National highway transportation capacity distribution prediction method based on spatiotemporal graph convolutional network

Abstract

The present invention discloses a method for predicting the distribution of national highway transport capacity based on a spatiotemporal graph convolutional network, including: obtaining and preprocessing the historical data of highway transport capacity in cities across the country; performing singular spectrum decomposition on the historical data of highway transport capacity to extract the trend, volatility and noise signals of highway transport capacity time series data; determining the causal matrix of highway transport capacity between cities by calculating transfer entropy; constructing an adjacency matrix using the causal matrix, and using the spatiotemporal graph convolutional network to learn and predict the trend data and volatility data of highway transport capacity; finally, fusing the trend prediction value and the volatility prediction value to obtain the predicted value of highway transport capacity in each city. The method of the present invention makes full use of the trend characteristics of highway transport capacity changes and takes into account the spatial causal characteristics of transport capacity distribution, has a good prediction ability for the distribution of highway transport capacity in cities across the country, and provides an effective and stable solution for the time series prediction of national transportation transport capacity distribution.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic capacity distribution prediction, and in particular relates to a national highway transportation capacity distribution prediction method based on a spatiotemporal graph convolutional network. Background Art

[0002] Realizing highway transport capacity forecasts for different cities across the country will help to effectively utilize and allocate highway transport capacity. However, highway transport capacity in different cities is affected by many factors such as local economy, geographical location, freight demand, and special holidays, and the impact of different factors on urban highway transport capacity is difficult to capture. Highway transport capacity mainly includes passenger capacity and freight capacity. Since passenger transport has a fixed schedule, route, and passenger capacity, highway passenger capacity is relatively stable; while freight transport has greater variability due to the diversity of routes and changes in customer demand, so the highway transport capacity in this application refers to highway freight capacity.

[0003] At present, there are few studies specifically targeting the prediction of highway capacity distribution in cities across the country, while there have been many studies at home and abroad on issues such as taxi passenger demand and freight demand prediction. The methods can be roughly divided into two categories: one is based on mathematical statistics, and the other is based on machine learning. The method based on mathematical statistics generally uses mathematical modeling, sets corresponding parameters, and uses mathematical expressions to calculate the predicted value. The advantage of this type of method is that the calculation is simple, but at the same time, this type of method is difficult to capture the deep-level characteristics of capacity distribution. Machine learning methods such as Bayesian networks, BP neural networks, CNN, RNN, LSTM, and spatiotemporal graph convolutional networks can learn the weights of each feature through training with a large amount of data, thereby realizing the prediction of freight demand.

[0004] Patent application publication number CN112819213A discloses a highway freight volume prediction method and system based on a deep learning network. It uses the flow data of highway toll stations to count the truck flow at the entrance station, the freight volume at the exit station, etc., and adopts a hybrid graph neural network combining a graph convolutional neural network and a long short-term memory network to fit the relationship between the input data such as the truck flow at the entrance station and the freight volume at the target exit station in the future. At the same time, considering the spatiotemporal correlation of the traffic volume at the entrance and exit of highway toll stations, a spatiotemporal attention mechanism is added to the deep learning network to achieve a high-precision prediction of the freight volume of all highway exit stations in the region.

[0005] Patent application publication number CN107392549A discloses an online highway transport capacity trading platform based on a fleet store and its construction method. The demander and the recipient can automatically adapt the capacity through a self-learning highway transport capacity matching model to achieve effective utilization of the capacity, thereby improving transportation efficiency and reducing costs at both ends.

[0006] For the highway capacity distribution prediction, since the highway capacity of cities across the country is affected by various nonlinear spatiotemporal factors, time, day, week, month, holidays, highway toll price fluctuations, economic development status of each city, social structure of each city, population size of each city, road network connection of each city, weather, policy, etc. are the main factors affecting the changes in highway capacity distribution values. The influence mechanism of each of these factors is difficult to capture accurately, and the changes in highway capacity values ​​caused by the existence of these influencing factors may exist in the form of noise. It is considered more meaningful to capture the trend of highway capacity changes and grasp the changing trend of highway capacity values ​​as a whole.

[0007] Therefore, accurately learning the trend changes of highway capacity will help to achieve effective and stable predictions of highway capacity in cities across the country. The present invention proposes to use the historical data of national highway capacity distribution, through singular spectrum analysis and spatiotemporal convolutional networks, to explore the trend and volatility of highway capacity and combine the temporal and spatial characteristics of capacity to achieve the prediction of national highway capacity distribution. In addition, the commonly used spatiotemporal graph convolutional network uses the Euclidean distance between nodes or the road network connection as the adjacency matrix, but in the scenario of national highway capacity distribution prediction, the capacity connection relationship between cities has a certain causality, so it is more reasonable to use a causal association matrix to represent the capacity relationship connection between cities. Summary of the invention

[0008] In order to solve the shortcomings of the above-mentioned existing technologies, the present invention proposes a national highway transportation capacity distribution prediction method based on a spatiotemporal graph convolutional network. The specific technical solution is as follows:

[0009] The national highway transportation capacity distribution prediction method based on spatiotemporal graph convolutional network includes the following steps:

[0010] S1: Data acquisition and preprocessing; the specific process is:

[0011] S1-1: Number each city in the country, and the highway capacity value of each city is X = [X1, X2, X3, ..., X c ,…,X G ], where X1, X2, X3, …, X c ,…,X G The highway transport capacity history values ​​of cities 1, 2, 3, …, c, …, G are obtained respectively. G is the total number of cities in the country. The daily highway transport capacity value X of city c within N+1 days before the prediction time is obtained. c =[x c1 ,x c2 ,x c3 ,…,x cn ,…,x c(N+1) ], where x c1 ,x c2 ,x c3 ,…,xcn ,…,x c(N+1) are the highway transport capacity values ​​of city c N+1 days, N days, N-1 days, …, N+1-n days, …, and 1 day before the prediction time;

[0012] S1-2: Clean the highway capacity data and fill in the missing data through linear interpolation;

[0013] S1-3: Calculate the difference of highway capacity:

[0014] Δx cn =x c(n+1) -x cn

[0015] ΔX c =[Δx c1 ,Δx c2 ,Δx c3 ,…,Δx cn ,…,Δx cN ]

[0016] Where Δx c1 ,Δx c2 ,Δx c3 ,…,Δx cn ,…,Δx cN are the highway capacity change values ​​of city c N days before the prediction time, N-1 days before, N-2 days before, ..., N+1-n days before, ..., 1 day before, ΔX c is the historical change value of highway transport capacity in city c;

[0017] S1-4: ΔX c Perform 0-1 normalization to obtain the normalized historical change value of highway capacity ΔZ c =[Δz c1 ,Δz c2 ,Δz c3 ,…,Δz cn ,…,Δz cN ], where Δz c1 ,Δz c2 ,Δz c3 ,…,Δz cn ,…,Δz cN are respectively the normalized historical change values ​​of highway transport capacity of city c N days before the prediction time, N-1 days before, N-2 days before, …, N+1-n days before, …, and 1 day before;

[0018] S1-5: Repeat steps S1-1 to S1-4 to obtain the historical change value of highway transport capacity of each city ΔZ=[ΔZ1, ΔZ2, ΔZ3, …, ΔZ c ,…,ΔZ G], where ΔZ1, ΔZ2, ΔZ3, …, ΔZ c ,…,ΔZ G is the normalized historical change value of highway transport capacity of cities 1, 2, 3, …, c, …, G;

[0019] S2: Determine the parameters of singular spectrum analysis, i.e., the size of the time sliding window, perform singular spectrum decomposition on the normalized historical change value of highway transport capacity of each city, and extract the trend signal, volatility signal and noise of the time series signal from the decomposed components;

[0020] S3: Calculate the transfer entropy between different cities and construct the causal matrix of highway transport capacity among cities across the country;

[0021] S4: Use the causal matrix of national highway capacity to construct an adjacency matrix, take the trend signal and volatility signal of each city's highway capacity value as the value of each node at different time slices in the spatiotemporal graph convolutional network, and build a spatiotemporal graph convolutional network;

[0022] S5: The trend signal prediction value and the volatility signal prediction value obtained after the spatiotemporal graph convolutional network are integrated to obtain the final urban highway capacity distribution prediction value.

[0023] Furthermore, in step S2, for the normalized historical change value of highway transport capacity of a city, the time window length of singular spectrum decomposition is selected as L,

[0024] S2-1: ΔZ c Convert to trajectory matrix:

[0025] Among them, X tr,c is the trajectory matrix of city c, Δz cL ,Δz c(L+1) ,Δz c(N-L+1) …,Δz cN They are the highway transport capacity change values ​​of city c N+1-L days, NL days, L days, …, 1 day before the prediction time;

[0026] S2-2: Covariance matrix of the trajectory matrix Perform eigenvalue decomposition to obtain eigenvalues ​​r1>r2>…r m …>r L ≥0 and the eigenvalues ​​corresponding to the eigenvectors U=U1,U2,…,U m ,…,U L , the singular spectrum is decomposed into:

[0027]

[0028] Where m is X tr,cThe eigenvalue number of , m = 1, 2, ..., L;

[0029] S2-3: Calculate the information entropy of each eigenvalue:

[0030]

[0031] Where, ΔE m is information entropy, j is a natural number from 1 to m, r j For X tr,c The characteristic value of

[0032] When the information entropy ΔE m <ε1, ε1 is the set threshold, and the m component and subsequent components are regarded as noise N c ; The signal component corresponding to the first eigenvalue represents the trend signal T of the time series signal c , the signal components corresponding to the second to the m-1th eigenvalues ​​represent the volatility signal W of the time series signal c ;

[0033] S2-4: Repeat steps S2-1 to S2-3 to obtain the trend signal, volatility signal, and noise signal of each city, and finally obtain the trend signal of the historical change value of highway transport capacity of cities across the country T = [T1, T2, T3, ..., T c ,...,T G ], volatility signal W=[W1,W2,W3,...,W c ,...,W G ] and noise, where T1, T2, T3, ..., T c ,...,T G and W1,W2,W3,...,W c ,...,W G They are the trend signals and volatility signals of the historical change values ​​of highway transport capacity of cities 1, 2, 3, …, c, …, G respectively.

[0034] Furthermore, the specific process of step S3 is as follows:

[0035] S3-1: Transfer entropy from the historical change value of highway transport capacity in city q to the historical change value of highway transport capacity in city p for:

[0036]

[0037] Where Δx pn ,Δx qn are the highway transport capacity changes of city p and city q N+1-n days before the prediction time;

[0038] Represents the change in highway capacity of city p N+1-n days before the prediction time and k days before that;

[0039] Represents the change in highway transport capacity of city q N+1-n days before the prediction time and l days before that;

[0040] is the joint probability, is the conditional probability;

[0041] The transfer entropy from the historical change value of highway transport capacity in city p to the historical change value of highway transport capacity in city q is: Calculation method and The difference between the two is:

[0042]

[0043] S3-2: Determine whether there is a causal relationship between the highway transport capacity of the two cities:

[0044] When TE ΔXp,ΔXq >ε2, ε2 is the set threshold, indicating that the highway capacity value of city q is the cause of the highway capacity value of city p, and city q is determined as the cause node of city p, and the number of cause nodes of each city is d;

[0045] S3-3: The final causal matrix TE is obtained as:

[0046]

[0047] The degree matrix D is:

[0048]

[0049] The causal matrix is ​​transformed into:

[0050] Among them, TE 1,1 ,TE 1,2 ,…,TE 1,G ,TE G,G They respectively represent the causal relationship between the historical change value of highway capacity of city 1 and itself, the causal relationship between the historical change value of highway capacity of city 2 and the historical change value of highway capacity of city 1, …, the causal relationship between the historical change value of highway capacity of city G and city 1, …, the causal relationship between the historical change value of highway capacity of city G and city G;

[0051] When TE ΔXp,ΔXq >ε2, that is, city q is the cause node of city p, let TE p,q =1; when TE ΔXp,ΔXq ≤ε2, that is, city q is not the cause node of city p, let TE p,q =0;

[0052] d1,d2,…,d G are the number of cause nodes of cities 1, 2, …, G respectively. A is the symmetric matrix of the causal matrix, that is, the adjacency matrix, which represents the connection relationship between the highway transport capacity of cities.

[0053] Furthermore, the spatiotemporal graph convolutional network in step S4 is expressed as The network input V includes the trend signal T and volatility signal W of the historical values ​​of highway transport capacity of G cities in the past N days, and finally outputs the predicted value of highway transport capacity of cities across the country for the next day;

[0054] The spatiotemporal graph convolution network includes at least one spatiotemporal graph convolution block, each of which includes two temporal convolution blocks and one spatial convolution block. The calculation process of each spatiotemporal graph convolution block is as follows:

[0055]

[0056] Among them, v l and v l+1 are the highway capacity values ​​of cities across the country on the lth spatiotemporal graph convolution block and the l+1th spatiotemporal graph convolution block, respectively. is the temporal convolution kernel in the lth spatiotemporal graph convolution block, Θ l is the spectral graph convolution kernel in the lth spatiotemporal graph convolution block, * τ is the temporal convolution operator, * g is the graph convolution operator, v 1 =V;

[0057] ①The calculation process of the spatial convolution block is:

[0058]

[0059] Among them, L0 = DA, θ is the spectral convolution kernel, θ k is the Chebyshev polynomial, θ0 is the first term of the Chebyshev polynomial, θ1 is the second term of the Chebyshev polynomial, L0 is the Laplace matrix, and D is the degree matrix;

[0060] ②The calculation process of the temporal convolution block is:

[0061] Γ* τ v=P⊙σ(Q)

[0062]

[0063] Among them, Г is the time convolution kernel, P and Q are the values ​​after convolution of v with the time convolution kernel Г, and ⊙ is the Hadamard product.

[0064] Furthermore, in step S5, the highway transport capacity of cities across the country is predicted by using a spatiotemporal graph convolutional network:

[0065]

[0066]

[0067] in, is the predicted value of highway transport capacity of cities across the country at time t+1, v t-N+1 is the actual value of highway transport capacity in cities across the country N days before the prediction time, v t is the actual value of highway transport capacity in cities across the country at time t, They are the predicted values ​​of the trend signal and volatility signal of highway transport capacity in cities across the country at the prediction time. The two are combined to obtain the final predicted value of urban highway transport capacity distribution.

[0068] The beneficial effects of the present invention are:

[0069] 1. The present invention provides a method for predicting the distribution of national highway transport capacity based on singular spectrum decomposition and spatiotemporal graph convolutional network. Singular spectrum decomposition is used to extract the trend and volatility of historical data of national highway truck capacity and remove the influence of noise. Transfer entropy is used to determine the causal relationship of changes in highway truck capacity between cities to construct a spatial graph structure. Spatiotemporal graph convolutional network is used to predict the trend signal and fluctuation signal of urban highway truck capacity, thereby realizing accurate prediction of future highway capacity data of each city.

[0070] 2. The present invention focuses on the trend of changes in highway truck capacity and the spatial causal correlation characteristics. Starting from the daily changes in highway truck transportation volume in cities across the country, the overall trend of capacity data changes is explored, and the adjacency matrix and degree matrix are constructed using the transfer entropy matrix to construct a spatiotemporal graph convolutional network, and a study on the national highway freight capacity distribution prediction method is carried out. The method of the present invention has important practical significance for overall mastering the highway capacity data of each city, improving the operating efficiency of national highways, stabilizing collaborative work, and promoting work scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention will be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0072] Figure 1 is a flow chart of the method of the present invention;

[0073] Figure 2 It is a flow chart of calculating the cause-effect relationship of highway transport capacity changes in various cities of the present invention;

[0074] Figure 3 It is a structural diagram of national highway transport capacity prediction of the present invention;

[0075] Figure 4 is a graph convolutional network of the present invention;

[0076] Figure 5 It is the signal after the singular spectrum decomposition of the present invention. DETAILED DESCRIPTION

[0077] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0078] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0079] The present invention provides a national highway transport capacity distribution prediction method based on a spatiotemporal graph convolutional network. First, according to the historical data of national highway capacity distribution, the data is preprocessed, the daily change value of the capacity is calculated and normalized; then the change value of the highway capacity of each city is subjected to singular spectrum decomposition, so as to extract the trend, volatility and noise of the highway capacity change in cities across the country; the highway capacity transfer entropy matrix between cities is calculated as the adjacency matrix to construct a graph structure; then the time series data after noise removal is used as the input of the spatiotemporal graph convolutional network, the spatiotemporal causal characteristics of the highway capacity distribution are fully explored, and the trend signal and the volatility signal are predicted; finally, the predicted trend signal and the volatility signal are fused to obtain the final national highway capacity distribution prediction value.

[0080] like Figure 1 and Figure 3 As shown in FIG, the national highway transport capacity distribution prediction method based on the spatiotemporal graph convolutional network includes the following steps:

[0081] S1: Data acquisition and preprocessing; the specific process is:

[0082] S1-1: Number each city in the country, and the highway capacity value of each city is X = [X1, X2, X3, ..., X c ,…,X G ], where X1, X2, X3, …, X c,…,X G The highway transport capacity history values ​​of cities 1, 2, 3, …, c, …, G are obtained respectively. G is the total number of cities in the country. The daily highway transport capacity value X of city c within N+1 days before the prediction time is obtained. c =[x c1 ,x c2 ,x c3 ,…,x cn ,…,x c(N+1) ], where x c1 ,x c2 ,x c3 ,…,x cn ,…,x c(N+1) are the highway transport capacity values ​​of city c N+1 days, N days, N-1 days, …, N+1-n days, …, and 1 day before the prediction time;

[0083] S1-2: Clean the highway capacity data and fill in the missing data through linear interpolation;

[0084] S1-3: Calculate the difference of highway capacity:

[0085] Δx cn =x c(n+1) -x cn

[0086] ΔX c =[Δx c1 ,Δx c2 ,Δx c3 ,…,Δx cn ,…,Δx cN ]

[0087] Where Δx c1 ,Δx c2 ,Δx c3 ,…,Δx cn ,…,Δx cN are the highway capacity change values ​​of city c N days before the prediction time, N-1 days before, N-2 days before, ..., N+1-n days before, ..., 1 day before, ΔX c is the historical change value of highway transport capacity in city c;

[0088] S1-4: ΔX c Perform 0-1 normalization to obtain the normalized historical change value of highway capacity ΔZ c =[Δz c1 ,Δz c2 ,Δz c3 ,…,Δz cn ,…,Δz cN ], where Δz c1 ,Δzc2 ,Δz c3 ,…,Δz cn ,…,Δz cN are respectively the normalized historical change values ​​of highway transport capacity of city c N days before the prediction time, N-1 days before, N-2 days before, …, N+1-n days before, …, and 1 day before;

[0089] S1-5: Repeat steps S1-1 to S1-4 to obtain the historical change value of highway transport capacity of each city ΔZ=[ΔZ1, ΔZ2, ΔZ3, …, ΔZ c ,…,ΔZ G ], where ΔZ1, ΔZ2, ΔZ3, …, ΔZ c ,…,ΔZ G is the normalized historical change value of highway transport capacity of cities 1, 2, 3, …, c, …, G;

[0090] S2: Determine the parameters of singular spectrum analysis, i.e., the size of the time sliding window, perform singular spectrum decomposition on the normalized historical change value of highway transport capacity of each city, and extract the trend signal, volatility signal and noise of the time series signal from the decomposed components;

[0091] In step S2, for the normalized historical change value of highway transport capacity of a city, the time window length of singular spectrum decomposition is selected as L.

[0092] S2-1: ΔZ c Convert to trajectory matrix:

[0093] Among them, X tr,c is the trajectory matrix of city c, Δz cL ,Δz c(L+1) ,Δz c(N-L+1) …,Δz cN They are the highway transport capacity change values ​​of city c N+1-L days, NL days, L days, …, 1 day before the prediction time;

[0094] S2-2: Covariance matrix of the trajectory matrix Perform eigenvalue decomposition to obtain eigenvalues ​​r1>r2>…r m …>r L ≥0 and the eigenvalues ​​corresponding to the eigenvectors U=U1,U2,…,U m ,…,U L , the singular spectrum is decomposed into:

[0095]

[0096] Where m is X tr,c The eigenvalue number of , m = 1, 2, ..., L;

[0097] S2-3: Calculate the information entropy of each eigenvalue:

[0098]

[0099] Where, ΔE m is information entropy, j is a natural number from 1 to m, r j For X tr,c The characteristic value of

[0100] When the information entropy ΔE m <ε1, ε1 is the set threshold, and the m component and subsequent components are regarded as noise N c ; The signal component corresponding to the first eigenvalue represents the trend signal T of the time series signal c , the signal components corresponding to the second to the m-1th eigenvalues ​​represent the volatility signal W of the time series signal c ;

[0101] Figure 5 The components of a city's truck capacity time series after singular spectrum analysis are decomposed. The upper left is the trend signal, and the components from the right to the back are the components from the volatility signal to the noise. The specific dividing line between the volatility signal and the noise is determined by step S2-3.

[0102] S2-4: Repeat steps S2-1 to S2-3 to obtain the trend signal, volatility signal, and noise signal of each city, and finally obtain the trend signal of the historical change value of highway transport capacity of cities across the country T = [T1, T2, T3, ..., T c ,...,T G ], volatility signal W=[W1,W2,W3,...,W c ,...,W G ] and noise, where T1, T2, T3, ..., T c ,...,T G and W1,W2,W3,...,W c ,...,W G They are the trend signals and volatility signals of the historical change values ​​of highway transport capacity of cities 1, 2, 3, …, c, …, G respectively.

[0103] S3: Calculate the transfer entropy between different cities and construct the causal matrix of highway transport capacity among cities across the country; Figure 2 As shown, the specific process is:

[0104] S3-1: Transfer entropy from the historical change value of highway transport capacity in city q to the historical change value of highway transport capacity in city p for:

[0105]

[0106] Where Δx pn ,Δx qn are the highway transport capacity changes of city p and city q N+1-n days before the prediction time;

[0107] Represents the change in highway capacity of city p N+1-n days before the prediction time and k days before that;

[0108] Represents the change in highway transport capacity of city q N+1-n days before the prediction time and l days before that;

[0109] is the joint probability, is the conditional probability;

[0110] The transfer entropy from the historical change value of highway transport capacity in city p to the historical change value of highway transport capacity in city q is: Calculation method and The difference between the two is:

[0111]

[0112] S3-2: Determine whether there is a causal relationship between the highway transport capacity of the two cities:

[0113] When TE ΔXp,ΔXq >ε2, ε2 is the set threshold, indicating that the highway capacity value of city q is the cause of the highway capacity value of city p, and city q is determined as the cause node of city p, and the number of cause nodes of each city is d;

[0114] S3-3: The final causal matrix TE is obtained as:

[0115]

[0116] The degree matrix D is:

[0117]

[0118] The causal matrix is ​​transformed into:

[0119] Among them, TE 1,1 ,TE 1,2 ,…,TE 1,G ,TE G,GThey respectively represent the causal relationship between the historical change value of highway capacity of city 1 and itself, the causal relationship between the historical change value of highway capacity of city 2 and the historical change value of highway capacity of city 1, …, the causal relationship between the historical change value of highway capacity of city G and city 1, …, the causal relationship between the historical change value of highway capacity of city G and city G;

[0120] When TE ΔXp,ΔXq >ε2, that is, city q is the cause node of city p, let TE p,q =1; when TE ΔXp,ΔXq ≤ε2, that is, city q is not the cause node of city p, let TE p,q =0;

[0121] d1,d2,…,d G are the number of cause nodes of cities 1, 2, …, G respectively. A is the symmetric matrix of the causal matrix, that is, the adjacency matrix, which represents the connection relationship between the highway transport capacity of cities.

[0122] S4: Use the causal matrix of national highway capacity to construct an adjacency matrix, take the trend signal and volatility signal of each city's highway capacity value as the value of each node at different time slices in the spatiotemporal graph convolutional network, and build a spatiotemporal graph convolutional network;

[0123] like Figure 4 As shown, the spatiotemporal graph convolutional network is represented as The network input V includes the trend signal T and volatility signal W of the historical values ​​of highway transport capacity of G cities in the past N days, and finally outputs the predicted value of highway transport capacity of cities across the country for the next day;

[0124] The spatiotemporal graph convolution network includes at least one spatiotemporal graph convolution block. Each spatiotemporal graph convolution block contains two temporal convolution blocks and one spatial convolution block. The calculation process of each spatiotemporal graph convolution block is:

[0125]

[0126] Among them, v l and v l+1 are the highway capacity values ​​of cities across the country on the lth spatiotemporal graph convolution block and the l+1th spatiotemporal graph convolution block, respectively. is the temporal convolution kernel in the lth spatiotemporal graph convolution block, Θ l is the spectral graph convolution kernel in the lth spatiotemporal graph convolution block, * τ is the temporal convolution operator, * g is the graph convolution operator, v 1 =V;

[0127] ①The calculation process of the spatial convolution block is:

[0128]

[0129] Among them, L0 = DA, θ is the spectral convolution kernel, θ k is the Chebyshev polynomial, θ0 is the first term of the Chebyshev polynomial, θ1 is the second term of the Chebyshev polynomial, L0 is the Laplace matrix, and D is the degree matrix;

[0130] ②The calculation process of the temporal convolution block is:

[0131] Γ* τ v=P⊙σ(Q)

[0132]

[0133] Among them, Г is the time convolution kernel, P and Q are the values ​​after convolution of v with the time convolution kernel Г, and ⊙ is the Hadamard product.

[0134] S5: The trend signal prediction value and the volatility signal prediction value obtained after the spatiotemporal graph convolutional network are integrated to obtain the final urban highway capacity distribution prediction value.

[0135] The highway transport capacity of cities across the country is predicted through the spatiotemporal graph convolutional network:

[0136]

[0137]

[0138] in, is the predicted value of highway transport capacity of cities across the country at time t+1, v t-N+1 is the actual value of highway transport capacity in cities across the country N days before the prediction time, v t is the actual value of highway transport capacity in cities across the country at time t, They are the predicted values ​​of the trend signal and volatility signal of highway transport capacity in cities across the country at the prediction time. The two are combined to obtain the final predicted value of urban highway transport capacity distribution.

[0139] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the distribution of national highway transportation capacity based on spatiotemporal graph convolutional networks, characterized in that: The following steps are involved: S1: Data acquisition and preprocessing; the specific process is: S1-1: Number each city in the country, and the highway capacity value of each city is X = [X1, X2, X3, ..., X c ,…,X G ], where X1, X2, X3, …, X c ,…,X G The highway transport capacity history values ​​of cities 1, 2, 3, …, c, …, G are obtained respectively. G is the total number of cities in the country. The daily highway transport capacity value X of city c within N+1 days before the prediction time is obtained. c =[x c1 ,x c2 ,x c3 ,…,x cn ,…,x c(N+1) ], where x c1 ,x c2 ,x c3 ,…,x cn ,…,x c(N+1) are the highway transport capacity values ​​of city c N+1 days, N days, N-1 days, …, N+1-n days, …, and 1 day before the prediction time; S1-2: Clean the highway capacity data and fill in the missing data through linear interpolation; S1-3: Calculate the difference of highway capacity: Δx cn =x c(n+1) -x cn ΔX c =[Δx c1 ,Δx c2 ,Δx c3 ,…,Δx cn ,…,Δx cN ] Where Δx c1 ,Δx c2 ,Δx c3 ,…,Δx cn ,…,Δx cN are the highway capacity change values ​​of city c N days before the prediction time, N-1 days before, N-2 days before, ..., N+1-n days before, ..., 1 day before, ΔX c is the historical change value of highway transport capacity in city c; S1-4: ΔX c Perform 0-1 normalization to obtain the normalized historical change value of highway capacity ΔZ c =[Δz c1 ,Δz c2 ,Δz c3 ,…,Δz cn ,…,Δz cN ], where Δz c1 ,Δz c2 ,Δz c3 ,…,Δz cn ,…,Δz cN are respectively the normalized historical change values ​​of highway transport capacity of city c N days before the prediction time, N-1 days before, N-2 days before, …, N+1-n days before, …, and 1 day before; S1-5: Repeat steps S1-1 to S1-4 to obtain the historical change value of highway transport capacity of each city ΔZ=[ΔZ1, ΔZ2, ΔZ3, …, ΔZ c ,…,ΔZ G ], where ΔZ1, ΔZ2, ΔZ3, …, ΔZ c ,…,ΔZ G is the normalized historical change value of highway transport capacity of cities 1, 2, 3, …, c, …, G; S2: Determine the parameters of singular spectrum analysis, i.e., the size of the time sliding window, perform singular spectrum decomposition on the normalized historical change value of highway transport capacity of each city, and extract the trend signal, volatility signal and noise of the time series signal from the decomposed components; S3: Calculate the transfer entropy between different cities and construct the causal matrix of highway transport capacity among cities across the country; S4: Use the causal matrix of national highway capacity to construct an adjacency matrix, take the trend signal and volatility signal of each city's highway capacity value as the value of each node at different time slices in the spatiotemporal graph convolutional network, and build a spatiotemporal graph convolutional network; S5: The trend signal prediction value and the volatility signal prediction value obtained after the spatiotemporal graph convolutional network are integrated to obtain the final urban highway capacity distribution prediction value; In step S2, for the normalized historical change value of highway transport capacity of a city, the time window length of singular spectrum decomposition is selected as L. S2-1: ΔZ c Convert to trajectory matrix: Among them, X tr,c is the trajectory matrix of city c, Δz cL ,Δz c(L+1) ,Δz c(N -L+1)…,Δz cN They are the highway transport capacity change values ​​of city c N+1-L days, NL days, L days, …, 1 day before the prediction time; S2-2: Covariance matrix of the trajectory matrix Perform eigenvalue decomposition to obtain eigenvalues ​​r1>r2>…r m …>r L ≥0 and the eigenvalues ​​corresponding to the eigenvectors U=U1,U2,…,U m ,…,U L , the singular spectrum is decomposed into: Where m is X tr,c The eigenvalue number of , m = 1, 2, ..., L; S2-3: Calculate the information entropy of each eigenvalue: Where, ΔE m is information entropy, j is a natural number from 1 to m, r j For X tr,c The characteristic value of When the information entropy ΔE m <ε1, ε1 is the set threshold, and the m component and subsequent components are regarded as noise N c ; The signal component corresponding to the first eigenvalue represents the trend signal T of the time series signal c , the signal components corresponding to the second to the m-1th eigenvalues ​​represent the volatility signal W of the time series signal c ; S2-4: Repeat steps S2-1 to S2-3 to obtain the trend signal, volatility signal, and noise signal of each city, and finally obtain the trend signal of the historical change value of highway transport capacity of cities across the country T = [T1, T2, T3, ..., T c ,...,T G ], volatility signal W=[W1,W2,W3,...,W c ,...,W G ] and noise, where T1, T2, T3, ..., T c ,...,T G and W1,W2,W3,...,W c ,...,W G They are the trend signals and volatility signals of the historical change values ​​of highway transport capacity of cities 1, 2, 3, …, c, …, G respectively; The specific process of step S3 is: S3-1: Transfer entropy of the historical change value of highway transport capacity from city q to the historical change value of highway transport capacity from city p for: Where Δx pn ,Δx qn are the highway transport capacity changes of city p and city q N+1-n days before the prediction time; Represents the change in highway capacity of city p N+1-n days before the prediction time and k days before that; Represents the change in highway transport capacity of city q N+1-n days before the prediction time and l days before that; is the joint probability, is the conditional probability; The transfer entropy from the historical change value of highway transport capacity in city p to the historical change value of highway transport capacity in city q is TE ΔXp→ΔXq , calculated in the same way as TE ΔXq→ΔXp The difference between the two is: yourself ΔXp,ΔXq =TE ΔXq→ΔXp yourself ΔXp→ΔXq S3-2: Determine whether there is a causal relationship between the highway transport capacity of the two cities: When TE ΔXp,ΔXq >ε2, ε2 is the set threshold, indicating that the highway capacity value of city q is the cause of the highway capacity value of city p, and city q is determined as the cause node of city p, and the number of cause nodes of each city is d; S3-3: The final causal matrix TE is obtained as: The degree matrix D is: The causal matrix is ​​transformed into: Among them, TE 1,1 ,TE 1,2 ,…,TE 1,G ,TE G,G They respectively represent the causal relationship between the historical change value of highway capacity of city 1 and itself, the causal relationship between the historical change value of highway capacity of city 2 and the historical change value of highway capacity of city 1, …, the causal relationship between the historical change value of highway capacity of city G and city 1, …, the causal relationship between the historical change value of highway capacity of city G and city G; When TE ΔXp,ΔXq >ε2, that is, city q is the cause node of city p, let TE p,q =1; when TE ΔXp,ΔXq ≤ε2, that is, city q is not the cause node of city p, let TE p,q =0; d1,d2,…,d G are the number of cause nodes of cities 1, 2, …, G respectively, and A is the symmetric matrix of the cause-effect matrix, i.e., the adjacency matrix, which represents the connection relationship of highway transport capacity between cities; The spatiotemporal graph convolutional network in step S4 is expressed as The network input V includes the trend signal T and volatility signal W of the historical values ​​of highway transport capacity of G cities in the past N days, and finally outputs the predicted value of highway transport capacity of cities across the country for the next day; The spatiotemporal graph convolution network includes at least one spatiotemporal graph convolution block, each of which includes two temporal convolution blocks and one spatial convolution block. The calculation process of each spatiotemporal graph convolution block is as follows: Among them, v l and v l+1 are the highway capacity values ​​of cities across the country on the lth spatiotemporal graph convolution block and the l+1th spatiotemporal graph convolution block, respectively. is the temporal convolution kernel in the lth spatiotemporal graph convolution block, Θ l is the spectral graph convolution kernel in the lth spatiotemporal graph convolution block, * τ is the temporal convolution operator, * g is the graph convolution operator, v 1 =V; ①The calculation process of the spatial convolution block is: Among them, L0 = DA, θ is the spectral convolution kernel, θ k is the Chebyshev polynomial, θ0 is the first term of the Chebyshev polynomial, θ1 is the second term of the Chebyshev polynomial, L0 is the Laplace matrix, and D is the degree matrix; ②The calculation process of the temporal convolution block is: Γ* τ v=P⊙σ(Q) Among them, Г is the time convolution kernel, P and Q are the values ​​after v is convolved with the time convolution kernel Г, and ⊙ is the Hadamard product; In step S5, the highway transport capacity of cities across the country is predicted by using a spatiotemporal graph convolutional network: in, is the predicted value of highway transport capacity of cities across the country at time t+1, v t-N+1 is the actual value of highway transport capacity in cities across the country N days before the prediction time, v t is the actual value of highway transport capacity in cities across the country at time t, They are the predicted values ​​of the trend signal and volatility signal of highway transport capacity in cities across the country at the prediction time. The two are combined to obtain the final predicted value of urban highway transport capacity distribution.

Images