Traffic prediction method and device based on spatio-temporal feature embedding and gate operation optimization

Abstract

The invention belongs to the technical field of traffic data prediction, and particularly relates to a traffic prediction method and device based on spatio-temporal feature embedding and gate operation optimization. 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 time and space at the same time, and combining historical traffic node data to obtain a fusion feature; constructing a traffic data prediction model, wherein the traffic data prediction model comprises a graph convolutional network and a gating circulation unit; wherein the graph convolutional network can capture a spatial relationship between nodes, and the gating loop unit can capture a dependency relationship in time; and training a traffic data prediction model based on the fusion features, and predicting future traffic data by adopting the trained traffic data prediction model. According to the method, the time-space relationship in the complex traffic network can be better captured, so that more accurate and reliable traffic prediction is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic data prediction, and in particular relates to a traffic prediction method and device based on spatiotemporal feature embedding and gate operation optimization. Background Art

[0002] With the acceleration of urbanization and the expansion of transportation networks, traffic congestion has become one of the major issues facing the world. 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, accurate prediction of traffic conditions is of great significance to improving traffic management efficiency, optimizing traffic flow distribution and improving travel experience.

[0003] Traditional prediction models based on statistical methods often have difficulty effectively capturing the nonlinear relationship and spatiotemporal dependency of traffic data, which limits their accuracy and stability. Real-world traffic prediction problems are often accompanied by complex spatiotemporal characteristics and changing patterns. Even on the same road, traffic conditions in different time periods can show significant differences. This is mainly because existing models fail to make full use of historical data information, making it difficult to accurately capture spatiotemporal dependencies. Summary of the invention

[0004] In order to overcome the problems in the prior art, the present invention proposes a traffic prediction method and device based on spatiotemporal feature embedding and door operation optimization.

[0005] The technical solution of the present invention to solve the above technical problems is as follows: In a first aspect, the present invention provides a traffic prediction method based on spatiotemporal feature embedding and gate operation optimization, comprising the following steps: Introduce time-step feature relationship embedding and time-step space relationship embedding to generate a comprehensive feature representation that considers both time and space, and then combine it with historical traffic node data to obtain fusion features; Constructing a traffic data prediction model, the traffic data prediction model includes a graph convolutional network and a gated recurrent unit; wherein the graph convolutional network can capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency; The traffic data prediction model is trained based on the fusion features, and the trained traffic data prediction model is used to predict future traffic data.

[0006] Furthermore, a node library is constructed, and a graph structure is constructed using the node library; the convolution operation output of the current time step is generated using the fused features and the graph structure; based on the convolution operation output of the current time step, features are further extracted to obtain a combination of weighted node features.

[0007] Furthermore, the construction of the graph structure using the node library specifically includes: generating node tensors by learning the node contents in the node library and performing linear transformation; obtaining a relationship similarity matrix by multiplying the node tensors and performing activation functions and normalization processing, wherein the relationship similarity matrix reflects the relationship similarity between nodes and is used to construct the graph structure.

[0008] Furthermore, the use of fusion features and graph structure to generate the convolution operation output of the current time step specifically includes: The fused feature input and the output of the convolution operation of the previous time step are spliced ​​together, and then used together with the graph structure as the input of the convolution operation of the current time step. After the graph convolution operation, normalization is performed to obtain a feature vector; the even and odd positions of the feature vector are used as two independent input streams to generate the update gate and reset gate of the corresponding position of the current time step; the reset gate is element-wise multiplied with the output of the convolution operation of the previous time step and spliced ​​with the fused feature input, and used together with the graph structure as the input of the convolution operation of the current time step. After the graph convolution operation, the candidate state of the current time step is obtained; the update gate is element-wise multiplied with the output of the convolution operation of the previous time step to obtain the retained information of the previous time step; the update gate is element-wise multiplied with the candidate state of the current time step to obtain the new information part; these two parts of information are combined to obtain the output of the convolution operation of the current time step.

[0009] Furthermore, the output of the convolution operation based on the current time step is further feature extracted to obtain a combination of weighted node features, specifically including: The hidden state of the normalized time step is projected to the query vector space through the parameter matrix to obtain the query vector; the query vector is dot-producted with each node feature in the node library to obtain the original attention score, and the original attention score is normalized to obtain the attention weight; the attention weight is weighted and summed with the node feature to obtain a combination of weighted node features.

[0010] Furthermore, after introducing the time-step feature relationship embedding and the time-step space relationship embedding, it also includes: reshaping the time-step node feature relationship embedding and the time-step space relationship embedding; adding the reshaped time-step feature embedding and the reshaped space feature embedding to obtain a comprehensive feature representation that considers both time and space.

[0011] In a second aspect, a traffic prediction device based on spatiotemporal feature embedding and gate operation optimization is provided, comprising: The 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 that considers both time and space, and then combines it with historical traffic node data to obtain fusion features; A traffic data prediction model module, wherein the traffic data prediction model includes a graph convolutional network unit and a gated recurrent unit; wherein the graph convolutional network unit is used to capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency; The traffic data prediction model training module is used to train the traffic data prediction model based on the fusion features, and use the trained traffic data prediction model to predict future traffic data.

[0012] Compared with the prior art, the present invention has the following technical effects: The present invention adopts the embedding of spatiotemporal feature relationships and integrates them with the input traffic data to more accurately predict the traffic conditions at different times and places. By increasing the feature dimension of node embedding, historical traffic information is fully utilized to better represent the traffic graph structure. At the same time, a new strategy is adopted in the gated recurrent unit, which separates the even and odd positions of the input sequence to generate update gates and reset gates 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 convolutionally normalized to further enhance the feature extraction capability, thereby helping to capture the spatiotemporal characteristics of traffic. These parts work together to better capture the spatiotemporal relationships in complex traffic networks, thereby providing more accurate and reliable traffic predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0014] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 Schematic diagram of a convolutional gated recurrent unit of the present invention; Figure 3 This is an experimental comparison diagram between the real data and the predicted data in the NYC-Bike test set of the present invention; Figure 4 This is an experimental comparison diagram of some real data and predicted data in the METR-LA test set of the present invention; Figure 5 This is a comparison result diagram of the present invention with the prior art using some data in the NYC-Bike test set as an example; Figure 6This is a comparison chart of all-day traffic flow prediction in the METR-LA test set between the present invention and the MegaCRN model. DETAILED DESCRIPTION

[0015] In order to further explain the technical means and effects taken by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. The specific features, structures or characteristics in one or more embodiments may be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present invention.

[0016] Traffic prediction problems in reality are often accompanied by complex spatiotemporal characteristics and changing patterns. Current models may encounter some challenges when processing these complex traffic data, and the prediction performance is often unsatisfactory. In order to address the limitations of existing traffic prediction methods in practice, a method for effectively predicting traffic data is designed.

[0017] In one embodiment of the present invention, referring to Figure 1-Figure 2 , provides a traffic prediction method based on spatiotemporal feature embedding and gate operation optimization, including the following steps: Step 100: Introduce time-step feature relationship embedding and time-step space relationship embedding to generate a comprehensive feature representation that considers both time and space, and then combine it with historical traffic node data to obtain fusion features; Step 200: constructing a traffic data prediction model, wherein the traffic data prediction model includes a graph convolutional network and a gated recurrent unit; wherein the graph convolutional network can capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency; Step 300: Train the traffic data prediction model based on the fusion features, and use the trained traffic data prediction model to predict future traffic data.

[0018] The following is a detailed explanation of each of the above steps: Step 100: Introduce time-step feature relationship embedding and time-step space relationship embedding to generate a comprehensive feature representation that considers both time and space, and then combine it with historical traffic node data to obtain fused features.

[0019] Before that, you need to define the work tasks first: In order to predict the data of future traffic nodes, a multi-step to multi-step prediction form is adopted. T A region in continuous time N The data of traffic nodes, that is, historical traffic node data, is expressed as ,O is the number of node data. By using historical data To train the model to predict data for several time steps in the future ,in a represents the number of historical time steps of observation, b Indicates the number of future time steps to predict.

[0020] In this embodiment, a specific implementation of step 100 may be: Step 1001: Introduce time-step feature relationship embedding TOE ( ) and time-step spatial relation embedding TSE( ).

[0021] In the initial state, the time-step feature relationship embedding and the time-step spatial relationship embedding are randomly initialized two-dimensional matrices. During the learning process, the time-step feature relationship embedding learns an embedding vector for each time step and each feature dimension to capture the impact of the time step on the feature expression. The time-step spatial relationship embedding learns an embedding vector for each time step and each spatial node to capture the impact of the time step on the spatial position.

[0022] Time-step feature relation embedding is used to capture the feature relation between time steps, and time-step spatial relation embedding is used to capture the feature relation at different spatial positions within a time step.

[0023] Step 1002: Reshape the time-step node feature relationship embedding and the time-step spatial relationship embedding.

[0024] Since the time-step feature relation embedding and the spatial relation embedding may have different dimensions, they need to be reshaped into the same dimension. In order to effectively integrate this information, we first reshape the time-step node feature relation embedding into a new dimension ( ), where the number of time steps N is set to 1. Similarly, the time-step spatial relationship embedding is also reshaped ( ), where the number of node data O Set to 1.

[0025] Step 1003: Add the reshaped time step feature embedding and the reshaped space feature embedding to obtain a comprehensive feature representation that considers both time and space.

[0026] Step 1004: Combine the comprehensive feature representation with the historical traffic node data X Splice to get the new tensor TOSE( ). This step is to preserve the information of the original input data while combining the new spatiotemporal comprehensive features.

[0027] Step 1005: Perform linear transformation on the concatenated tensor to obtain a new input , that is, fusion feature input, as the input data of the subsequent steps. The purpose of linear transformation is to adjust the feature dimension to make it suitable for subsequent model processing.

[0028] Step 200: Construct a traffic data prediction model, which includes a graph convolutional network and a gated recurrent unit. The graph convolutional network can capture the spatial relationship between nodes, while the gated recurrent unit can capture the temporal dependency. The specific structure of the convolutional gated recurrent unit is as follows: Figure 2 shown.

[0029] As an example, step 200 may include the following sub-steps: Step 210: Use the node library to build a graph structure.

[0030] The node library is used to store important node information. The node library is defined as ,in n Indicates the number of important nodes, h The dimension of the feature vector representing the node. The node library is randomly assigned in the initial stage and continuously optimized and updated in the subsequent training process.

[0031] Nodes are intersections and transportation hubs with large traffic volumes. Node information includes the node's feature vectors, which can reflect key information such as the node's 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 that reflects the relationship similarity between nodes. The graph structure is the basis for the graph convolutional network to extract spatial features, which helps the model capture the spatial relationship between nodes in the transportation network.

[0032] In the present invention, a specific implementation of step 200 may be: Step 2101: Generate node tensors by learning the node contents in the node library and performing linear transformation and .

[0033] Step 2102: Multiply the two node tensors and pass them through the activation function and normalization , we get two matrices, namely the relationship similarity matrix , , the relationship similarity matrix reflects the relationship similarity between nodes and can be used to construct a graph structure. The formula is as follows: (1) ; In formula (1), the parameter matrix W 1 andW 2 are the key parameters for learning the node library, and the parameter matrix W Essentially, the learned node library is projected into a new embedding space.

[0034] Through these two relationship similarity matrices, a graph structure G reflecting the relationship similarity between nodes is obtained. In the graph structure G, the nodes correspond to the nodes in the node library; if the value of two nodes in the relationship similarity matrix is ​​greater than a certain threshold, an edge is added between the two nodes. The weight of the edge can be set to the value in the similarity matrix.

[0035] Step 220: Based on the fused feature input, the graph convolution operation extracts the spatial features, the gated recurrent unit processes the temporal features, and generates update gates and reset gates to control the flow of information; the update gates and reset gates 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 retained information of the previous time step, and the output of the convolution operation of the current time step is obtained by merging.

[0036] In the present invention, a specific implementation of step 220 may be: Input the fusion features And the output of the convolution operation at the previous time step Stitched together and then with the graph structure Together as the input of the convolution operation of the current time step, after the graph convolution operation, apply The function is normalized to obtain the eigenvector ; Eigenvector The even and odd positions of the s are used as two independent input streams to generate the update gates for the corresponding positions of the current time step. and reset gate ; Reset the gate and convolve the output of the previous time step Perform element-wise multiplication and combine with the fused feature input Splicing and graph structure Together as the input of the convolution operation of the current time step, after the graph convolution operation, apply Function, get the candidate state of the current time step ; Update the gate with the output of the convolution operation at the previous time step The element-wise multiplication operation is performed to obtain the information of the previous time step. The candidate state of the current time step is multiplied element-wise by 1 minus the update gate to obtain the new information part. The two parts of information are combined to obtain the output of the convolution operation of the current time step. The specific formula is as follows: (2); In formula (2), represents the graph convolution operation, Represents the concatenation of two tensors.

[0037] Step 230: Based on the output of the convolution operation at the current time step, further enhance feature extraction to obtain a combination of weighted node features.

[0038] In the present invention, a specific implementation of step 230 may be: Step 2301: Based on the output of the convolution operation at the current time step, after two relu activations, two-dimensional convolution operations and normalization, the hidden state of the normalized time step is obtained.

[0039] (3); In the above formula, Represents a two-dimensional convolution operation; Represents the preliminary feature extraction. This stage reduces the interference of useless information by suppressing negative responses and highlights the important signals in the input features. Indicates further feature extraction. In this stage, the spatiotemporal correlation patterns are further extracted by strengthening the significant features. represents the hidden state at the normalized time step.

[0040] 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.

[0041] The hidden state of the normalized time step Through the parameter matrix Projected into the query vector space, the query vector is obtained ; The query vector Each node feature in the node library Dot product, get the original attention score, apply the original attention score The function is normalized to obtain the attention weight ; The attention weight With node features Perform weighted summation to obtain a combination of weighted node features , the specific formula is as follows: (4); In formula (4), Indicates transpose.

[0042] The weighted node feature combination provides the model with contextual information about the traffic state, and the query vector dynamically adjusts the degree of attention paid to these features according to the current state. The final generated context-aware enhanced feature is the weighted node feature combination. val . valIt is the basis for predicting future traffic data. It combines current status with historical information to help the model predict future traffic conditions more accurately.

[0043] Step 300: Train the traffic data prediction model based on the fusion features, and use the trained traffic data prediction model to predict future traffic data.

[0044] After the design of the traffic data prediction model is completed, the traffic data obtained based on step 100 is split in a ratio of 6:2:2 to generate a training data set, a validation data set, and a test data set to train the traffic data prediction model. After the training of the traffic data prediction model is completed, future traffic data is predicted based on the real-time collected traffic data.

[0045] The present invention conducts experiments on three datasets, METR-LA, NYC-Bike, and NYC-Taxi. The three datasets are all public datasets. The METR-LA dataset integrates the traffic speed data monitored in real time at 207 traffic nodes in Los Angeles. The NYC-Bike and NYC-Taxi datasets record in detail the rental conditions of bicycles and taxis at different locations in New York City.

[0046] This experiment randomly selects some data from two datasets for visualization. Figure 3-Figure 4 A quantitative comparison of real data and predicted data is presented. Figure 3 It shows the experimental comparison between some real data and predicted data in the NYC-Bike test set. Figure 4 The experimental comparison between some real data and predicted data in the METR-LA test set is shown in Figure 1. Among them, the bicycle demand quantity prediction is an important quantitative mapping of the dynamic changes of traffic flow.

[0047] In order to further highlight the prediction advantage of this model, some NYC-Bike data are randomly selected to compare the prediction accuracy with the existing MegaCRN and RGDAN models. Figure 5 In addition, in order to highlight the prediction advantage of the present invention during the peak traffic flow period, the full-day data of a traffic monitoring station on June 5, 2016 in the METR-LA dataset was selected as a sample, and a refined visualization analysis of the dynamic changes in traffic flow was performed, as shown in the figure. Figure 6 As shown. The prediction trajectory of the present invention is more consistent with the actual observation value than the MegaCRN model, especially in the prediction accuracy of the peak area, showing better performance. In terms of key evaluation indicators, the present invention shows significant improvement in MAE (mean absolute error): when the prediction time steps are 3, 6, and 12, the prediction accuracy is improved by 16%, 18%, and 19% respectively compared with the MegaCRN model.

[0048] In the experimental process after model construction, a batch training strategy was adopted, and 64 batches of data were analyzed simultaneously each time to ensure computational efficiency. The initial learning rate was set to 0.01 to balance learning speed and stability. When training to the 15th and 35th rounds, the learning rate was reduced to one-tenth of the original. If the model performed stably within 20 rounds, the training was terminated in advance; if it did not meet the standard, the learning continued to 200 rounds. All experiments were completed on a computer equipped with a 3060Ti graphics card. This configuration has been repeatedly tested and can efficiently complete model training while maintaining prediction accuracy.

[0049] Based on the same inventive concept, the embodiment of the present invention also provides a traffic prediction device based on spatiotemporal feature embedding and door operation optimization for implementing the above-mentioned traffic prediction method based on spatiotemporal feature embedding and door operation optimization. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more device embodiments provided below can refer to the above-mentioned limitations on a traffic prediction method based on spatiotemporal feature embedding and door operation optimization, and will not be repeated here.

[0050] In one embodiment, a traffic prediction device based on spatiotemporal feature embedding and gate operation optimization is provided, the device comprising a spatiotemporal feature relationship embedding module, a traffic data prediction model module, and a traffic data prediction model training module; The 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 that considers both time and space, and then combines it with historical traffic node data to obtain fusion features; A traffic data prediction model module, wherein the traffic data prediction model includes a graph convolutional network unit and a gated recurrent unit; wherein the graph convolutional network unit is used to capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency; The traffic data prediction model training module is used to train the traffic data prediction model based on the fusion features, and use the trained traffic data prediction model to predict future traffic data.

[0051] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A traffic prediction method based on spatiotemporal feature embedding and gate operation optimization, characterized in that: The following steps are involved: Introduce time-step feature relationship embedding and time-step space relationship embedding to generate a comprehensive feature representation that considers both time and space, and then combine it with historical traffic node data to obtain fusion features; Constructing a traffic data prediction model, the traffic data prediction model includes a graph convolutional network and a gated recurrent unit; wherein the graph convolutional network can capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency; The traffic data prediction model is trained based on the fusion features, and the trained traffic data prediction model is used to predict future traffic data.

2. The traffic prediction method based on spatiotemporal feature embedding and gate operation optimization according to claim 1 is characterized in that: The construction of the traffic data prediction model specifically includes: constructing a node library, and using the node library to construct a graph structure; using fusion features and the graph structure to generate a 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.

3. The traffic prediction method based on spatiotemporal feature embedding and gate operation optimization according to claim 2 is characterized in that: The method of constructing a graph structure using a node library specifically includes: generating a node tensor by learning the node content in the node library and performing a linear transformation; obtaining a relationship similarity matrix by performing a multiplication operation on the node tensor and performing an activation function and normalization processing. The relationship similarity matrix reflects the relationship similarity between nodes and is used to construct a graph structure.

4. The traffic prediction method based on spatiotemporal feature embedding and gate operation optimization according to claim 3 is characterized in that: The method of using the fusion features and the graph structure to generate the convolution operation output of the current time step specifically includes: The fused feature input and the output of the convolution operation of the previous time step are spliced ​​together, and then used together with the graph structure as the input of the convolution operation of the current time step. After the graph convolution operation, normalization is performed to obtain a feature vector; the even and odd positions of the feature vector are used as two independent input streams to generate the update gate and reset gate of the corresponding position of the current time step; the reset gate is element-wise multiplied with the output of the convolution operation of the previous time step, and spliced ​​with the fused feature input, and used together with the graph structure as the input of the convolution operation of the current time step. After the graph convolution operation, the candidate state of the current time step is obtained; the update gate is element-wise multiplied with the output of the convolution operation of the previous time step to obtain the retained information of the previous time step; the update gate is element-wise multiplied with the candidate state of the current time step to obtain the new information part; these two parts of information are combined to obtain the output of the convolution operation of the current time step.

5. The traffic prediction method based on spatiotemporal feature embedding and gate operation optimization according to claim 4 is characterized in that: The output of the convolution operation based on the current time step is further feature extracted to obtain a combination of weighted node features, specifically including: The hidden state of the normalized time step is projected to the query vector space through the parameter matrix to obtain the query vector; the query vector is dot-producted with each node feature in the node library to obtain the original attention score, and the original attention score is normalized to obtain the attention weight; the attention weight is weighted and summed with the node feature to obtain a combination of weighted node features.

6. The traffic prediction method based on spatiotemporal feature embedding and gate operation optimization according to claim 4 is characterized in that: After the time-step feature relationship embedding and the time-step space relationship embedding are introduced, the method further includes: reshaping the time-step node feature relationship embedding and the time-step space relationship embedding; and adding the reshaped time-step feature embedding and the reshaped space feature embedding to obtain a comprehensive feature representation that considers both time and space.

7. A traffic prediction device based on spatiotemporal feature embedding and gate operation optimization, used to implement a traffic prediction method based on spatiotemporal feature embedding and gate operation optimization as claimed in any one of claims 1 to 6, characterized in that: include: The 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 that considers both time and space, and then combines it with historical traffic node data to obtain fusion features; A traffic data prediction model module, wherein the traffic data prediction model includes a graph convolutional network unit and a gated recurrent unit; wherein the graph convolutional network unit is used to capture the spatial relationship between nodes, and the gated recurrent unit can capture the temporal dependency; The traffic data prediction model training module is used to train the traffic data prediction model based on the fusion features, and use the trained traffic data prediction model to predict future traffic data.

Images