Urban rail transit passenger flow short-term OD passenger flow prediction method

Abstract

The invention relates to a short-term OD passenger flow prediction method for urban rail transit passenger flow, and the method comprises the steps: 1) constructing a label of a scene feature based on an influence factor set which influences the short-term OD passenger flow prediction of multiple scenes, and embedding the scene feature into a time sequence of historical OD passenger flow; 2) constructing a subway network space-time diagram according to the time sequence of the historical OD passenger flow embedded with the scene features, and constructing a sparse graph convolutional network for capturing spatial features; 3) according to the time sequence of the historical OD passenger flow embedded with the scene features, constructing a cross attention mechanism for capturing the time features; and 4) based on the sparse graph convolutional network and a cross attention mechanism, constructing a Transform-based multi-scene short-term passenger flow prediction combination model, and outputting a final predicted OD passenger flow matrix.

Description

Technical Field

[0001] The present invention relates to the technical field of passenger flow prediction, and in particular to a method for predicting short-term OD passenger flow in urban rail transit. Background Art

[0002] With the development of informatization in various industries, the rail transit systems in most cities have entered the network operation stage. The ever-expanding road network scale and the increasing passenger flow have made the operation management and transportation organization strategy formulation more complicated. To form a reasonable and scientific transportation organization, it is necessary to conduct in-depth research on the dynamic changes of passenger flow in daily network operations, and accurately predict the road network OD passenger flow in future operation periods, provide real-time passenger flow data support for dynamic operation management and control, and assist the operation department in making network operation decisions, which is of great significance for optimizing the allocation of public transportation resources. The huge scale of the road network makes the passenger flow distribution complex and dynamic, so how to timely and accurately predict the road network OD passenger flow has become a difficult problem to be overcome.

[0003] In view of the difficulty in obtaining OD passenger flow, large-scale manual surveys are often used to count OD passenger flow, which is not only extremely time-consuming and labor-intensive, but also has large deviations in the final statistical data. When using traditional prediction models to predict OD passenger flow, model constraints make it difficult to fully utilize the nonlinear relationship between OD data, and machine learning methods for predicting OD passenger flow are mostly based on shallow network structures, which are limited in extracting complex nonlinear spatiotemporal relationships between OD passenger flows on large road networks.

[0004] At the same time, the various factors that affect the changes in OD passenger flow are complex and changeable. Collecting the corresponding influencing factors while obtaining OD passenger flow data also brings challenges to the study of OD passenger flow, resulting in relatively few related prediction studies. The short-term OD passenger flow matrix in multiple scenarios still faces the problems of large data dimensions and severe data sparsity, which causes the integrated scene features to be diluted and the learning results are not significant.

[0005] At present, the AFC card swiping system is widely deployed in urban rail transit in various cities in my country, which can collect historical passenger flow data more comprehensively and accurately. The rise of deep learning provides technical support for solving the problem of low prediction accuracy of traditional OD passenger flow prediction models. Summary of the invention

[0006] The present invention relates to the field of urban rail transit transportation organization theory and technology, and specifically provides a short-term OD passenger flow prediction method applicable to multiple scenarios. According to historical OD passenger flow, spatial distribution of stations on the road network, weather, special events and time factors, a feature fusion prediction model under multiple scenarios is constructed based on a deep learning framework to obtain the predicted OD passenger flow, and provide OD passenger flow information for operation organization management and path planning design.

[0007] A short-term OD passenger flow prediction method for urban rail transit passenger flow, comprising:

[0008] Step 1) Based on the influencing factor set that affects the short-term OD passenger flow prediction in multiple scenarios, construct the labels of the scene features, and embed the scene features into the time series of the historical OD passenger flow;

[0009] Step 2) constructing a spatiotemporal graph of the subway network based on the time series of historical OD passenger flows embedded with scene features, and building a sparse graph convolutional network for capturing spatial features;

[0010] Step 3) Based on the time series of historical OD passenger flows embedded with scene features, a cross-attention mechanism is constructed to capture temporal features;

[0011] Step 4) Based on the sparse graph convolutional network and cross-attention mechanism, a Transformer-based multi-scenario short-term passenger flow prediction combination model is constructed to output the final predicted OD passenger flow matrix.

[0012] Optionally, the influencing factor set affecting the multi-scenario short-term OD passenger flow prediction includes the following influencing factors: historical OD passenger flow, spatial distribution of road network stations, weather, special events and time

[0013] Optionally, the historical OD passenger flow is an OD passenger flow matrix generated using passenger card swiping data in the AFC system;

[0014] The spatial distribution of the road network stations is represented as a completely directed graph, where stations are represented by nodes, and connections between stations are represented by edges formed by connecting nodes, and each edge contains historical OD passenger flow and road network spatial information;

[0015] The weather characteristics include the degree of sunshine or rain and the air quality;

[0016] The weather conditions include sunny, cloudy, overcast, light rain, moderate rain, heavy rain and rainstorm; the air quality includes excellent, good, lightly polluted, moderately polluted, heavily polluted and severely polluted.

[0017] The special events mentioned mainly include urban rail transit operation events and large-scale events;

[0018] The time includes time, working day and non-working day information.

[0019] Optionally, the subway network space-time graph includes the following variables: nodes, edges, space-time points, days, stations, and series;

[0020] The subway network spatiotemporal graph includes the following sets: an OD matrix of an input sequence, a feature set, a research time range, a station set, a set of edges with stations as points, and the number of matrices in a matrix sequence.

[0021] The subway network structure includes the following structures:

[0022] The complete directed graph G(V,E,A) represents the graph network at time t;

[0023] Where V t is the OD passenger flow value matrix of the subway network at time t, V t ∈R n×n , n is the number of stations, E is the set of edges with stations as points in the subway network, representing the connection between stations, A∈R n×n is the adjacency matrix of graph G, where A 0 Indicates the physical connection relationship between sites. 1 Indicates the travel time between stations.

[0024] Optionally, the method of capturing spatial features using a sparse convolutional network is:

[0025] The sparse graph convolution SGCN and graph convolution GCN are combined to capture the spatial characteristics of the subway network, and upsampling convolution and downsampling convolution are added to the sparse graph convolution module.

[0026] Optionally, the step 3) includes the following steps:

[0027] Step 1) Set the three Transformer latent spaces of ST, IS, and IT with dimensions m, and transform the original OD passenger flow data X t ∈R 1×n×n Mapped to different latent spaces, we get The calculation formula is:

[0028]

[0029] In formula (1) It represents the OD passenger flow statistics from station i to station j that have completed the trip at the t-th time interval on the d-th day; where i,j∈[1,n], n represents the number of stations in the road network.

[0030]

[0031] In formula (2), Spare∈{ST,IS,IT} represents different latent spaces, ReLU(·) represents the activation function, and Conv 1×1 (·) represents a convolution operation with a convolution kernel of 1×1.

[0032] Step 2) Divide the input matrix sequence of embedded features into three vector spaces, and use differentiated feature extraction operations in different spaces to capture the features of the input matrix sequence in three time dimensions: long-term, short-term, and medium-term:

[0033] The ST space is the long-term feature space. On this model branch:

[0034] α-ST=Cross Attention{SGCN ST [Pool ST (input ST )]} (3)

[0035] IS space is a short-term feature space. On this model branch:

[0036] α-IS=Cross Attention{SGCN IS [Pool IS (Slipt(input IS ))]} (4)

[0037] IT space is a medium- to long-term feature space. On this model branch:

[0038] α-IT=Cross Attention{SGCN IT [Pool IT (GCN IT (input IT ))]}))] (5)

[0039] In the formula, SGCN represents sparse graph convolutional network, GCN represents graph convolutional network, Pool represents pooling operation, Slipt represents separation operation in formula (3), and both the network and operation layers contain weight parameters that can be used for training;

[0040] Step 3) Mapped to the output Chead through a Cross-Attention mechanism t , respectively mapping the corresponding query space on ST, IS, and IT space Keyspace With value space V t i For the Cross-Attention mechanism, the query space of different spaces is obtained by mapping the information of the space itself, and the key space and value space on different spaces are obtained by mapping the information of other spaces. The calculation formula is:

[0041] α-IT=Cross Attention{SGCN IT [Pool IT(GCN IT (input IT ))]}))] (5)

[0042]

[0043]

[0044] In formula (6), (χ∈{q,k,v},i∈{1,2,…,h}) correspond to the feature extraction functions in the ST, IS, and IT branch spaces respectively; Q t STi Represents the query vector of the i-th element in the ST space at time t, represented by the function φ i q according to Calculated; Represents the query vector of the i-th element in the IS space at time t, which is represented by the function according to Calculated; Represents the query vector of the i-th element in the IT space at time t, represented by the function according to Calculated; Respectively represent the bond vectors in ST, IS, and IT spaces; V t STi ,V t ISi ,V t ITi Respectively represent the value vectors in ST, IS, and IT spaces; W in formula (8) m Represents a trainable parameter, d = m / h, where d represents days, m represents dimension, and h represents the length of the input sequence.

[0045] Optionally, the step 4) includes the following steps: combining the sparse graph convolutional network and the cross attention mechanism, and performing residual connection and normalization processing to generate a combined prediction model based on the Transformer network Encoder-Decoder architecture.

[0046] Optionally, the model combination process is as follows:

[0047] Step 1: Set the three Transformer latent spaces of ST, IS, and IT to be m in dimension, and transform the original OD passenger flow data X t ∈R 1×n×n Mapped to different latent spaces, we get

[0048] Step 2: Mapped to the output Chead through a Cross-Attention mechanism t .

[0049] Step 3: After residual connection and normalization operations, the generalization of the model network is enhanced and the effectiveness of training deep learning networks is improved.

[0050] Step 4: The output Y is obtained through a feed-forward neural network FCs consisting of two fully connected layers t n×n .

[0051] Step 5: Repeat step 4 to obtain OD passenger flow related information Y t od .

[0052] Step 6: Let X t =Y t od , repeat step 2 to step 5 k+2 times to deeply extract feature information.

[0053] Optionally, the residual connection and normalization operations are:

[0054]

[0055] In formula (9), Norm(·) represents the normalization function, and GCN(·) represents the graph convolution function.

[0056] The output Y is obtained by a feed-forward neural network FCs consisting of two fully connected layers t n×n .

[0057]

[0058] Compared with the prior art, the present invention has the following advantages and effects:

[0059] Based on the spatiotemporal attention mechanism model, the cross attention mechanism is combined with spatiotemporal feature extraction and expanded into the spatiotemporal attention mechanism, which improves the correlation of spatiotemporal features and the quality of feature extraction.

[0060] Embedding external environmental features into the OD historical passenger flow sequence and combining the prediction model with the Transformer framework can enhance the information interaction between multi-scenario external environmental features and the corresponding OD passenger flow, and meet the prediction of multi-scenario short-term OD passenger flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of the present invention.

[0062] Figure 2This is a schematic diagram of the subway network structure.

[0063] Figure 3 This is a schematic diagram of the SGCN module.

[0064] Figure 4 It is a schematic diagram of feature capture of the input matrix sequence in three time dimensions: long-term, short-term, and medium- to long-term. DETAILED DESCRIPTION

[0065] In order to make the purpose and technical solution of the present invention more clear, the present invention is further described in detail. The experimental methods described in the following examples are all conventional methods unless otherwise specified: if no specific techniques or conditions are specified in the examples, they are carried out according to the techniques or conditions described in the literature in the field or according to the product instructions; the reagents and materials, unless otherwise specified, can be obtained from commercial channels.

[0066] Example 1

[0067] A short-term OD passenger flow prediction method for large-scale networked urban rail transit based on spatiotemporal multimodality includes:

[0068] Step 1) determine the influencing factor set that affects the short-term OD passenger flow prediction in multiple scenarios;

[0069] Calibrate the factors that affect short-term OD passenger flow prediction: including historical OD passenger flow, spatial distribution of network stations, weather, special events and time.

[0070] The historical OD passenger flow is an OD passenger flow matrix at a certain historical interval generated by using the passenger card swiping data in the AFC system;

[0071] The spatial distribution of the road network stations can be represented as a completely directed graph, where stations are represented by nodes, and the connections between stations are represented by edges connected by nodes, and each edge contains historical OD passenger flow and road network spatial information;

[0072] The weather characteristics mainly examine the degree of sunshine and rain, which are divided into 7 degrees: sunny, cloudy, overcast, light rain, moderate rain, heavy rain, and rainstorm. The air quality index AQI is divided into 6 levels, namely excellent, good, light pollution, moderate pollution, heavy pollution, and severe pollution;

[0073] The special events mentioned mainly include urban rail transit operation events and large-scale events;

[0074] The time includes time, working days and non-working days.

[0075] Step 2) Based on the influencing factor set, construct a label of the scene feature, and embed the scene feature into the time series of the historical OD passenger flow;

[0076] 2. Construct a label of scene features based on the influencing factor set and embed it into the time series of historical OD passenger flow, which is characterized by including label-related parameters:

[0077] The tag-related parameters include the following indexes:

[0078]

[0079] The above embedding into the time series of historical OD passenger flow, the scene at time t includes the following feature vectors:

[0080]

[0081] in Respectively represent the category corresponding to each feature in the table.

[0082] Step 3) construct a spatiotemporal graph of the subway network and use a sparse graph convolutional network to capture spatial features;

[0083] The subway network spatiotemporal diagram includes the following variables:

[0084]

[0085] The subway network space-time diagram includes the following sets:

[0086]

[0087] The subway network structure includes the following structures:

[0088] Complete directed graph G(V,E,A): represents the graph network at time t, where V t is the OD passenger flow value matrix of the subway network at time t, V t ∈R n×n , n is the number of stations, E is the set of edges with stations as points in the subway network, representing the connection between stations, A∈R n×n is the adjacency matrix of graph G, where A 0 Indicates the physical connection relationship between sites. 1 Indicates the travel time between stations.

[0089] The spatial features captured by the sparse convolutional network are:

[0090] The sparse graph convolution (SGCN) and GCN are mainly used to capture the spatial characteristics of the subway network, and upsampling convolution (UPConv) and downsampling convolution (DownConv) are added to the SGCN module.

[0091] Step 4) Use the cross-attention mechanism to capture temporal features;

[0092] The spatiotemporal sequence of the historical OD passenger flow is used to generate the query in the attention mechanism in a nonlinear function manner. key Sum The convolution and pooling operations, SGCN and GCN networks are integrated into the Cross-Attention mechanism to capture temporal features. The features are:

[0093] (1) Set the three Transformer latent spaces of ST, IS, and IT to be m in dimension, and transform the original OD passenger flow data X t ∈R 1×n×n Mapped to different latent spaces, we get X t ST ∈R m×n×n , The calculation formula is:

[0094]

[0095] In formula (1) It represents the OD passenger flow statistics from station i to station j that have completed the trip at the tth time interval on the dth day. Where i,j∈[1,n], n represents the number of stations in the network.

[0096]

[0097] In formula (2), Spare∈{ST,IS,IT} represents different latent spaces, ReLU(·) represents the activation function, and Conv 1×1 (·) represents a convolution operation with a convolution kernel of 1×1.

[0098] (2) The input matrix sequence of embedded features is divided into three vector spaces, and differentiated feature extraction operations are used in different spaces to capture the features of the input matrix sequence in the three time dimensions of long-term, short-term, and medium-term:

[0099] 1) ST space is a long-term feature space. On this model branch:

[0100] α-ST=Cross Attention{SGCN ST [Pool ST (input ST )]} (3)

[0101] 2) IS space is a short-term feature space. On this model branch:

[0102] α-IS=Cross Attention{SGCN IS [Pool IS(Slipt(input IS ))]} (4)

[0103] 3) IT space is a medium- and long-term feature space. On this model branch:

[0104] α-IT=Cross Attention{SGCN IT [Pool IT (GCN IT (input IT ))]}))] (5)

[0105] In the formula, SGCN represents sparse graph convolutional network, GCN represents graph convolutional network, Pool represents pooling operation, Slipt represents separation operation in formula (3), and both the network and operation layers contain weight parameters that can be used for training.

[0106] (3) Mapped to the output Chead through a Cross-Attention mechanism t , respectively mapping the corresponding query space on ST, IS, and IT space Keyspace With value space V t i For the Cross-Attention mechanism, the query space of different spaces is obtained by mapping the information of the space itself, and the key space and value space on different spaces are obtained by mapping the information of other spaces. The calculation formula is:

[0107] α-IT=Cross Attention{SGCN IT [Pool IT (GCN IT (input IT ))]}))] (5)

[0108]

[0109] In formula (6), (χ∈{q,k,v},i∈{1,2,…,h}) correspond to the feature extraction functions in the ST, IS, and IT branch spaces respectively; Q t STi Represents the query vector of the i-th element in the ST space at time t, represented by the function φ i q According to X t ST Calculated; Q t ISi Represents the query vector of the i-th element in the IS space at time t, which is represented by the function according to Calculated; Represents the query vector of the i-th element in the IT space at time t, represented by the function according to Calculated; Respectively represent the bond vectors in ST, IS, and IT spaces; V t STi ,V t ISi ,V t ITi Respectively represent the value vectors in ST, IS, and IT spaces; W in formula (8) m Represents a trainable parameter, d=m / h, where d represents day; m represents dimension; h represents the length of the input sequence, and the value is 5 in the present invention.

[0110] Step 5) Build a Transformer-based multi-scenario short-term passenger flow prediction combination model and output the final predicted OD passenger flow matrix. The specific steps include:

[0111] The above cross-attention mechanism is combined with the above feedforward neural network, and residual connection and normalization are performed to generate a combined prediction model based on the Transformer network Encoder-Decoder architecture:

[0112] The residual connection and normalization operation are:

[0113] Y t m =Norm[X t +GCN(Chead t )] (9)

[0114] In formula (9), Norm(·) represents the normalization function, and GCN(·) represents the graph convolution function.

[0115] The output Y is obtained by a feed-forward neural network FCs consisting of two fully connected layers t n×n .

[0116] Y t n×n =FCs(Y t m ) (10)

[0117] The model combination process is as follows:

[0118] Step 1: Set the three Transformer latent spaces of ST, IS, and IT to be m in dimension, and transform the original OD passenger flow data X t ∈R1×n×n Mapped to different latent spaces, we get

[0119] Step 2: Mapped to the output Chead through a Cross-Attention mechanism t .

[0120] Step 3: After residual connection and normalization operations, the generalization of the model network is enhanced and the effectiveness of training deep learning networks is improved.

[0121] Step 4: The output Y is obtained through a feed-forward neural network FCs consisting of two fully connected layers t n×n .

[0122] Step 5: Repeat step 4 to obtain OD passenger flow related information Y t od .

[0123] Step 6: Let X t =Y t od , repeat step 2 to step 5 k+2 times to deeply extract feature information.

[0124] In addition, it should be noted that the above are only preferred embodiments of the present invention and are not limited to the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein by equivalents. 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 short-term OD passenger flow prediction method for urban rail transit passenger flow, characterized in that: include: Step 1) Based on the influencing factor set that affects the short-term OD passenger flow prediction in multiple scenarios, construct the labels of the scene features, and embed the scene features into the time series of the historical OD passenger flow; Step 2) constructing a spatiotemporal graph of the subway network based on the time series of historical OD passenger flows embedded with scene features, and building a sparse graph convolutional network for capturing spatial features; Step 3) Based on the time series of historical OD passenger flows embedded with scene features, a cross-attention mechanism is constructed to capture temporal features; Step 4) Based on the sparse graph convolutional network and cross-attention mechanism, a Transformer-based multi-scenario short-term passenger flow prediction combination model is constructed to output the final predicted OD passenger flow matrix.

2. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The influencing factor set affecting the multi-scenario short-term OD passenger flow prediction includes the following influencing factors: historical OD passenger flow, spatial distribution of road network stations, weather, special events and time.

3. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The historical OD passenger flow is an OD passenger flow matrix generated using passenger card swiping data in the AFC system; The spatial distribution of the road network stations is represented as a completely directed graph, where stations are represented by nodes, and connections between stations are represented by edges formed by connecting nodes, and each edge contains historical OD passenger flow and road network spatial information; The weather characteristics include the degree of sunshine or rain and the air quality; The weather conditions include sunny, cloudy, overcast, light rain, moderate rain, heavy rain and rainstorm; the air quality includes excellent, good, lightly polluted, moderately polluted, heavily polluted and severely polluted. The special events mentioned mainly include urban rail transit operation events and large-scale events; The time includes time, working day and non-working day information.

4. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The subway network space-time graph includes the following variables: nodes, edges, space-time points, days, stations, and series; The subway network spatiotemporal graph includes the following sets: an OD matrix of an input sequence, a feature set, a research time range, a station set, a set of edges with stations as points, and the number of matrices in a matrix sequence. The subway network structure includes the following structures: Complete directed graph G(V t ,E,A), represents the graph network at time t; Where V t is the OD passenger flow value matrix of the subway network at time t, V t ∈R n×n , n is the number of stations, E is the set of edges with stations as points in the subway network, representing the connection between stations, A∈R n×n is the adjacency matrix of the complete directed graph G.

5. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The spatial features captured by the sparse convolutional network are: The sparse graph convolution SGCN and graph convolution GCN are combined to capture the spatial characteristics of the subway network, and upsampling convolution and downsampling convolution are added to the sparse graph convolution module.

6. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The step three) comprises the following steps: Step 1) Set the three Transformer latent spaces of ST, IS, and IT with dimensions m, and transform the original OD passenger flow data X t ∈R 1×n×n Mapped to different latent spaces, we get The calculation formula is: In formula (1) It represents the OD passenger flow statistics from station i to station j that have completed the trip at the t-th time interval on the d-th day; Where i,j∈[1,n], n represents the number of stations in the network. In formula (2), Spare∈{ST,IS,IT} represents different latent spaces, ReLU(·) represents the activation function, and Conv 1×1 (·) represents a convolution operation with a convolution kernel of 1×1. Step 2) Divide the input matrix sequence of embedded features into three vector spaces, and use differentiated feature extraction operations in different spaces to capture the features of the input matrix sequence in three time dimensions: long-term, short-term, and medium-term: The ST space is the long-term feature space. On this model branch: α-ST=Cross Attention{SGCN ST [Pool ST (input ST )]} (3) IS space is a short-term feature space. On this model branch: α-IS=Cross Attention{SGCN IS [Pool IS (Slipt(input IS ))]} (4) IT space is a medium- to long-term feature space. On this model branch: α-IT=Cross Attention{SGCN IT [Pool IT (GCN IT (input IT ))]}))] (5) In the formula, SGCN represents sparse graph convolutional network, GCN represents graph convolutional network, Pool represents pooling operation, Slipt represents separation operation in formula (3), and both the network and operation layers contain weight parameters that can be used for training; Step 3) Mapped to the output Chead through a Cross-Attention mechanism t , respectively mapping the corresponding query space on ST, IS, and IT space Keyspace With value space V t i For the Cross-Attention mechanism, the query space of different spaces is obtained by mapping the information of the space itself, and the key space and value space on different spaces are obtained by mapping the information of other spaces. The calculation formula is: α-IT=Cross Attention{SGCN IT [Pool IT (GCN IT (input IT ))]}))] (5) In formula (6), (χ∈{q,k,v}, i∈{1,2,…,h}) correspond to the feature extraction functions in the ST, IS, and IT branch spaces respectively; Represents the query vector of the i-th element in the ST space at time t, which is represented by the function according to Calculated; Represents the query vector of the i-th element in the IS space at time t, which is represented by the function according to Calculated; Represents the query vector of the i-th element in the IT space at time t, represented by the function according to Calculated; Respectively represent the bond vectors in ST, IS, and IT spaces; V t STi ,V t ISi ,V t ITi Respectively represent the value vectors in ST, IS, and IT spaces; W in formula (8) m Represents a trainable parameter, d = m / h, where d represents days, m represents dimension, and h represents the length of the input sequence.

7. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The step four) includes the following steps: combining the sparse graph convolutional network and the cross attention mechanism, and performing residual connection and normalization processing to generate a combined prediction model based on the Transformer network Encoder-Decoder architecture.

8. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The model combination process is as follows: Step 1: Set the three Transformer latent spaces of ST, IS, and IT to be m in dimension, and transform the original OD passenger flow data X t ∈R 1×n×n Mapped to different latent spaces, we get Step 2: Mapped to the output Chead through a Cross-Attention mechanism t ; Step 3: After residual connection and normalization operations, the generalization of the model network is enhanced and the effectiveness of training deep learning networks is improved; Step 4: The output Y is obtained through a feed-forward neural network FCs consisting of two fully connected layers t n×n ; Step 5: Repeat step 4 to obtain OD passenger flow related information Y t od ; Step 6: Let X t =Y t od , repeat step 2 to step 5 k+2 times to deeply extract feature information.

9. The short-term OD passenger flow prediction method for urban rail transit according to claim 1 is characterized in that: The residual connection and normalization operation are: Y t m =Norm[X t +GCN(Permission t )] (9) In formula (9), Norm(·) represents the normalization function, and GCN(·) represents the graph convolution function; The output Y is obtained by a feed-forward neural network FCs consisting of two fully connected layers t n×n ; Y t n×n =FCs(Y t m ) (10)。

Images