invention, the prediction model for flight delay propagation at a single airport only extracts the maximum value of all node features in the last layer of the feature extraction layer of the model, that is, the most significant feature is selected for subsequent prediction. In order to predict all nodes, the extraction operation of this layer of maximum value is removed, and flight delay prediction is performed for multiple airports. However, the prediction effect is not ideal, and the prediction sequence patterns of future delayed flights at different airports show similar patterns. The parameter sharing nature of the graph attention network for all nodes may cause this situation. Therefore, the parameter learning part in the graph attention network is modified accordingly, so that each node can learn different parameters according to its different patterns, which is more conducive to learning the heterogeneous features between different airports. The following formula is the process of each node of the graph attention network capturing the information of surrounding nodes to form weight edges:

[0146]

[0147] In the formula, e ij is the attention score of node i to node j, a and W are both learnable parameters, and Represent the feature vectors of node i and node j respectively. The above formula means that after linear transformation of the features of nodes i and j, dot multiplication with an attention matrix is ​​performed to obtain the attention score of node i to node j. In this case, the linear transformation W is the same for each node, which will lead to the above problem. Therefore, this process is improved. Since there may be multiple nodes in the graph neural network, if different linear transformations are used for each node feature, the number of model parameters will be greatly increased. Therefore, an embedding vector E is first used to linearly transform the features, and then a weight pool W is set. p , so that the linearly transformed features can learn different parameters from the weight pool, so as to achieve the purpose of learning different parameters for different nodes. The following is the improved formula:

[0148]

[0149] In the formula, e ij is the attention score of node i to node j; a is a learnable parameter; E i and E j are different embedding vectors for different nodes; W p is the weight pool; is the feature vector representing node i; is the feature vector of node j;

[0150] Furthermore, the new weight model is improved, including: combining the propagation dynamics to add dynamic state transition, so that the state transition of the entire multi-airport flight delay propagation forms a closed loop, which is more in line with the actual situation that the flight delay state is constantly changing during the flight delay propagation process. It includes the following formulas:

[0151]

[0152] In the formula, and is the change of the three states from the current moment to the next moment; α t is the propagation rate of delayed flights at time t; β t is the flight recovery rate at time t, S(t) is the number of normal flights at the previous moment; I(t) is the number of state S and state I at the previous moment; γR(t) is the flight state conversion process, γ represents the normal rate, which is the conversion rate from the restored flight state to the normal flight state. Fig.12 As shown, it becomes the propagation dynamics state transition equation of FSIRS. By modifying the propagation dynamics state transition equation, the characteristics of flight delays in the propagation process can be better learned.

[0153] According to the above improvements, the flight data and weather data are processed and input into the model. The initial learning rate is set to 0.001, the dropout rate is 0.2, the GAT weight pool embedding dimension is 10, the historical window is 16, the prediction window is 4, and the sliding window is 4. The Adam optimizer is used for gradient update. MSE is selected as the loss function. After 50 rounds of forward propagation and back propagation, the new weight model is trained, and the parameters are continuously optimized to obtain a network model with hyperparameters.

[0154] Specifically, the divided test set airport data is input into the trained flight delay impact prediction model to obtain the final output result. The test results are measured by the model's performance indicators MSE, MAE, and CCC values ​​to see whether the results meet the expected requirements. If not, the network model is retrained. The specific meaning and formula of each indicator are as follows:

[0155] The mean square error (MSE) refers to the sum of the squares of the errors between the true value of the regression result and the predicted value, and then takes the average, which reflects the effectiveness of the prediction result.

[0156]

[0157] Mean-Absolute Error (MAE) refers to the sum of the absolute values ​​of the errors between the true value of the regression result and the predicted value, and then taking the average, which also reflects the effectiveness of the prediction results.

[0158]

[0159] The average consistency correlation coefficient (average Concordance Correlation Coefficient, CCC) is as follows, where μ x and μ y is the mean of the two variables, σ x 2 and σ y 2 is the corresponding variance, ρ is the Pearson correlation coefficient between the two variables (a statistical indicator that measures the degree of linear correlation between two variables), and CCC is often used to measure the consistency between two variables in regression problems.

[0160]

[0161] As shown in Table 1, the model part trains and tests the original model STAN, the spatio-temporal attention network with adaptive node parameter learning (STAN_ANL), the spatio-temporal attention network with propagation dynamics SIRS model (STAN_SIRS), and the spatio-temporal attention network weight model with adaptive node learning and propagation dynamics (STAN-ANLSIRS) with two parts modified at the same time. At the same time, the test results of the four parameter models using different functions sigmoid and tanh when predicting dynamic parameters are compared. The values ​​in the brackets in each cell represent the numerical range of the prediction results of the 21 airports. This is due to the characteristics of different airports, such as the number of flights or the size of the airport. The middle value of each cell represents the average level of the prediction results of the 21 airports.

[0162] Table 1

[0163]

[0164] From a global perspective, no matter one or two modules are modified, the model effect is better than the original model in terms of mean square error, mean absolute error and average consistency correlation coefficient. When one module is modified alone, the modified model for GAT node adaptive parameter learning is better than the modified model for the dynamic state transition equation. At the same time, it can be seen that when the dynamic state transition equation is not improved to SIRS, the model using the tanh function to normalize the dynamic parameter prediction results is better than the case of using the sigmoid function (STAN, STAN_ANL). After improving the dynamic state transition equation to SIRS, the case of using the sigmoid function is better than the case of using the tanh function. This is because the SIR model before improving the dynamic state transition equation has a rough conversion process, which may be in the learning process. The propagation process in some cases is lost. At this time, using the tanh function to map the dynamic parameters to (-1,1) can make up for the learning of the propagation process of the SIR state transition equation on the negative semi-axis to a certain extent compared with the sigmoid function mapping to (0,1). After being modified to the dynamic state transition equation of SIRS, it can itself learn various different propagation processes more comprehensively. Using the tanh function will increase the instability of the prediction results compared to the sigmoid function, resulting in negative optimization (STAN_SIRS, STAN_ANLSIRS). Therefore, after the GAT node parameter adaptive learning module and the dynamic state transition equation are improved at the same time, the sigmoid function is selected to normalize the dynamic parameter prediction results. At this time, as shown in Table 1, the model test prediction results in this case are optimal.

[0165] The prediction window is set to 4, that is, the number of delayed flights in the next four time slices is predicted. The length of a time slice is 6 hours, so four time slices can just represent the flight delays in the next day. Fig.13As shown in the figure, the grayscale matrix constructed according to the prediction window values ​​of 21 airports has a lighter color when the number of delayed flights is larger, and a darker color when the number of delayed flights is smaller. In terms of time, the 21 airports have more delayed flights in the first three time periods of the day, and fewer delayed flights in the last time period. This is because the number of flights at each airport is relatively small at night, and the number of delayed flights is the largest in the second time period, that is, most airports often have the largest number of flights from 6 am to 12 pm, and are most likely to be affected by various factors and cause flight delays. From the perspective of airports, the number of delayed flights at some airports throughout the day is not as large as that of flights at other airports in a certain time period. This is caused by the scale and function of the airport. Some airports are feeder airports, and the airports themselves are not large, so the number of flights per day is small, and they are not easily affected by delay propagation. However, some airports are large in scale and bear the function of hub airports, accepting flights from multiple airports. The number of flights per day is large, and they are also easily affected by delay propagation.

[0166] Fig.14 and Fig.15 The predictions for Beijing Capital Airport (ZBAA) and Xi'an Xianyang Airport (ZLXY) are shown respectively. The improved model is compared with the original model and the actual value. It can be seen that the improved model is closer to the true value than the original model. It has good prediction ability for predicting the number of future flight delays at multiple airports, making the reference to the prediction results of large-scale flight delay propagation more practical.

[0167] In summary, the embodiments of the present invention divide the flight data into time slices by analyzing the time series autocorrelation of the acquired flight data and meteorological message data; and parse and convert the meteorological message data to obtain indexed meteorological message data; analogously to the input of the spatiotemporal attention network model, the flight data divided into time slices and the meteorological message data with indexes are preprocessed to obtain a spatiotemporal data sequence; the spatiotemporal data sequence obtained in this way is more standardized and easier to converge for model training; the spatiotemporal data sequence is used as modeling data, and a new weight model is obtained according to the spatiotemporal attention network model and the propagation dynamics method; according to the prediction results of the new weight model, the new weight model is improved to obtain a multi-airport flight delay prediction model; the present invention combines the spatiotemporal attention network model with the propagation dynamics model, and constructs a spatiotemporal network model combined with propagation dynamics through the advantages of deep learning and physical information learning; the delay propagation results are analyzed using quantitative indicators, which helps airlines and airports improve the quality of assessment of next-day flight delays, and assists decision makers in making wise decisions in air traffic management, thereby reducing air traffic delays and reducing economic losses.

[0168] Furthermore, the multi-airport flight delay prediction model of the present invention combines adaptive node learning with the spatiotemporal attention network model of propagation dynamics, so that each node can learn different characteristics, avoiding the homogenization of prediction results due to shared weight parameters, and making the network have stronger nonlinear capabilities. In addition, the combined propagation dynamics state transition equation is improved to better fit the characteristics of the multi-airport delay propagation network, and a better prediction effect is achieved by combining data learning and physical information learning, thereby realizing a high-accuracy prediction of flight delays at multiple airports.

[0169] Embodiment 2

[0170] Corresponding to the method embodiment, on the other side of the embodiment of the present invention, a large-area flight delay propagation prediction system based on a spatiotemporal attention deep learning network is also provided. Fig.16 A schematic diagram of the structure of a large-area flight delay propagation prediction system based on a spatiotemporal attention deep learning network provided by an embodiment of the present invention is shown. The large-area flight delay propagation prediction system based on a spatiotemporal attention deep learning network is Figure 1 The system corresponding to the large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network described in the corresponding embodiment is implemented by means of a virtual system Figure 1 In the corresponding embodiment, the large-area flight delay propagation prediction method based on the spatiotemporal attention deep learning network, each virtual module constituting the large-area flight delay propagation prediction system based on the spatiotemporal attention deep learning network can be executed by an electronic device, such as a network device, a terminal device or a server. Specifically, the large-area flight delay propagation prediction system based on the spatiotemporal attention deep learning network in the embodiment of the present invention includes:

[0171] Historical data acquisition module 01, used to obtain flight data and weather message data of multiple airports;

[0172] The historical data analysis and processing module 02 is used to analyze the time series autocorrelation of the flight data and the weather message data, divide the flight data into time slices, and parse and convert the weather message data to obtain weather message data with indexes;

[0173] The spatiotemporal data sequence generation module 03 is used to pre-process the flight data and the indexed meteorological message data divided into time slices by analogy with the input of the spatiotemporal attention network model to obtain a spatiotemporal data sequence; the spatiotemporal data sequence includes a three-dimensional data set of airports, time and features;

[0174] In the embodiment of the present invention, the flight propagation dynamics model is defined by analogy with the spatiotemporal attention network model; and the flight status is defined by analogy with the input of the spatiotemporal attention network model; wherein the dynamic data needs to label the dynamic status of the flight, such as Figure 8 As shown in the figure, in the SIR disease transmission dynamics, there are normal population S, infected population I and recovered population R, which can be divided into three states in the flight data: normal flight S, delayed flight I and recovered flight R. The take-off and landing delay time of the aircraft is used as the division standard. If the take-off delay is less than or equal to 15 minutes, it is state S. If the take-off delay time is greater than 15 minutes and the landing delay is greater than 0 minutes, it is state I. The rest of the flights are in state R.

[0175] The number of flights in each flight status at each airport and the number of each aircraft type at each airport are counted; after dividing the historical flight data by airport, the number of flights in the three dynamic states and the number of different aircraft types at each airport are counted using the same time interval as the weather data as the time slice.

[0176] Then, the missing values ​​in the flight data divided into time slices are cleaned by deleting them; the indexed meteorological message data are cleaned by filling in adjacent values; the cleaned flight data divided into time slices and the indexed meteorological message data are fused; since the time slices used for statistics are the same as the weather data, each flight data has the meteorological data at the corresponding time after fusion. According to the data used in the disease transmission prediction model, the three flight statuses of each airport in the fused multi-airport data are differentially calculated to obtain the quantity changes dS, dI and dR of each time slice of the three flight statuses as regression task labels, and the static data of each airport is fused into each dynamic data by copying. All flight data that do not need to extract intermediate data features are encoded, and the standardized method is used to encode the numerical data. The term frequency-inverse document frequency (TF-IDF) method is used to process the non-numerical data and then standardize them to obtain the encoded three-dimensional data set consisting of airports, times and features.

[0177] Furthermore, in an embodiment of the present invention, the uncoded data in the original airport flight delay data is screened out for extraction; and each data is optimized and sorted according to the historical window, the prediction window and the sliding step size.

[0178] A new weight model construction module 04 is used to obtain a new weight model based on the spatiotemporal attention network model and the propagation dynamics method using the spatiotemporal data sequence as modeling data; the new weight model is used for single airport flight delay prediction;

[0179] In the embodiment of the present invention, the prediction result of the new weight model includes:

[0180] The fully connected layer Linear of the output part of the spatiotemporal attention network model predicts the number of delayed flights at the airport in the future. There are two fully connected layers. The input dimension of these two fully connected layers is +2. The dI and dR of the historical window are added as the additional features of the prediction results of the fully connected layer. The output dimension is 1. The dI and dR values ​​of each time slice in the window are predicted respectively. These two values ​​are the changes in the number of delayed flights and resumed flights in the future.

[0181] The method separates the layers and uses another fully connected layer to predict the parameters in the dynamic model. The input dimension is the same as the above two fully connected layers, and the output dimension is 2, which is the two parameters α and β in the dynamic model. The number of delayed flights at the airport is:

[0182] N d =N0+(dI+dR)

[0183] In the formula, N d is the number of delayed flights in the future prediction window, N0 is the number of delayed flights at the airport in the last time slice of the historical window, dI and dR are the predicted changes in the number of delayed flights and restored flights in the prediction window, and the sum of the two is the change in the number of delayed flights at the airport in the future.

[0184] The flight delay prediction model building module 05 is used to improve the new weight model according to the prediction result of the new weight model to obtain a flight delay prediction model; the flight delay prediction model is used for multi-airport flight delay prediction.

[0185] Improvements to the new weight model include: adding dynamic state transitions in combination with propagation dynamics, so that the state transitions of the entire multi-airport flight delay propagation form a closed loop, which is more in line with the actual situation that the flight delay state is constantly changing during the flight delay propagation process. The following formulas are included:

[0186]

[0187] In the formula, and is the change of the three states from the current moment to the next moment; α t is the propagation rate of delayed flights at time t; β t is the flight recovery rate at time t, S(t) is the number of normal flights at the previous moment; I(t) is the number of state S and state I at the previous moment; γR(t) is the flight state conversion process, γ represents the normal rate, which is the conversion rate from the restored flight state to the normal flight state. Fig.12As shown, it becomes the propagation dynamics state transition equation of FSIRS. By modifying the propagation dynamics state transition equation, the characteristics of flight delays in the propagation process can be better learned.

[0188] It should be noted that the specific implementation and technical effects of the large-area flight delay propagation prediction system based on spatiotemporal attention deep learning network in the embodiment of the present invention can be referred to Figure 1 The corresponding large-scale flight delay propagation prediction method based on spatiotemporal attention deep learning network will not be elaborated here.

Claims

1. A large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network, characterized in that: Includes steps: S11, obtaining flight data and weather message data of multiple airports; S12, analyzing the time series autocorrelation of the flight data and the meteorological message data, and dividing the flight data into time slices; and parsing and converting the meteorological message data to obtain meteorological message data with index; S13, analogizing the input of the spatiotemporal attention network model, preprocessing the flight data divided into time slices and the indexed meteorological message data to obtain a spatiotemporal data sequence; the spatiotemporal data sequence includes a three-dimensional data set of airports, time and features; S14, using the spatiotemporal data sequence as modeling data, obtaining a new weight model according to the spatiotemporal attention network model and the propagation dynamics method; the new weight model is used for single airport flight delay prediction; S15. According to the prediction result of the new weight model, the new weight model is improved to obtain a flight delay prediction model; the flight delay prediction model is used for multi-airport flight delay prediction.

2. The large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network according to claim 1 is characterized in that: The preprocessing of the flight data and the indexed meteorological message data divided into time slices includes: By analogy with the spatiotemporal attention network model and input, the flight propagation dynamics model and flight status are defined; the number of flights in each flight status at each airport and the number of each aircraft type at each airport are counted; the flight status includes: normal flight, delayed flight and resumed flight; The missing values ​​in the flight data divided into time slices are cleaned by deletion; the indexed meteorological message data is cleaned by adjacent value filling; Merging the cleaned flight data divided into time slices and the indexed meteorological message data; Perform differential calculation on each flight status in the fused data to obtain the change in the number of flights in each time slice in each flight status; The quantity change of each time slice of each flight status is used as the regression task label; and the static flight data of each airport is integrated into each dynamic flight data.

3. The large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network according to claim 2 is characterized in that: The preprocessing of the flight data and the indexed meteorological message data divided into time slices includes: Encode all flight data that does not require the extraction of intermediate data features, including: A standardized method was used to encode numerical data; a word frequency-inverse document frequency method was used to process non-numerical data and then standardize it; Extract the uncoded data and optimize and organize the data according to the historical window, prediction window and sliding step size.

4. The large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network according to claim 3 is characterized in that: Extract the uncoded data; And optimize and organize the data according to the historical window, prediction window and sliding step size, including: According to the step length of each learning slide of the gated recurrent unit, the time dimension of each uncoded data is sorted into the total number of steps; The feature dimension is changed from the original dimension of each time slice to a feature set that contains all time slices at each step; The number of flight states in each step is extracted as the starting value for prediction by the flight propagation dynamics model; the number of changes in the number of delayed flights and the number of changes in the number of restored flights in each historical window are used as added features of the prediction results of the fully connected layer; the number of changes in the number of delayed flights and the number of changes in the number of restored flights in each prediction window are used as the true value of the loss function calculated with the predicted value.

5. The large-area flight delay propagation prediction method based on spatiotemporal attention deep learning network according to claim 1 is characterized in that: The new weight model comprises: Forward propagation training model and back propagation training model, where The total loss of the spatiotemporal attention network model is calculated in the forward propagation training model using the following formula: L=MSE(L pred )+α phy MSE(L phy )1 In the formula, L is the total loss of the spatiotemporal attention network model, MSE represents the loss function of the mean square error, and L pred is the result predicted by the fully connected layer, α phy is the weight of the prediction result of the dynamics model participating in the mean square error loss function, L phy Predict results for kinetic models; In the back propagation training model, the loss function obtained in the forward propagation training model is chain-derived by the following formula: In the formula, θ k is the network parameter of the kth layer, η is the learning rate, L is the total loss function of the network model, is the gradient operator.

6. The method for predicting large-area flight delay propagation based on spatiotemporal attention deep learning network according to claim 5, characterized in that: The new weight model comprises: The input layer is used to input the preprocessed encoded data into the spatiotemporal attention network model; The feature extraction layer is used to extract features through the forward propagation training model and the back propagation training model, and perform weighted optimization using the same loss function with different weights according to the prediction values ​​of different methods; The method separation layer is used to output the results of different prediction sequences according to different methods.

7. The method for predicting large-area flight delay propagation based on spatiotemporal attention deep learning network according to claim 6, characterized in that: The prediction results of the new weight model include: Predict the number of delayed flights at the airport in the future through the fully connected layer of the spatiotemporal attention network model output; The method separates the layer and uses another fully connected layer to predict the parameters in the flight propagation dynamics model.

8. The method for predicting large-area flight delay propagation based on spatiotemporal attention deep learning network according to claim 7, characterized in that: The new weight model is improved, including: In the graph attention network of the spatiotemporal attention network model, the parameters of each node are learned by the following formula: In the formula, e ij is the attention score of node i to node j; a is a learnable parameter; E i and E j are different embedding vectors for different nodes; W p is the weight pool; is the feature vector representing node i; is the feature vector of node j.

9. The method for predicting large-area flight delay propagation based on spatiotemporal attention deep learning network according to claim 8, characterized in that: The new weight model is improved, including: Combining the propagation dynamics to increase the dynamic state transition, including the following formula: In the formula, and is the change of the three states from the current moment to the next moment; α t is the propagation rate of delayed flights at time t; β t is the flight recovery rate at time t, S(t) is the number of normal flights at the previous moment; I(t) is the number of state S and state I at the previous moment; γR(t) is the flight state conversion process, γ represents the normal rate, which is the conversion rate from the restored flight state to the normal flight state.

10. A large-area flight delay propagation prediction system based on spatiotemporal attention deep learning network, characterized in that: include: Historical data acquisition module, used to obtain flight data and weather message data of multiple airports; A historical data analysis and processing module, used to analyze the time series autocorrelation of the flight data and the meteorological message data, and divide the flight data into time slices; and parsing and converting the meteorological message data to obtain meteorological message data with index; A spatiotemporal data sequence generation module is used to pre-process the flight data and the indexed meteorological message data divided into time slices by analogy with the input of the spatiotemporal attention network model to obtain a spatiotemporal data sequence; the spatiotemporal data sequence includes a three-dimensional data set of airports, time and features; A new weight model construction module is used to obtain a new weight model based on the spatiotemporal attention network model and the propagation dynamics method using the spatiotemporal data sequence as modeling data; the new weight model is used for single airport flight delay prediction; The flight delay prediction model construction module is used to improve the new weight model according to the prediction results of the new weight model to obtain a flight delay prediction model; the flight delay prediction model is used for multi-airport flight delay prediction.

Images