Short-term OD passenger flow prediction method for urban rail transit based on multi-task learning

Abstract

The present invention discloses a method for predicting short-term OD passenger flow of urban rail transit based on multi-task learning. The method first constructs a spatiotemporal propagation module, takes in-station and out-station passenger flow as node features, and OD passenger flow as edge features; performs spatiotemporal feature learning on node features and edge features to obtain spatiotemporal features of nodes and edges; transmits and aggregates the spatiotemporal feature information of nodes and edges layer by layer to obtain feature representations of nodes and edges; constructs an OD passenger flow prediction module and an in-station and out-station passenger flow prediction module to obtain passenger flow prediction results. The present invention can better capture the spatiotemporal dependency in the rail transit network by combining the joint modeling of OD flow and in-station and out-station flow, and learns OD passenger flow prediction and in-station and out-station passenger flow prediction as a joint task, which can fully utilize the spatiotemporal features of in-station and out-station flow, improve the accuracy of OD passenger flow prediction and the generalization ability of the model, have strong practical application value, and provide more reliable technical support for rail transit management and decision-making.

Description

Technical Field

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

[0002] With the rapid development of urban rail transit systems, passenger flow forecasting has become a key process in rail transit management, serving as an important means to improve operational efficiency, alleviate traffic congestion, optimize resource allocation, and enhance the passenger experience. Short-term OD (origin-destination) passenger flow forecasting is crucial for predicting passenger flow demand for different lines and stations within a specific timeframe, guiding operational scheduling and resource allocation. However, traditional forecasting methods based on statistics and simple machine learning often overlook the complex spatiotemporal dependencies within rail transit networks and the dynamic connections between stations, resulting in insufficient performance when dealing with complex passenger flow patterns and external influencing factors.

[0003] In recent years, deep learning, particularly graph neural networks (GNNs), has made significant progress in processing spatiotemporal data with complex structures. GNNs can effectively model relationships between nodes, demonstrating significant advantages, particularly in systems with significant spatial and temporal connections, such as urban rail transit. Among them, message-passing-based graph neural networks (MPNNs), as an innovative graph learning method, can capture spatial dependencies and temporal dynamics by transferring information between graph nodes, and have gradually become a research hotspot in the field of rail transit passenger flow prediction. However, existing research has mostly focused on model optimization for a single task, often overlooking the interconnectedness and synergy between multiple prediction tasks in rail transit systems. Summary of the Invention

[0004] The purpose of the present invention is to provide a short-term OD passenger flow prediction method for urban rail transit based on multi-task learning in response to the above-mentioned problems. By combining the joint modeling of OD flow and in- and out-station flow, it can better capture the spatiotemporal dependencies in the rail transit network, and learn OD passenger flow prediction and in- and out-station passenger flow prediction as a joint task. It can make full use of the spatiotemporal characteristics of in- and out-station flow, improve the accuracy of OD passenger flow prediction and the generalization ability of the model, have strong practical application value, provide more reliable technical support for rail transit management and decision-making, and provide new ideas and methods for the development of the field of short-term passenger flow prediction. It has broad application prospects.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is as follows:

[0006] According to one aspect of the present invention, a method for predicting short-term OD passenger flow in urban rail transit based on multi-task learning is provided, comprising the following steps:

[0007] S1. Using inbound and outbound passenger flow as node features and outbound passenger flow as edge features, a spatiotemporal propagation module is constructed. The spatiotemporal propagation module is composed of multiple spatiotemporal learning units stacked in descending time steps.

[0008] S2. Use the spatiotemporal learning unit to perform spatiotemporal feature learning on node features and edge features to obtain spatiotemporal features of nodes and edges;

[0009] S3, the spatiotemporal feature information of nodes and edges is transmitted and aggregated layer by layer in the spatiotemporal propagation module to obtain the feature representation of nodes and edges;

[0010] S4. Using the feature representations of edges and nodes as input, we construct an OD passenger flow prediction module as the main task and an inbound and outbound passenger flow prediction module as the auxiliary task. We use the OD passenger flow prediction module to obtain the OD matrix prediction result, and the inbound and outbound passenger flow prediction module to obtain the inbound and outbound passenger flow.

[0011] Preferably, in step S2, the spatiotemporal learning unit includes a node spatial feature learning subunit, a node temporal feature learning subunit and an edge spatiotemporal feature learning subunit;

[0012] The node space feature learning subunit is used to complete node space feature learning;

[0013] The node time feature learning subunit is used to complete node time feature learning;

[0014] The edge spatiotemporal feature learning subunit is used to complete edge spatiotemporal feature learning.

[0015] Preferably, the node spatial feature learning includes the following steps:

[0016] Calculate the neighbor node attention coefficient:

[0017] The neighbor node attention coefficient is calculated by the following formula:

[0018]

[0019] in, is the set of neighbor nodes of node i; j is the neighbor node, α i,j is the neighbor node attention coefficient; H i and H j are the features of nodes i and j respectively, W is the learnable weight matrix, a is the attention vector; k is any neighbor node of node i; H k is the feature vector of neighbor node k; | is vector splicing;

[0020] Calculate the node connection edge attention coefficient:

[0021] The node connection edge attention coefficient is calculated by the following formula:

[0022]

[0023] Among them, e i,j is the edge feature between nodes i and j; β i,j is the node connection edge attention coefficient; W e is the edge feature transformation matrix; b is the attention vector; e i,k Represents the edge features between node i and its adjacent node k;

[0024] Message aggregation:

[0025] Based on the attention coefficient, the information of neighboring nodes and edges is aggregated to obtain the spatial features of node i:

[0026]

[0027] Among them, H space is the spatial feature of node i.

[0028] Preferably, the edge spatiotemporal feature learning includes the following steps:

[0029] Node information aggregation:

[0030] Based on edge e i,j , aggregate the features of ingress node i and egress node j:

[0031]

[0032] Among them, φ node It is the node information aggregation function;

[0033] Historical side information fusion:

[0034] Set the current edge feature Edge features at the previous moment To perform the fusion:

[0035]

[0036] Among them, φ temp is the historical side information aggregation function;

[0037] Edge feature update:

[0038] Fuse node information and historical edge information and update edge features:

[0039]

[0040] Among them, φ edge Update function for edge features.

[0041] Preferably, in step S4, the OD passenger flow prediction module includes the following prediction steps:

[0042] Edge feature extraction:

[0043] Get the edge feature matrix E from the spatiotemporal propagation module i,h ∈R N×N×H , where E i,j represents the temporal and spatial implicit features from site i to site j, and H is the hidden layer dimension.

[0044] Fully connected mapping:

[0045] Perform a nonlinear transformation on each edge feature and map it to the prediction space:

[0046]

[0047] Among them, W OD1 ∈R H×H and b OD1 ∈R H is a learnable parameter;

[0048] OD matrix prediction:

[0049] The OD passenger flow prediction value for the next K time steps is output through the second-layer fully connected network:

[0050]

[0051] in, and b OD2 ∈R is the prediction layer parameter; the final output is Among them, each matrix dimension is N×N.

[0052] Preferably, in step S4, the inbound and outbound passenger flow prediction module includes the following prediction steps:

[0053] Node feature extraction:

[0054] Get the node feature matrix H from the spatiotemporal propagation module L ∈R N×H , where H L [i] represents the temporal and spatial implicit features of site i;

[0055] Feature transformation and activation:

[0056] Perform nonlinear transformation on node features:

[0057]

[0058] Among them, W IOt ∈R H×Hand b IOt ∈R H is a learnable parameter; LayerNorm is a layer normalization operation;

[0059] Passenger flow forecast:

[0060] Output the inbound and outbound passenger flow in the next K time steps through the fully connected network:

[0061]

[0062] in, and b IOt ∈R is the prediction layer parameter;

[0063] The final output is Each vector has dimension N.

[0064] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0065] By combining the joint modeling of OD flow and in- and out-station flow, the present invention can better capture the spatiotemporal dependencies in the rail transit network, and learn OD passenger flow prediction and in- and out-station passenger flow prediction as a joint task, which can make full use of the spatiotemporal characteristics of in- and out-station flow, improve the accuracy of OD passenger flow prediction and the generalization ability of the model, and has strong practical application value. It provides more reliable technical support for rail transit management and decision-making, and provides new ideas and methods for the development of the field of short-term passenger flow prediction, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic flow diagram of the present invention;

[0067] Figure 2 It is a schematic structural diagram of the space-time propagation module of the present invention;

[0068] Figure 3 It is a structural diagram of the prediction model of the present invention. DETAILED DESCRIPTION

[0069] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described below with reference to the accompanying drawings and by way of preferred embodiments. However, it should be noted that many of the details listed in this specification are merely provided to help the reader gain a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be practiced even without these specific details.

[0070] See also Figure 1 The present invention provides a method for predicting short-term OD passenger flow in urban rail transit based on multi-task learning. The technical solution is as follows:

[0071] A method for predicting short-term OD passenger flow in urban rail transit based on multi-task learning includes the following steps:

[0072] S1. Construct a space-time propagation module.

[0073] A spatiotemporal propagation module is constructed based on a message-passing graph neural network (MPNN). The overall structure of the spatiotemporal propagation module is an upward-pointing triangle, consisting of multiple stacked spatiotemporal learning units with decreasing time steps. The spatiotemporal propagation module uses inbound and outbound passenger flow as node features and outbound and outbound passenger flow as edge features. The inbound and outbound passenger flow and outbound and outbound passenger flow serve as inputs to the spatiotemporal propagation module, which is used to capture the spatiotemporal dependencies of passenger flow data. MPNN models the spatiotemporal dependencies between stations through a graph structure, effectively capturing the dynamic spatial relationships between stations and their time-varying characteristics, thereby improving the accuracy of passenger flow prediction.

[0074] S2. Use the spatiotemporal learning unit to perform spatiotemporal feature learning on node features and edge features to obtain the spatiotemporal features of nodes and edges.

[0075] After the node features and edge features are input, the spatiotemporal learning unit is used to implement spatiotemporal feature learning of nodes and edges in the message passing graph neural network. The spatiotemporal learning unit sequentially completes node spatial feature learning, node temporal feature learning, and edge spatiotemporal feature learning. Specifically:

[0076] Node spatial feature learning

[0077] Node spatial feature learning is used to aggregate information about neighboring nodes to capture the spatial dependencies between sites. The specific steps are as follows:

[0078] 1. Calculate the neighbor node attention coefficient:

[0079] For node i and its neighboring nodes is the set of neighbor nodes of node i, and calculates the attention coefficient α i,j :

[0080]

[0081] Among them, H i and H j are the features of nodes i and j respectively, W is the learnable weight matrix, a is the attention vector; k is any neighbor node of node i; H k is the feature vector of neighbor node k; | is vector splicing.

[0082] 2. Calculate the node connection edge attention coefficient:

[0083] For an edge e between nodes i and ji,j , calculate the edge attention coefficient β i,j :

[0084]

[0085] Among them, e i,j is the edge feature between nodes i and j; W e is the edge feature transformation matrix; b is the attention vector; e i,k Represents the edge feature between node i and its adjacent node k.

[0086] 3. Message aggregation:

[0087] Based on the attention coefficient, the information of neighboring nodes and edges is aggregated to obtain the spatial features of node i:

[0088]

[0089] Node time feature learning

[0090] Node time feature learning is used to fuse node information at the current moment and the previous moment to capture the temporal evolution of passenger flow data. The specific steps are as follows:

[0091] 1. Time information fusion:

[0092] The node features at the current moment Compared with the node characteristics of the previous moment To perform the fusion:

[0093]

[0094] Among them, φ temp It is a temporal fusion function, which is usually implemented using a gating mechanism (such as GRU or LSTM).

[0095] 2. Spatial-temporal feature fusion:

[0096] The spatial feature H space and time characteristics Perform fusion and update node features:

[0097]

[0098] Among them, φ fuse It is a feature fusion function, which usually adopts the method of splicing or weighted summation.

[0099] Learning spatiotemporal features of edges

[0100] Edge spatiotemporal feature learning is used to aggregate information from incoming and outgoing nodes and update edge features based on historical edge features. The specific steps are as follows:

[0101] 1. Node information aggregation:

[0102] For edge e i,j , aggregate the features of ingress node i and egress node j:

[0103]

[0104] Among them, φ node It is a node information aggregation function, which is usually implemented using a fully connected network.

[0105] 2. Historical side information fusion:

[0106] Set the current edge feature Edge features at the previous moment To perform the fusion:

[0107]

[0108] Among them, φ temp is the historical side information aggregation function;

[0109] 3. Edge feature update:

[0110] Fuse node information and historical edge information and update edge features:

[0111]

[0112] Among them, φ edge It is the edge feature update function, which is usually implemented using a fully connected network.

[0113] Through the above computational process, the spatiotemporal learning unit can simultaneously capture the spatiotemporal dependencies of nodes and edges, thereby comprehensively capturing the spatiotemporal dependencies of passenger flow data and providing efficient feature representation for passenger flow prediction tasks. This unit has significant advantages in modeling complex spatiotemporal patterns, effectively improving prediction accuracy.

[0114] S3. The spatiotemporal feature information of nodes and edges is transmitted and aggregated layer by layer in the spatiotemporal propagation module to obtain the feature representation of nodes and edges.

[0115] Specifically, such as Figure 2As shown in Figure 1, assume that the input data consists of L time steps. The spatiotemporal propagation module is composed of multiple stacked spatiotemporal learning units with decreasing time steps. Each layer in the spatiotemporal propagation module transmits the time step information of the current layer to the adjacent time step of the next layer through a message passing mechanism. For example, information at time steps t1 and t2 in the first layer will flow to time step t_2 in the second layer, and so on. As the time step decreases layer by layer, the information is propagated L times in time and space, ultimately converging to the last layer. By transferring and aggregating information layer by layer, it is possible to effectively capture the long-term dependencies and spatial correlations in passenger flow data, achieving efficient learning of spatiotemporal features.

[0116] The output of the spatiotemporal propagation module serves as the feature input for the subsequent passenger flow prediction task. Through a hierarchical and progressive information transfer mechanism, the spatiotemporal dynamic characteristics of passenger flow data are simultaneously learned, thereby improving prediction accuracy. The module has a significant advantage in capturing complex spatiotemporal patterns, providing a reliable feature representation for passenger flow prediction tasks.

[0117] S4. Construct the OD passenger flow prediction module and the in-and-out passenger flow prediction module, with the feature representation of nodes and edges as input, the OD passenger flow prediction module and the in-and-out passenger flow prediction module.

[0118] The prediction module includes the OD passenger flow prediction module and the inbound and outbound passenger flow prediction module. The OD passenger flow prediction is the main task, and the inbound and outbound passenger flow prediction is the auxiliary task. The task-specific network structure is used to achieve refined prediction. Specifically:

[0119] OD passenger flow prediction module

[0120] The OD passenger flow prediction module uses the edge features (i.e., the implicit representation of OD relationships) output by the spatiotemporal propagation module as input and generates the OD matrix prediction results for future time steps through a multi-layer network structure. The specific implementation steps are as follows:

[0121] 1. Edge feature extraction:

[0122] Get the edge feature matrix E from the spatiotemporal propagation module i,j ∈R N×N×H , where E i,j represents the temporal and spatial implicit features from site i to site j, and H is the hidden layer dimension.

[0123] 2. Fully connected mapping:

[0124] Perform a nonlinear transformation on each edge feature and map it to the prediction space:

[0125]

[0126] Among them, W OD1 ∈R H×H and b OD1 ∈RH are learnable parameters.

[0127] 3. OD matrix prediction:

[0128] The OD passenger flow prediction value for the next K time steps is output through the second-layer fully connected network:

[0129]

[0130] in, and b OD2 ∈R is the prediction layer parameter. The final output is Among them, each matrix dimension is N×N.

[0131] Assume that the OD matrix at time t is M t , where element m i,j represents the number of passengers who enter station i at time t and finally exit at station j. Assuming there are N stations in the rail transit network, then M t The dimension is N×N, which constitutes the OD demand graph of the entire network at that time point.

[0132] Inbound and outbound passenger flow prediction module

[0133] The inbound and outbound passenger flow prediction module uses the node features output by the spatiotemporal propagation module as input to predict the future inbound and outbound passenger flow at each station. The specific implementation steps are as follows:

[0134] 1. Node feature extraction:

[0135] Get the node feature matrix H from the spatiotemporal propagation module L [i]∈R N×H , where H L [i] represents the temporal and spatial implicit features of site i.

[0136] 2. Feature transformation and activation:

[0137] Perform nonlinear transformation on node features:

[0138]

[0139] Among them, W IOt ∈R H×H and b IOt ∈R H is a learnable parameter; LayerNorm is a layer normalization operation.

[0140] 3. Passenger flow forecast:

[0141] Output the inbound and outbound passenger flow in the next K time steps through the fully connected network:

[0142]

[0143] in, and b IOt ∈R is the prediction layer parameter. The final output is The dimension of each vector is N.

[0144] Assume that the passenger flow vector at time t is F t , where the elements represents the total passenger flow in and out of station i at time t. By modeling passenger flow data at different times using time series, short-term passenger flow forecasting can provide rail transit managers with accurate demand forecasting support, ensuring sufficient capacity to meet passenger demand in the upcoming time period.

[0145] The present invention optimizes short-term OD passenger flow prediction and short-term inbound and outbound passenger flow prediction as two subtasks in multi-task learning, so that the two can share the spatiotemporal features extracted from the model. This joint learning can not only obtain additional information from inbound and outbound traffic, but also improve the accuracy of OD passenger flow prediction, especially during peak hours or special events, when changes in inbound and outbound traffic play an important auxiliary role in OD passenger flow prediction. Compared with traditional single-task prediction methods, the present invention effectively improves the accuracy of OD passenger flow prediction, especially in the face of scarce or unbalanced data, and can provide more reliable prediction results.

[0146] Learning shared spatiotemporal features not only reduces the number of model parameters but also improves computational efficiency and avoids overfitting. Within the framework of multi-task learning, the joint learning of outbound and inbound / outbound traffic provides the model with a more comprehensive representation of spatiotemporal features, promoting mutual learning between the two and improving overall prediction accuracy. This multi-task learning architecture can effectively enhance short-term passenger flow forecasting capabilities in rail transit systems, providing more accurate decision support for urban traffic scheduling and operational management.

[0147] In actual tests, the OD passenger flow prediction method based on multi-task learning significantly outperformed traditional single-task methods, particularly in prediction tasks across different time periods and stations, demonstrating high prediction accuracy and strong generalization capabilities. This method not only provides a new solution for short-term OD passenger flow prediction, but also provides more accurate data support for intelligent scheduling and resource allocation of rail transit.

[0148] In summary, the present invention can better capture the spatiotemporal dependencies in the rail transit network by combining the joint modeling of OD flow and in- and out-station flow, has strong practical application value, and provides more reliable technical support for rail transit management and decision-making. The multi-task learning framework provides new ideas and methods for the development of the field of short-term passenger flow prediction and has broad application prospects.

[0149] The present invention also discloses a short-term OD passenger flow prediction model for urban rail transit based on multi-task learning, such as Figure 3 As shown in Figure 2, the model includes a spatiotemporal propagation module, consisting of multiple spatiotemporal learning units stacked in descending time steps, as well as OD passenger flow prediction and inbound and outbound passenger flow prediction modules. Based on a multi-task learning framework, the model is designed to improve the accuracy of short-term OD passenger flow prediction.

[0150] The model uses short-term outbound passenger flow prediction as the primary task and inbound and outbound passenger flow prediction as an auxiliary task. The model leverages inter-task synergy to improve overall prediction performance. First, the model constructs a spatiotemporal propagation module using a message-passing-based graph neural network (MPNN). In this module, inbound and outbound passenger flows are modeled as node features, and outbound passenger flows are modeled as edge features. MPNN models the spatiotemporal dependencies between stations through a graph structure, effectively capturing the dynamic spatial relationships and their time-varying characteristics, thereby improving passenger flow prediction accuracy. Within a multi-task learning framework, outbound and inbound passenger flow prediction are jointly trained by sharing network parameters. Inbound and outbound passenger flow serves as an auxiliary task, complementing outbound passenger flow prediction. This shared parameter structure facilitates knowledge sharing between tasks during training. In particular, changes in inbound and outbound passenger flows during peak hours or special events can help improve the accuracy of outbound passenger flow prediction. Furthermore, the model incorporates external features such as date type and weather data, which can reflect the impact of passenger flow over different time periods. Combining these features with spatiotemporal data further enhances the model's ability to perceive complex spatiotemporal features and improves forecasting performance. Through collaborative optimization between tasks and the integration of external features, the model can simultaneously predict both outbound and inbound passenger flows, providing more accurate short-term passenger flow forecasts. This not only improves the accuracy of outbound passenger flow forecasts but also enhances the model's robustness in complex environments, providing reliable decision-making support for the scheduling and management of urban rail transit.

[0151] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A short-term OD passenger flow prediction method for urban rail transit based on multi-task learning, characterized by: The following steps are involved: S1. Using inbound and outbound passenger flow as node features and outbound passenger flow as edge features, a spatiotemporal propagation module is constructed. The spatiotemporal propagation module is composed of multiple spatiotemporal learning units stacked in descending time steps. S2. Use the spatiotemporal learning unit to perform spatiotemporal feature learning on node features and edge features to obtain spatiotemporal features of nodes and edges; S3, the spatiotemporal feature information of nodes and edges is transmitted and aggregated layer by layer in the spatiotemporal propagation module to obtain the feature representation of nodes and edges; S4. Using the feature representations of edges and nodes as input, we construct an OD passenger flow prediction module (the main task) and an inbound and outbound passenger flow prediction module (the auxiliary task). We use the OD passenger flow prediction module to obtain the OD matrix prediction result, and the inbound and outbound passenger flow prediction module to obtain the inbound and outbound passenger flow. The OD passenger flow prediction module includes the following prediction steps: Edge feature extraction: Get the edge feature matrix from the space-time propagation module ,in Indicates a site To the site The temporal and spatial implicit characteristics of is the hidden layer dimension; Fully connected mapping: Perform a nonlinear transformation on each edge feature and map it to the prediction space: ， in, is a learnable parameter; Matrix prediction: Output the future through the second layer of fully connected network time steps Passenger flow forecast: ， in, and is the prediction layer parameter; the final output is , where each matrix dimension is ; The inbound and outbound passenger flow prediction module includes the following prediction steps: Node feature extraction: Get node feature matrix from spatiotemporal propagation module ,in Indicates a site The temporal and spatial implicit characteristics of Feature transformation and activation: Perform nonlinear transformation on node features: ， in, and is a learnable parameter; is the layer normalization operation; Passenger flow forecast: Outputting the future through a fully connected network Passenger flow in and out of the station at each time step: ， in, and is the prediction layer parameter; The final output is , each vector dimension is .

2. The urban rail transit short-term OD passenger flow prediction method based on multi-task learning according to claim 1 is characterized by: In step S2, the spatiotemporal learning unit includes a node spatial feature learning subunit, a node temporal feature learning subunit, and an edge spatiotemporal feature learning subunit; The node space feature learning subunit is used to complete node space feature learning; The node time feature learning subunit is used to complete node time feature learning; The edge spatiotemporal feature learning subunit is used to complete edge spatiotemporal feature learning.

3. The urban rail transit short-term OD passenger flow prediction method based on multi-task learning according to claim 2 is characterized by: The node spatial feature learning includes the following steps: Calculate the neighbor node attention coefficient: The neighbor node attention coefficient is calculated by the following formula: ， in, For nodes The set of neighbor nodes of is the neighbor node, is the neighbor node attention coefficient; and Node and Features, is the learnable weight matrix, is the attention vector; For nodes Any neighbor node of ; Neighbor nodes The eigenvector of For vector splicing; Calculate the node connection edge attention coefficient: The node connection edge attention coefficient is calculated by the following formula: ， in, For nodes and The edge features between is the node connection edge attention coefficient; is the edge feature transformation matrix; is the attention vector; Representation node Its adjacent nodes The edge features between Message aggregation: Based on the attention coefficient, the information of neighbor nodes and edges is aggregated to obtain the node Spatial characteristics: ， in, For nodes spatial characteristics.

4. The urban rail transit short-term OD passenger flow prediction method based on multi-task learning according to claim 3 is characterized by: The node time feature learning includes the following steps: Temporal information fusion: Fuse the node features at the current moment with the node features at the previous moment: ， in, is the node feature at the current moment; is the node feature of the previous moment; is the time fusion function; Spatiotemporal feature fusion: The spatial features and time characteristics Perform fusion and update node features: ， in, is the feature fusion function.

5. The urban rail transit short-term OD passenger flow prediction method based on multi-task learning according to claim 4 is characterized by: The edge spatiotemporal feature learning includes the following steps: Node information aggregation: Edge-based , aggregate into the node and egress nodes Features: ， in, It is the node information aggregation function; Historical side information fusion: Set the current edge feature Edge features at the previous moment To perform the fusion: ， in, is the historical side information aggregation function; Edge feature update: Fuse node information and historical edge information and update edge features: ， in, Update function for edge features.

Images