Traffic prediction method and device based on spatiotemporal feature embedding and gate operation optimization

Abstract

The application belongs to the technical field of traffic data prediction, and particularly relates to a traffic prediction method and device based on space-time feature embedding and gate operation optimization. The method comprises the following steps: introducing time step feature relationship embedding and time step space relationship embedding, generating a comprehensive feature representation considering time and space, and combining historical traffic node data to obtain fused features; constructing a traffic data prediction model, wherein the traffic data prediction model comprises a graph convolution network and a gated recurrent unit; the graph convolution network can capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency; training the traffic data prediction model based on the fused features, and predicting future traffic data by using the trained traffic data prediction model. The application can better capture the space-time relationship in a complex traffic network, thereby providing more accurate and reliable traffic prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of traffic data prediction, and particularly relates to a traffic prediction method and device based on spatiotemporal feature embedding and gate operation optimization. BACKGROUND

[0002] With the acceleration of urbanization and the expansion of transportation networks, traffic congestion has become one of the important problems faced globally. Traffic congestion not only brings inconvenience to people's travel, but also has a negative impact on the environment, economy and social sustainable development. Therefore, accurately predicting traffic conditions is of great significance for improving traffic management efficiency, optimizing traffic flow distribution and improving travel experience.

[0003] Traditional prediction models based on statistical methods often have difficulty in effectively capturing the nonlinear relationship and spatiotemporal dependence of traffic data, which limits their accuracy and stability. Real-world traffic prediction problems are often accompanied by complex spatiotemporal features and variable patterns. Even the same road, the traffic conditions at different time periods will show significant differences. This is mainly due to the fact that existing models fail to fully utilize historical data information, making it difficult to accurately capture spatiotemporal dependence. SUMMARY

[0004] In order to overcome the problems in the prior art, the application provides a traffic prediction method and device based on spatiotemporal feature embedding and gate operation optimization.

[0005] The technical solution of the application to solve the above technical problems is as follows:

[0006] In a first aspect, the application provides a traffic prediction method based on spatiotemporal feature embedding and gate operation optimization, comprising the following steps:

[0007] Introducing time step feature relationship embedding and time step space relationship embedding to generate a comprehensive feature representation considering both time and space, and then combining historical traffic node data to obtain fusion features;

[0008] Constructing a traffic data prediction model, the traffic data prediction model comprising a graph convolution network and a gated recurrent unit; wherein the graph convolution network can capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependence relationship;

[0009] Training the traffic data prediction model based on the fusion features, and using the trained traffic data prediction model to predict future traffic data.

[0010] Further, a node library is constructed, and a graph structure is constructed using the node library; a convolution operation output of the current time step is generated using the fused feature and the graph structure; and further feature extraction is performed based on the convolution operation output of the current time step to obtain a combination of weighted node features.

[0011] Further, the construction of the graph structure using the node library specifically includes: generating a node tensor through linear transformation by learning the node content in the node library; and obtaining a relationship similarity matrix through multiplication operation on the node tensor and activation function and normalization processing, the relationship similarity matrix reflecting the relationship similarity between nodes and being used for constructing the graph structure.

[0012] Further, the generation of the convolution operation output of the current time step using the fused feature and the graph structure specifically includes:

[0013] The fused feature input and the output of the convolution operation of the previous time step are spliced together, and then are input together with the graph structure as the input of the convolution operation of the current time step; after the graph convolution operation, normalization processing is performed to obtain a feature vector; the even positions and the odd positions of the feature vector are respectively taken as two independent input streams, and are used for generating an update gate and a reset gate of the corresponding positions of the current time step; the reset gate is multiplied element-wise with the output of the convolution operation of the previous time step, and is spliced with the fused feature input, and is input together with the graph structure as the input of the convolution operation of the current time step; after the graph convolution operation, a candidate state of the current time step is obtained; the update gate is multiplied element-wise with the output of the convolution operation of the previous time step to obtain the preserved information of the previous time step; the update gate is multiplied element-wise with the candidate state of the current time step to obtain a new information part; and the two parts of information are combined to obtain the output of the convolution operation of the current time step.

[0014] Further, the further feature extraction based on the output of the convolution operation of the current time step to obtain the combination of weighted node features specifically includes:

[0015] The hidden state of the normalized time step is projected to a query vector space through a parameter matrix to obtain a query vector; the query vector is dot multiplied with each node feature in the node library to obtain an original attention score; the original attention score is normalized to obtain an attention weight; and the attention weight is weighted summed with the node feature to obtain the combination of weighted node features.

[0016] Further, after the introduction of the time step feature relationship embedding and the time step spatial relationship embedding, it further includes: remodeling the time step feature relationship embedding and the time step spatial relationship embedding; and performing addition operation on the remodeled time step feature embedding and the remodeled spatial feature embedding to obtain a comprehensive feature representation considering both time and space.

[0017] In a second aspect, a traffic prediction device based on spatio-temporal feature embedding and gate operation optimization is provided, comprising:

[0018] a time-space feature relationship embedding module for introducing time step feature relationship embedding and time step space relationship embedding to generate a comprehensive feature representation considering both time and space, and combining historical traffic node data to obtain fused features;

[0019] a traffic data prediction model module, wherein the traffic data prediction model comprises a graph convolution network unit and a gated recurrent unit; the graph convolution network unit is used to capture the spatial relationship between nodes, and the gated recurrent unit is capable of capturing the temporal dependency;

[0020] a traffic data prediction model training module for training the traffic data prediction model based on the fused features, and using the trained traffic data prediction model to predict future traffic data.

[0021] Compared with the prior art, the present application has the following technical effects:

[0022] The present application adopts spatio-temporal feature relationship embedding and fuses with input traffic data to more accurately predict traffic conditions at different times and locations. By increasing the feature dimension of node embedding, historical traffic information is fully utilized to better represent the traffic graph structure. Meanwhile, a new strategy is adopted in the gated recurrent unit, which separates the even and odd positions of the input sequence to generate the update gate and reset gate at the corresponding positions, so that the model can adaptively adjust the retention and forgetting degree of traffic information, thereby better transmitting and retaining long-term dependent traffic information. Finally, the extracted features are normalized by convolution to further enhance the feature extraction capability, thereby helping to capture traffic spatio-temporal features. These parts work together to better capture the spatio-temporal relationship in complex traffic networks, thereby providing more accurate and reliable traffic prediction. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0024] Figure 1 is a flowchart of the present application;

[0025] Figure 2 is a schematic diagram of the convolution gated recurrent unit of the present application;

[0026] Figure 3 Figure 1 is an experimental comparison chart between the real data and the predicted data of the present application in the NYC-Bike test set;

[0027] Figure 4 Figure 2 is an experimental comparison chart between the real data and the predicted data of the present application in the METR-LA test set;

[0028] Figure 5 Figure 3 is a comparison result chart of the present application and the prior art using partial data in the NYC-Bike test set as an example;

[0029] Figure 6 Figure 4 is a comparison chart of the present application and the MegaCRN model in the all-day traffic flow prediction in the METR-LA test set. DETAILED DESCRIPTION

[0030] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined object, the specific embodiments, structures, features and effects of the technical solutions proposed by the present application are described in detail below in combination with the drawings and preferred embodiments. The specific features, structures or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art of the technology to which the present application belongs.

[0031] Real-world traffic prediction problems often involve complex spatio-temporal characteristics and variable patterns. Current models may encounter some challenges when dealing with these complex traffic data, and the prediction performance is often unsatisfactory. In order to solve the limitations encountered by existing traffic prediction methods in practice, an effective method for predicting traffic data is designed.

[0032] In one embodiment of the present application, referring to Figures 1-2 , a traffic prediction method based on spatio-temporal feature embedding and gate operation optimization is provided, comprising the following steps:

[0033] Step 100: Introduce time step feature relationship embedding and time step space relationship embedding to generate a comprehensive feature representation considering both time and space, and then combine historical traffic node data to obtain fused features;

[0034] Step 200: Construct a traffic data prediction model, wherein the traffic data prediction model comprises a graph convolution network and a gated recurrent unit; wherein the graph convolution network can capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency;

[0035] Step 300: Train the traffic data prediction model based on the fused features, and use the trained traffic data prediction model to predict future traffic data.

[0036] The above steps are described in detail as follows:

[0037] Step 100: Introduce time step feature relationship embedding and time step space relationship embedding to generate a comprehensive feature representation considering both time and space, and combine historical traffic node data to obtain fused features.

[0038] Before that, the work task needs to be defined first:

[0039] In order to predict the data of future traffic nodes, a multi-step to multi-step prediction form is adopted. Given T the data of traffic nodes in a certain area within a continuous time, i.e. historical traffic node data, represented as N , , O is the number of node data. By using historical data to train the model, the data of future time steps is predicted, where a represents the number of observed historical time steps, b represents the number of predicted future time steps.

[0040] In this embodiment, a specific implementation of step 100 can be:

[0041] Step 1001: Introduce time step feature relationship embedding TOE( ) and time step space relationship embedding TSE( ).

[0042] In the initial state, the time step feature relationship embedding and the time step space relationship embedding are randomly initialized two-dimensional matrices. In the learning process, the time step feature relationship embedding learns an embedding vector for each time step and each feature dimension to capture the influence of time steps on feature expression. The time step space relationship embedding learns an embedding vector for each time step and each spatial node, aiming to capture the influence of time steps on spatial positions.

[0043] The time step feature relationship embedding is used to capture the feature relationship between time steps, and the time step space relationship embedding is used to capture the feature relationship of different spatial positions within a time step.

[0044] Step 1002: Reshape the time step node feature relationship embedding and the time step space relationship embedding.

[0045] Since the time step feature relationship embedding and the space relationship embedding may have different dimensions, they need to be reshaped to the same dimension. In order to effectively integrate these information, the time step node feature relationship embedding is first reshaped to a new dimension ( ), where the number of time stepsN is set to 1, and similarly, the time step space relation embedding is also reshaped , where the number of node data O is set to 1.

[0046] Step 1003: Perform an addition operation on the reshaped time step feature embedding and the reshaped space feature embedding to obtain a comprehensive feature representation considering both time and space.

[0047] Step 1004: Concatenate the comprehensive feature representation with the historical traffic node data X to obtain a new tensor TOSE . This step is to retain the information of the original input data while combining the new spatio-temporal comprehensive features.

[0048] Step 1005: Perform a linear transformation on the concatenated tensor to obtain a new input , i.e., a fusion feature input, as the input data for subsequent steps. The purpose of linear transformation is to adjust the feature dimension to make it suitable for subsequent model processing.

[0049] Step 200: Construct a traffic data prediction model, which includes a graph convolution network and a gated recurrent unit; the graph convolution network can capture the spatial relationship between nodes, while the gated recurrent unit can capture the temporal dependence relationship. The specific structure of the convolution gated recurrent unit is as shown in Figure 2 .

[0050] As an example, step 200 can include the following sub-steps:

[0051] Step 210: Use the node library to construct a graph structure.

[0052] The node library is used to store important node information. The node library is defined as , where n represents the number of important nodes, h represents the dimension of the node feature vector. The node library is randomly assigned in the initial stage and continuously optimized and updated in the subsequent training process.

[0053] Nodes are intersections, transportation hubs, etc. with heavy traffic. Node information includes node feature vectors, which can reflect key information such as node traffic attributes and geographic location. After learning the node information in the node library, a relationship similarity matrix can be generated to construct a graph structure reflecting the relationship similarity between nodes. The graph structure is the basis for spatial feature extraction by the graph convolution network and helps the model to capture the spatial relationship between nodes in the traffic network.

[0054] In the present application, one specific implementation of step 200 can be:

[0055] Step 2101: generating node tensor by learning the content of nodes in the node library and linear transformation and .

[0056] Step 2102: obtaining two matrices, i.e. relationship similarity matrices and normalization processing , by multiplying two node tensors and passing through an activation function 、 The relationship similarity matrices reflect the relationship similarity between nodes and can be used to construct a graph structure. The formula is as follows:

[0057] (1) ;

[0058] In formula (1), parameter matrices W 1 and W 2 are key parameters learned for the node library, and parameter matrix W is essentially a projection of the learned node library into a new embedding space.

[0059] Through the two relationship similarity matrices, a graph structure G reflecting the relationship similarity between nodes is obtained. In the graph structure G, nodes correspond to nodes in the node library; if the values of two nodes in the relationship similarity matrix are greater than a certain threshold, an edge is added between the two nodes. The weight of the edge can be set as the value in the similarity matrix.

[0060] Step 220: based on the fusion feature input, the spatial feature is extracted by graph convolution operation, the time feature is processed by the gating recurrent unit, and the update gate and the reset gate are generated to control the flow of information; the update gate and the reset gate are combined with the output of the previous time step and the current input to generate the candidate state of the current time step and the preserved information of the previous time step, and the output of the current time step convolution operation is obtained by merging.

[0061] In the present application, a specific implementation mode of step 220 can be:

[0062] The fusion feature input and the output of the previous time step convolution operation are spliced together, and then are taken together with the graph structure as the input of the current time step convolution operation, after graph convolution operation, the function is applied for normalization processing to obtain the feature vector ; the even positions and the odd positions of the feature vector are taken as two independent input streams respectively, which are used to generate the update gate and reset gate ; the reset gate is multiplied with the output of the previous time step convolution operation element-wise multiplication operation, and is fused with the feature input spliced, and the graph structure together as the input of the current time step convolution operation, after the graph convolution operation, the function is applied to obtain the candidate state of the current time step update gate and the output of the previous time step convolution operation element-wise multiplication operation to obtain the retained previous time step information; 1 minus the update gate is element-wise multiplied with the candidate state of the current time step to obtain a new information part; the two parts of information are combined to obtain the output of the current time step convolution operation The specific formula is as follows:

[0063] (2)

[0064] In formula (2), denotes the graph convolution operation, denotes splicing of two tensors.

[0065] Step 230: Based on the output of the current time step convolution operation, further strengthen feature extraction, and obtain a combination of weighted node features.

[0066] In the present application, a specific implementation mode of step 230 can be:

[0067] Step 2301: Based on the output of the current time step convolution operation, after two relu activations, two-dimensional convolution operation and normalization, the hidden state of the normalized time step is obtained.

[0068] (3)

[0069] In the above formula, denotes two-dimensional convolution operation; denotes preliminary feature extraction, which reduces the interference of useless information by suppressing negative response, and highlights important signals in the input features; denotes further feature extraction, which further extracts the spatiotemporal correlation pattern by strengthening the significant features; denotes the hidden state of the normalized time step.

[0070] Step 2302: The hidden state of the normalized time step is interacted and fused with the features in the node library to obtain a combination of weighted node features.

[0071] The hidden state of the normalized time step is interacted with the features in the node library through a parameter matrix Projecting to the query vector space to obtain a query vector ; the query vector is multiplied point by point with each node feature in the node library to obtain the original attention score, and a function is applied to the original attention score to normalize and obtain the attention weight ; the attention weight is weighted and summed with the node feature to obtain the combination of weighted node features, and the specific formula is as follows:

[0072] (4);

[0073] In formula (4), denotes transposition.

[0074] The combination of weighted node features provides the model with the context information of the traffic state, and the query vector dynamically adjusts the degree of attention to these features according to the current state. The finally generated enhanced feature with context awareness capability, i.e. the combination of weighted node features val . val is the prediction basis for future traffic data, which combines the current state and historical information to help the model more accurately predict future traffic conditions.

[0075] Step 300: Based on the fusion features, a traffic data prediction model is trained, and the trained traffic data prediction model is used to predict future traffic data.

[0076] After the design of the traffic data prediction model is completed, the traffic data obtained based on step 100 is split, and the training data set, the verification data set and the test data set are generated in the ratio of 6:2:2, and the traffic data prediction model is trained. After the training of the traffic data prediction model is completed, the future traffic data is predicted according to the real-time collected traffic data.

[0077] The present application carries out experiments on three data METR-LA, NYC-Bike and NYC-Taxi, and the three data sets are public data sets. The METR-LA data set integrates the real-time monitored traffic speed data of 207 traffic nodes in Los Angeles. The NYC-Bike and NYC-Taxi data sets record the rental situation of bicycles and taxis at different locations in New York City in detail.

[0078] In this experiment, part of the data in two data sets is randomly selected for visualization, Figures 3-4 and the quantitative comparison between real data and predicted data is presented. Figure 3 ​Experimental comparison of part of the real data and predicted data in the NYC-Bike test set, Figure 4 Experimental comparison of part of the real data and predicted data in the METR-LA test set. The bicycle demand quantity prediction is an important quantitative mapping of traffic flow dynamic changes.

[0079] In order to further highlight the prediction advantage of the model, part of the data of NYC-Bike is randomly selected for comparison of prediction accuracy with the existing MegaCRN and RGDAN models, as shown in Figure 5 In addition, in order to highlight the prediction advantage of the application in the traffic flow peak period, the all-day data of a traffic monitoring station on June 5, 2016 in the METR-LA data set is selected as a sample, and a fine visualization analysis of the dynamic changes of the traffic flow is carried out, as shown in Figure 6 The fitting degree of the predicted trajectory of the application to the real observation value is higher than that of the MegaCRN model, and the prediction accuracy in the peak value area exhibits better performance. In terms of key evaluation indicators, the application shows significant improvement in MAE (mean absolute error): compared with the MegaCRN model, the prediction accuracy is increased by 16%, 18% and 19% when the prediction time step is 3, 6 and 12, respectively.

[0080] In the experimental process after the model is constructed, a batch training strategy is adopted, 64 batches of data are analyzed at the same time each time to ensure the calculation efficiency, and the initial learning rate is set to 0.01 to balance the learning speed and stability. When the training reaches the 15th round and the 35th round, the learning rate is reduced to one tenth of the original, and if the model is stable within 20 rounds, the training is ended in advance; if it does not meet the standard, it continues to learn to 200 rounds. All experiments are completed on a computer equipped with a 3060Ti graphics card. The configuration is repeatedly tested and can efficiently complete the model training while maintaining the prediction accuracy.

[0081] Based on the same inventive concept, the embodiment of the application also provides a traffic prediction device based on spatio-temporal feature embedding and gate operation optimization for implementing the traffic prediction method based on spatio-temporal feature embedding and gate operation optimization. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more device embodiments provided below can be referred to the limitations of the traffic prediction method based on spatio-temporal feature embedding and gate operation optimization in the above, which will not be repeated here.

[0082] In one embodiment, a traffic prediction device based on spatio-temporal feature embedding and gate operation optimization is provided, which comprises a time-space feature relationship embedding module, a traffic data prediction model module and a traffic data prediction model training module.

[0083] The time-space feature relationship embedding module is configured to introduce time step feature relationship embedding and time step space relationship embedding, generate a comprehensive feature representation considering both time and space, and obtain fused features in combination with historical traffic node data.

[0084] The traffic data prediction model module includes a graph convolution network unit and a gated recurrent unit; the graph convolution network unit is configured to capture spatial relationships between nodes, and the gated recurrent unit is capable of capturing time-dependent relationships.

[0085] The traffic data prediction model training module is configured to train the traffic data prediction model based on the fused features, and predict future traffic data by using the trained traffic data prediction model.

[0086] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A traffic prediction method based on spatio-temporal feature embedding and gate operation optimization, characterized in that, The method comprises the following steps: Introducing time step feature relationship embedding and time step space relationship embedding to generate a comprehensive feature representation considering both time and space, and combining historical traffic node data to obtain fused features; wherein, in the initial state, the time step feature relationship embedding and the time step space relationship embedding are two-dimensional matrices initialized randomly; A traffic data prediction model is constructed, which comprises a graph convolution network and a gated recurrent unit, and further comprises: constructing a node library to construct a graph structure; using the fused features and the graph structure to generate the convolution operation output of the current time step; based on the convolution operation output of the current time step, further extracting features to obtain a combination of weighted node features; wherein, the graph convolution network can capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency relationship, and specifically comprises: The constructing the graph structure by using the node library comprises: learning the node content in the node library D, and generating a node tensor through linear transformation 、 ; performing multiplication operation on the node tensor, and performing activation function relu and normalization processing softmax to obtain a relationship similarity matrix E 1、 E 2, the relationship similarity matrix reflects the relationship similarity between nodes, and is used for constructing the graph structure: ； In the above formula, the parameter matrix W 1 and W 2 are key parameters for learning the node library; W denotes the parameter matrix; The use of fused features and graph structure to generate the convolution operation output of the current time step specifically comprises: ； In the above formula, denotes a feature vector; denotes a fused feature; denotes the output of the previous time step convolution operation; denotes a graph structure; denotes an update gate; In the above formula, i denotes a row, denotes a column; denotes a reset gate; denotes a candidate state; denotes the output of the current time step convolution operation; denotes a graph convolution operation; denotes the concatenation of two tensors; Based on the convolution operation output of the current time step, further extracting features to obtain a combination of weighted node features, including: after two relu activations, two-dimensional convolution operation and normalization, the normalized hidden state of the time step is obtained; the normalized hidden state of the time step is projected to a query vector space through a parameter matrix to obtain a query vector; the query vector is dot producted with each node feature point in the node library to obtain an original attention score, and the original attention score is normalized to obtain an attention weight; the attention weight is weighted summed with the node feature to obtain a combination of weighted node features; Based on the fused features, the traffic data prediction model is trained, and the trained traffic data prediction model is used to predict future traffic data.

2. The traffic prediction method based on spatio-temporal feature embedding and gate operation optimization according to claim 1, characterized in that, After introducing the time step feature relationship embedding and the time step space relationship embedding, further comprising: remodeling the time step node feature relationship embedding and the time step space relationship embedding; adding the remodeled time step feature embedding and the remodeled space feature embedding to obtain a comprehensive feature representation considering both time and space.

3. A device for traffic prediction based on spatio-temporal feature embedding and gate operation optimization, for implementing a method for traffic prediction based on spatio-temporal feature embedding and gate operation optimization according to any one of claims 1 to 2, characterized in that, Comprise: A time-space feature relationship embedding module is used to introduce time step feature relationship embedding and time step space relationship embedding to generate a comprehensive feature representation considering both time and space, and combine historical traffic node data to obtain fused features; A traffic data prediction model module, wherein the traffic data prediction model comprises a graph convolution network unit and a gated recurrent unit; wherein, the graph convolution network unit is used to capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency relationship; A traffic data prediction model training module is used to train the traffic data prediction model based on the fused features, and the trained traffic data prediction model is used to predict future traffic data.

Images