A traffic flow prediction method based on predefined spatio-temporal joint graph decoding

Abstract

The application discloses a traffic flow prediction method based on predefined space-time joint graph decoding and belongs to the technical field of public transport data analysis and prediction. The method comprises the following steps: collecting rail transit AFC gate passage data and train operation timetable, constructing a space-time enhanced OD matrix through abnormal value filtering and trip chain completion, and constructing a time dimension heterogeneous graph connection by using a dynamic time warping algorithm; designing a space-time physical coding module, introducing a time attention mechanism and a space attention mechanism, replacing a feature point multiplication form with feature splicing, reserving effective information obtained after passing through the two attention layers, and balancing the space-time feature fusion proportion in combination with a gate aggregation structure; constructing a space-time physical decoding module, modeling the space-time dependence of target flow on a sequence under the condition that the target flow is unknown, and reflecting comprehensive and dynamic space-time correlation by using a convolution predefined and adaptive space-time joint graph.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of public transportation data analysis and prediction, and relates to a traffic flow prediction method based on predefined spatio-temporal joint graph decoding. BACKGROUND

[0002] Traffic flow prediction is the basis of traffic management, Internet of Vehicles and many other applications in intelligent transportation systems (ITS); traffic data is a time series structure, and early statistical methods are used to predict traffic flow, such as seasonal autoregressive integrated moving average (SARIMA); these methods have been successfully applied to traffic prediction tasks; however, they face the challenge of being unable to extract inherent limitations of traffic data; traditional machine learning algorithms gradually appear in the flow prediction task, such as K-nearest neighbor (KNN), support vector regression (SVR) and hidden Markov model (HMM); to a certain extent, they break through the limitations of neural network-based methods; however, these methods rely on complex mathematical algorithms and prior knowledge, and can only obtain shallow features, and cannot extract deep and complex spatio-temporal correlations.

[0003] The rise of deep learning methods automatically processes high-dimensional data and extracts complex spatio-temporal correlations, providing new insights for completing the traffic flow prediction task; these methods include LSTM-BILSTM, DELA and ConvLSTM; however, traffic networks are a topological structure, and traffic data is subject to discrete distribution; recently, graph neural networks (GNN) have been widely used for flow prediction; although they have made significant progress in spatio-temporal correlation extraction and prediction accuracy, due to the need for physical structure information of traffic networks, spatial modeling is limited by methods that only consider static spatial dependencies or dynamic spatial correlations. SUMMARY

[0004] The purpose of the application is to provide a traffic flow prediction method based on predefined spatio-temporal joint graph decoding that can stably capture coupled information and maintain high accuracy.

[0005] Technical solution: the traffic flow prediction method based on predefined spatio-temporal joint graph decoding comprises the following operation steps:

[0006] (1) AFC punch card data collection and processing;

[0007] (2) Constructing a multi-modal convolution feature map;

[0008] (3) Constructing a spatio-temporal physical encoding module;

[0009] (4) Constructing a spatio-temporal semantic decoding module.

[0010] Further, the data collection in step (1) takes the rail transit route as the research object, records the passenger punch data into the database file with a half-hour interval.

[0011] Further, the data processing in step (1) is to filter out redundant values, abnormal values and missing values in the collected data, and finally obtain passenger OD data.

[0012] Further, the implementation process of step (1) is as follows: first, the AFC punch-in and punch-out behaviors at the subway station need to be performed twice, and a passenger OD trip is defined by this rule; second, the passenger OD trip is located by the card number, transaction station number and transaction time, and the records of missing corresponding information or long transaction time are deleted; finally, all OD data are counted with time intervals of 15 minutes, 30 minutes and 60 minutes to form passenger OD data.

[0013] Further, in step (2), the multi-modal convolution feature map is based on the real rail transit physical topology, and the OD data before and after dynamic time warping are decoupled into passenger flow signals in traffic lines and traffic stations to establish a multi-modal convolution feature map.

[0014] Further, the established multi-modal convolution feature map includes a physical topology feature map, a passenger flow similarity map and a correlation ratio map.

[0015] The physical topology feature map G p is directly constructed according to the subway physical topology structure of the selected area, and the implementation process is as follows: a physical connection matrix P∈R N×N is constructed, if the corresponding stations i and j are connected in the real world, then P(i,j) = 1, otherwise P(i,j) = 0, and after determining the physical connection matrix P, a row normalization operation is performed on P to obtain an edge weight matrix W p , the elements W p (i,j) at each position of the weight matrix are calculated as follows:

[0016]

[0017] In the formula, P(i,k) represents the connection relationship between stations i and j.

[0018] The passenger flow similarity map G S is defined by the similarity of passenger flow, and the implementation process is as follows: first, the DTW is used to evaluate the deviation between two time series, and the similarity score S(i,j) between stations i and j is calculated to obtain a similarity score matrix S∈R N×N , and the calculation formula of DTW is as follows:

[0019] S(i,j) = exp(-DTW(X i ,X j ))

[0020] where X i and X j represent the complete passenger flow time series of station i and j, and exp(·) represents the exponential function calculation formula;

[0021] Then, normalize S to calculate the edge weight matrix W s , whose calculation formula is as follows:

[0022]

[0023] where, when E s contains the edge connecting nodes i and k, L(E s , i, k) = 1, otherwise L(E s , i, k) = 0;

[0024] The correlation ratio graph G c is constructed from the origin and destination distribution of passengers, and the implementation process is as follows: by establishing a correlation ratio matrix C ∈ R N×N , whose calculation formula is as follows:

[0025]

[0026] where D(i,j) represents the total number of passengers from station i to j, and then is the total number of passengers from station i.

[0027] Further, the step (3) time-space physical coding module is composed of a time attention layer, a space attention layer, a multi-head graph convolution layer and an aggregation gate layer. The basic features before the multi-head graph convolution layer are spliced in the channel, and then the outflow features are processed by the aggregation gate layer.

[0028] Further, the time attention layer adopts a multi-head attention mechanism to calculate its correlation, and its expression is as follows:

[0029]

[0030] where, represents the correlation between the t i and t j time nodes, and τ represents the total length of the input time series; this correlation coefficient is obtained by the inner product of the Q and K vectors of the detection station v at time steps t i and t j :

[0031]

[0032] where, and denote the Q and K vectors in the nonlinear transformation function, respectively, and denotes the in-out degree of the monitoring station v;

[0033] The correlation coefficient is calculated After that, the head attention is calculated by the following formula:

[0034]

[0035] where, denotes the nonlinear transformation function for obtaining the V vector in the head attention, || denotes the channel concatenation operation, BN denotes the batch normalization, and HDT∈R τ×N×d denotes the output of the final temporal attention layer;

[0036] The spatial attention layer calculates the weight through the multi-head attention mechanism with physical information, and its expression is as follows:

[0037]

[0038] where, denotes the correlation between the node t i and t j , and v denotes the monitoring point of the input module;

[0039] The multi-head graph convolution layer focuses on the traffic flow in different subspaces of the monitoring point, and its expression is as follows:

[0040]

[0041] where, denotes the normalized adjacency matrix with self-connection, A vi denotes the node adjacency matrix with self-connection;

[0042] The output HDS∈R τ×N×d of the spatial attention and the output HSS∈R τ×N×d of the multi-head graph convolution are adaptively fused:

[0043] HST l = We HDS l +(1-W)e HSS l +HST l-1

[0044] W=σ(HDS l ⊙HSS l )

[0045] where σ denotes a Sigmoid activation function, W ∈ R T×N×d is a weight vector.

[0046] Further, the step (4) includes an input with a mask mechanism, an adaptive spatio-temporal joint graph convolution layer, a predefined spatio-temporal joint graph convolution layer, and an aggregation gate layer.

[0047] Further, the adaptive spatio-temporal joint graph convolution layer and the predefined spatio-temporal joint graph convolution layer are implemented as follows: first, a predefined spatio-temporal joint convolution graph is defined according to the spatio-temporal distance between nodes, and then an adaptive spatio-temporal joint graph is constructed.

[0048] The predefined spatio-temporal joint graph defines a spatial adjacency matrix based on a pair of road network distances:

[0049]

[0050] where dist(v i ,,v j ) denotes the shortest travel distance between the sites v i and v j , and σ denotes the travel distance standard deviation.

[0051] The constructed spatio-temporal joint graph preserves comprehensive spatio-temporal correlation by connecting different nodes between two time steps:

[0052]

[0053] where k denotes the time difference between two time steps, A i,t-k；j,t denotes the edge weight of the node v i at the time steps t-k and t.

[0054] The adaptive spatio-temporal joint graph convolution is as follows:

[0055]

[0056] where L t-k；t denotes the adaptive adjacency matrix between the steps t-k and t, denotes the updated feature at the time step t, and b adt denotes the learnable weight and bias parameters.

[0057] The mean squared error is used as the model loss function, and the mean absolute error and the mean absolute percentage error are used as the evaluation indicators of the final prediction results, and the mathematical expressions are as follows:

[0058]

[0059] Beneficial effects: compared with the prior art, the significant advantages of the present application are: 1, the present application decouples the original subway passenger flow diffusion and the inherent signal, constructs a multi-modal convolution feature map, proposes a dynamic decoupling and multi-modal reconstruction technology, and effectively reduces the spatial feature confusion error through signal separation; 2, the present application proposes a space-time attention collaborative coding module, which designs a cascade of time-space double attention layers in the coding module for complex space-time characteristics in traffic flow prediction, cooperates with the parallel feature extraction of multi-head graph convolution, realizes the precise modeling of space-time interaction mode of passenger flow in peak period; 3, the present application proposes a double graph convolution decoding architecture, which adopts adaptive and predefined double graph convolution joint decoding, maintains the physical topology constraint of rail transit, dynamically responds to real-time passenger flow fluctuation, increases the understanding and feature extraction ability of the correlation between sequences, and finally improves the prediction accuracy and robustness of the overall network. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is the network framework diagram of the present application;

[0061] Figure 2 is the calculation schematic diagram of the space attention layer in the present application;

[0062] Figure 3 is the calculation schematic diagram of the time attention layer in the present application;

[0063] Figure 4 is the aggregation gate structure diagram in the present application;

[0064] Figure 5 is the graph convolution process schematic diagram in the present application;

[0065] Figure 6 is the calculation schematic diagram of the space-time joint graph convolution layer in the present application;

[0066] Figure 7 is the traffic flow prediction result diagram in the present application. DETAILED DESCRIPTION

[0067] The specific technical solutions of the present application will be further described in detail below in combination with specific examples.

[0068] As shown in the figure, the operation steps of the traffic flow prediction method based on the pre-defined space-time joint graph decoding according to the present application are as follows:

[0069] Step (1), AFC card punching data collection and processing; the data collection takes the rail transit line in the area around the second ring expressway of a city as the research object, records the passenger card punching data to the database file with half an hour as the statistical interval;

[0070] The data processing is to filter out redundant values, abnormal values and missing values in the collected data, and finally obtain passenger OD data.

[0071] Step (2), constructing a multi-modal convolution feature map; the multi-modal convolution feature map is based on the real rail transit physical topology in the ring region of the second ring expressway of a city, and the OD data before and after dynamic time warping is decoupled into passenger flow signals of traffic lines and traffic stations to establish a multi-modal convolution feature map;

[0072] Step (3), designing a space-time physical coding module; the space-time physical coding module is composed of a time attention layer, a space attention layer, a multi-head graph convolution layer and an aggregation gate layer, and the features before the multi-head graph convolution layer are spliced in the channel, and then the features are processed by the aggregation gate layer;

[0073] Step (4), designing a space-time physical decoding module; the space-time physical decoding module is composed of an input with a mask mechanism, an adaptive space-time joint graph convolution layer, a pre-defined space-time joint graph convolution layer and an aggregation gate layer;

[0074] The implementation process of step (1) is as follows: AFC card punching data collection takes the rail transit line in the ring region of the second ring expressway of a city as the research object, takes half an hour as the statistical interval, and saves the passenger card punching records to the database file; the data processing is as follows: first, two AFC card punching behaviors are required for entering and leaving the subway station, so a passenger OD travel behavior can be defined according to this rule; second, the passenger OD travel can be located by the passenger card number, the station number where the transaction occurs and the transaction time; and third, records with missing corresponding information of entering and leaving the station or with too long transaction time of entering and leaving the station are deleted; finally, all OD data is counted with time intervals of 15 minutes, 30 minutes and 60 minutes to form passenger OD data.

[0075] The multi-modal convolution feature map in step (2) includes a physical topology feature map, a passenger flow similarity map and a correlation ratio map:

[0076] The physical topology feature map G p is directly constructed according to the subway physical topology structure of the selected region; first, a physical connection matrix P∈R N×N is constructed, if the corresponding stations i and j are connected in the real world, then P(i,j)=1, otherwise P(i,j)=0; after determining the physical connection matrix P, the row normalization operation is performed on P to obtain the edge weight matrix W p The element W p (i,j) at each position of the weight matrix is calculated as follows:

[0077]

[0078] where P(i, k) represents the connection relationship between stations i and j;

[0079] Passenger flow similarity graph G S The similarity of passenger flow is defined; first, the deviation between two time series is evaluated using dynamic time warping (DTW), and the similarity score S(i, j) between stations i and j is calculated, and the similarity score matrix S ∈ R N×N is obtained, and the calculation formula of DTW is as follows:

[0080] S(i, j) = exp(-DTW(X i , X j ))

[0081] where X i and X j represent the complete passenger flow time series of stations i and j, and exp(·) represents the calculation formula of the exponential function; then, S is normalized to calculate the edge weight matrix W s , and the calculation formula is as follows:

[0082]

[0083] where E s contains the edge connecting nodes i and k, L(E s , i, k) = 1, otherwise L(E s , i, k) = 0;

[0084] Correlation ratio graph G c is constructed from the origin and destination distribution of passengers; first, a correlation ratio matrix C ∈ R N×N is established, and the calculation formula is as follows:

[0085]

[0086] where D(i, j) represents the total number of passengers from station i to j, and then is the total number of passengers from station i.

[0087] In step (3), the time attention acts on the past, present and future domains of each monitoring point in the traffic network; the internal relationship of the monitoring data will change between different time steps, and this internal relationship is called dynamic time correlation; the time attention layer uses a multi-head attention mechanism to calculate its correlation, and the expression is as follows:

[0088]

[0089] where represents the node t iwith t j The correlation between time nodes, τ represents the total length of the input time series, this correlation coefficient can be calculated by the detection site v at time step t i with t j The inner product of Q, K vectors at t

[0090]

[0091] In the formula, and Q and K vectors in the nonlinear transformation function respectively, <**> represents the inner product operator, is the out-degree of the monitoring site v. The correlation coefficient After that, the head attention Can be calculated by the following formula:

[0092]

[0093] In the formula, Indicates the nonlinear transformation function of V vector in the head attention, || indicates the channel splicing operation, BN indicates the batch normalization, HDT∈R τ×N×d is the output of the final time attention layer;

[0094] The spatial attention acts on a monitoring point in the traffic network, is affected by other monitoring points, and its influence weight changes with time, which defines the dynamic spatial correlation; considering the influence of the inherent physical information of the traffic network on the calculation, the spatial attention layer designs a multi-head attention mechanism with physical information to calculate the weight, its expression is as follows:

[0095]

[0096] In the formula, Indicates the correlation between node t i and t j The correlation between time nodes, v represents the monitoring point of the input module.

[0097] Similar to the time attention layer, the spatial correlation of traffic flow is also affected by the out-degree of the monitoring site, and the out-degree of the monitoring site Is embedded as an exogenous information input into the spatial attention layer, and its correlation coefficient is calculated as follows:

[0098]

[0099] At the same time, the correlation between monitoring sites is affected by the shortest path distance and the shortest path, which is defined as and K = |E|, for each pair of reachable stations, find the shortest path SP, and then calculate the average point product value of the eigenvalue of the edge on the path and the weight matrix as the bias term and incorporate it into the attention module, whose calculation formula is as follows:

[0100]

[0101] In the formula, And The direct distance can be calculated by the Dijkstra algorithm, Indicates the weight matrix; similarly, the correlation coefficient of the spatial attention layer can be represented as:

[0102]

[0103] In the formula, Indicates a nonlinear transformation, Indicates the spatial correlation coefficient of station v; the input of the spatial attention layer is HDT∈R τ×N×d , and the output is HDS∈R τ×N×d .

[0104] The multi-head graph convolution considers the connection relationship between nodes in the traffic network, i.e., the origin and destination of traffic flow at the monitoring station, which is the static spatial correlation of traffic flow; a multi-head graph convolution is used to focus on the traffic flow in different subspaces of the monitoring station, and its expression is as follows:

[0105]

[0106] In the formula, Indicates a normalized adjacency matrix with self-connection, A vi Indicates a node adjacency matrix with self-connection;

[0107] In order to obtain the final spatio-temporal correlation HST∈R τ×N×d between traffic flows, the output HDS∈R τ×N×d of the spatial attention and the output HSS∈R τ×N×d of the multi-head graph convolution are adaptively fused using a gating fusion method, and its expression is as follows:

[0108] HST l = We HDS l +(1-W)e HSS l +HST l-1

[0109] W = σ(HDS l ⊙HSS l )

[0110] In the formula, σ represents the Sigmoid activation function, W∈R T×N×drepresents a weight vector.

[0111] The implementation process of the adaptive spatio-temporal joint graph convolution layer and the predefined spatio-temporal joint graph convolution layer in the step (4) is as follows: first, a predefined spatio-temporal joint convolution graph is defined according to the spatio-temporal distance between nodes, however, the predefined graph may not reflect the potential correlation between nodes, and an adaptive spatio-temporal joint graph is proposed to construct two types of graphs to represent the comprehensive and dynamic spatio-temporal correlation between nodes;

[0112] The predefined spatio-temporal joint graph defines a spatial adjacency matrix based on the distance of the pair of road networks:

[0113]

[0114] In the formula, dist(v i, ,v j, ) represents the shortest travel distance between the stations v i, and v j, , and σ represents the travel distance standard deviation;

[0115] The spatio-temporal joint graph is constructed by connecting different nodes between two time steps to retain comprehensive spatio-temporal correlation:

[0116]

[0117] In the formula, k represents the time difference between two time steps, A i,t-k；j,t represents the edge weight of the node v i at the time steps t-k and t, and when k=0, the spatio-temporal joint graph is restored to the ordinary spatial adjacency matrix;

[0118] Before defining the adaptive spatio-temporal joint graph, two learnable matrices U and B are set for the very close nodes, U represents the potential attributes of the nodes, B represents the interaction mode of each attribute, and UBU T can represent the connection between nodes; thus, it can be defined as:

[0119]

[0120] In the formula, U T represents the spatio-temporal embedding of the node at the time step t, and the softmax function is used to eliminate the weights less than the threshold δ adt , represents the spatial connection, which changes with the time step and can be deformed as:

[0121]

[0122] In the formula, L t-k；tdenotes the normalized adaptive spatio-temporal joint matrix, the standard graph convolution only models the spatial graph correlation, in order to match the spatio-temporal joint graph, the spatio-temporal joint graph convolution is proposed; the calculation formula of the traditional graph convolution is as follows:

[0123]

[0124] wherein X and Z denote the input and output, denotes the normalized adjacency matrix, wherein A is the adjacency matrix with self-loop, and D is the degree matrix;

[0125] The convolution mode on the predefined spatio-temporal joint graph is as follows:

[0126]

[0127] wherein, denotes the normalized predefined spatio-temporal joint matrix, denotes the aggregation of the information in the time and space, denotes the aggregation of the information in the space field only; the information of K time steps can be integrated, and the above formula can be reduced to:

[0128]

[0129] In the case of a directed graph:

[0130]

[0131] wherein, and denote the normalized predefined spatio-temporal joint matrix in two directions respectively, X t-k and X t denote the input graph signal at time points t-k and t, is the updated feature at the time step t, and b pdf are the learnable weight and bias parameters;

[0132] Similarly, the adaptive spatio-temporal joint graph convolution is as follows:

[0133]

[0134] wherein, denotes the adaptive adjacency matrix between the steps t-k and t, denotes the updated feature at the time step t, and b adt are the learnable weight and bias parameters;

[0135] The mean square error (MSE) is used as the model loss function, and the mean absolute error (MAE) and the mean absolute percentage error are used as the final prediction result evaluation indexes, and the mathematical expressions are as follows:

[0136]

Claims

1. A traffic flow prediction method based on decoding of a predefined spatio-temporal joint graph, characterized in that, The operation steps are as follows: (1) AFC clock-in data collection and processing; (2) Constructing a multi-modal convolution feature map; The construction process is: based on the real rail transit physical topology, the OD data before and after dynamic time warping is used as the content value, the passenger flow signal of the traffic line and the traffic station is decoupled, and finally the multi-modal convolution feature map is established; The established multi-modal convolution feature map includes physical topology feature map, passenger flow similarity map and correlation ratio map; Wherein, the physical topology feature map According to the selected area subway physical topology structure is directly constructed, its implementation process is as follows: by constructing the physical connection matrix If the corresponding station in the real world With Communicate, then Not connected then Determine the physical connection matrix After Line normalization operation gets the edge weight matrix The element of each position on the weight matrix The calculation formula is as follows: ; in which, represents a communication relationship of a station with ; Passenger flow similarity graph The similarity score between two stations is calculated by the similarity of passenger flow, and the process is as follows: first, the DTW is used to evaluate the deviation between two time series, and the similarity score between stations and is calculated , and the similarity score matrix is obtained , and the calculation formula of DTW is as follows: ; in the formula, and represent the complete passenger flow time series of stations and , and represents the calculation formula of the exponential function Then, the edge weight matrix W is calculated by normalizing the matrix W = (wij) with the formula ; where wij= 1 if the edge between nodes i and j contains a connection and , otherwise wij= 0 ; Correlation matrix is constructed from the distribution of the passengers' origins and destinations, and the implementation process is as follows: a correlation matrix is established, and the calculation formula is as follows: ; in the formula, represents the total number of passengers from the station to , and is the total number of passengers from the station ; (3) Constructing a space-time physical coding module; The space-time physical coding module is composed of a time attention layer, a space attention layer, a multi-head graph convolution layer and an aggregation gate layer, which splices each basic feature before the multi-head graph convolution layer in the channel, and then processes the outflow feature by the aggregation gate layer; Wherein, the time attention layer adopts a multi-head attention mechanism to calculate its correlation, and its formula is as follows: ; wherein, represents the correlation between the node and the time node, represents the total length of the input time sequence; this correlation coefficient is obtained by the inner product of the vectors at the time steps , and , : ; wherein, and respectively represent the and vectors in the nonlinear transformation function, represents the inner product operator, represents the in-degree and out-degree of the monitoring station ; correlation coefficient is calculated after which the head attention is obtained from wherein, denotes the head attention a non-linear transformation function of the vector, denotes a channel concatenation operation, BN denotes batch normalization, denotes the output of the final temporal attention layer; The spatial attention layer calculates the weight through the multi-head attention mechanism with physical information, and the formula is as follows: ; wherein, represents the correlation between the node and the time node, represents the monitoring point of the input module; The multi-head graph convolutional layer focuses on the traffic flow of different subspaces of the monitoring points, and its formula is as follows: ; wherein, denotes a normalized adjacency matrix with self-connection, denotes a node adjacency matrix with self-connection; Output of spatial attention using gated fusion method Output of multi-head graph convolution Adaptive fusion, which is formulated as follows: wherein, denotes a Sigmoid activation function, is a weight vector; (4) Constructing a space-time object decoding module; The space-time object decoding module includes input with mask mechanism, adaptive space-time joint graph convolution layer, predefined space-time joint graph convolution layer and aggregation gate layer; The implementation process of the adaptive space-time joint graph convolution layer and the predefined space-time joint graph convolution layer is as follows: first, the predefined space-time joint convolution graph is defined according to the space-time distance between nodes, and then the adaptive space-time joint graph is constructed. wherein the predefined spatio-temporal joint graph defines a spatial adjacency matrix based on pairwise road network distances, which is formulated as follows: ; wherein, denotes the shortest travel distance between sites and , denotes the travel distance standard deviation; The spatio-temporal joint graph is constructed by connecting different nodes between two time steps, preserving comprehensive spatio-temporal correlations, and is formulated as follows: wherein, denotes the time difference between two time steps, denotes the time step and the edge weight of the node at the time step The adaptive spatio-temporal joint graph convolution is in the following manner: wherein denotes the step size and an adaptive adjacency matrix between denotes the updated feature at time step , and denotes the learnable weight and bias parameters; The mean square error is used as the model loss function, and the mean absolute error and the mean absolute percentage error are used as the final prediction result evaluation indexes, and the formulas are as follows: .

2. The traffic flow prediction method based on decoding of predefined spatio-temporal joint graph according to claim 1, characterized in that, In step (1), the data collection uses rail transit lines as the research object, and records the passenger clock-in data to the database file with a half-hour statistical interval.

3. The traffic flow prediction method based on decoding of predefined spatio-temporal joint graph according to claim 1, characterized in that: In step (1), the data processing is: filtering out the redundant values, abnormal values and missing values in the collected data, so as to obtain passenger OD data.

4. The traffic flow prediction method based on decoding of pre-defined spatio-temporal joint graph according to claim 1, characterized in that: The implementation process of step (1) is: first, two AFC clock-in behaviors are required for entering and leaving the subway station, and a passenger OD travel behavior is defined by this rule; second, the passenger card number, transaction station number and transaction time are used to locate an OD trip; finally, all OD data is counted with a time interval of 15 minutes, 30 minutes and 60 minutes, and passenger OD data is formed.

Images