Traffic flow prediction method based on decoupling fusion network

Through the traffic flow prediction method of the decoupling fusion network, the noise enhancement timing model and the bidirectional diffusion feedback space model are used, combined with the bilinear space-time fusion model, the shortcomings of the existing technology in spatiotemporal feature modeling are solved, and more efficient Traffic flow forecast.

CN119049285APending Publication Date: 2024-11-29HANGZHOU DIANZI UNIV
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411206531.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing traffic flow prediction methods have shortcomings in spatiotemporal feature modeling. The increased complexity of deep learning methods has not brought about the expected performance improvement. The statistical methods are limited by the data linear correlation assumption. The spatial and temporal graph neural networks are used in spatial feature modeling. Poor performance.

Method used

Using a traffic flow prediction method based on a decoupled fusion network, the timing features are extracted through the noise-enhanced timing model, and a bidirectional diffusion feedback spatial model and a bilinear space-time fusion model are constructed, combining multi-layer residual multi-layer perceptron for feature fusion, abandoning tradition Graph convolution method to capture space-time dependencies.

Benefits of technology

Effectively alleviate the impact of non-periodic peak flow on the prediction results, improve the performance and efficiency of traffic flow prediction, concisely capture space-time dependence, and improve prediction performance and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119049285A_ABST
    Figure CN119049285A_ABST
Patent Text Reader

Abstract

The invention discloses a traffic flow prediction method based on a decoupling fusion network, and the method comprises the following steps: 1, obtaining traffic flow data, and obtaining time sequence features through a noise enhancement time sequence model; step 2, constructing a bidirectional diffusion feedback space model composed of a forward diffusion feedback space model and a backward diffusion feedback space model; 3, constructing a predefined graph, wherein the predefined graph is output to the bidirectional diffusion feedback space model; 4, constructing a bilinear space-time fusion model, wherein the bilinear space-time fusion model comprises three networks, namely a forward space-time fusion network, a reverse space-time fusion network and an overall fusion network; and step 5, outputting N prediction results from the extracted features through an N-layer bilinear space-time fusion model, splicing all the prediction results, and then obtaining a final prediction result through a feature fusion device. The method can effectively improve the prediction performance and efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a traffic flow prediction method based on a decoupled fusion network. Background Art

[0002] With the acceleration of urbanization, traffic flow prediction is of great significance for alleviating traffic congestion and improving road use efficiency. Existing prediction methods mainly include statistical methods and deep learning-based methods. Statistical methods are limited by the assumption of linear correlation of data, and although deep learning methods can capture time dependence, they are insufficient in modeling spatial features. As a mainstream method in recent years, spatiotemporal graph neural network (STGNN) has taken spatiotemporal features into account to a certain extent, but its increased complexity has not brought the expected performance improvement. Summary of the invention

[0003] In view of the deficiencies of the prior art, the present invention proposes a traffic flow prediction method based on a decoupled fusion network, which can effectively improve the performance and efficiency of the prediction.

[0004] In order to solve the above technical problems, the technical solution of the present invention is:

[0005] A traffic flow prediction method based on a decoupled fusion network includes the following steps:

[0006] Step 1: Obtain traffic flow data and obtain time series features through noise enhancement time series model;

[0007] Step 2: construct a bidirectional diffusion feedback space model consisting of a forward diffusion feedback space model and a reverse diffusion feedback space model, wherein the forward diffusion feedback space model and the reverse diffusion feedback space model have the same structure, both including a two-layer fully connected network and a residual connection;

[0008] Step 3: Get the known road network adjacency matrix A and the transpose A of the road network adjacency matrix A T , according to A and A respectively T Get an A for and A back The predefined graph A for As input, the forward graph G is obtained through the forward diffusion feedback space model. for , the predefined graph A back As input, the reverse graph G is obtained by back-diffusion feedback space model. back ;

[0009] Step 4: construct a bilinear space-time fusion model, which includes three networks: forward space-time fusion, reverse space-time fusion and overall fusion;

[0010] Step 5: F, T, Gfor and G back All the prediction results are equally and uniformly regarded as the features in the traffic flow prediction task, and N prediction results are output through the N-layer bilinear spatiotemporal fusion model. All the prediction results are spliced, and then the final prediction result is obtained through the feature fuser.

[0011] Preferably, in step 1, the traffic flow data X is used as input, and the periodic component F is extracted by the noise enhancement time series model. Per , then the traffic flow data X and the periodic component F Per The difference between Noi , and finally the periodic component F Per and the noise component F Noi Concatenate to obtain the time series feature F = concat(F Noi , F Per ).

[0012] Preferably, in step 1, the method of extracting the periodic component by using the noise enhanced time series model is:

[0013] Phase 1: Input traffic flow data X, and then pass it through a self-attention module and a residual multi-layer perceptron at a time;

[0014] The second stage: take the result of the first stage as input, perform multi-head cross attention with the result of the self-attention module in the second stage, and finally obtain the periodic component F through RMLP Per .

[0015] In the above technical solution, the periodic characteristics in the traffic flow data are decoupled from the noise through the noise enhancement time series model to alleviate the impact of non-periodic peak flow on the prediction results.

[0016] Preferably, in step 3, the predefined graph is a normalized weighted adjacency matrix, that is, A for =A / rowsum(A), A back =A r / rowsum(A T ).

[0017] In the above technical solution, a bidirectional diffusion feedback space model is used to retain the explicit association characteristics between nodes in the predefined graph, and the transmission and feedback between upstream and downstream traffic are used to describe the characteristics of the road network.

[0018] Preferably, in step 4, the three networks all use multi-layer RMLP as feature fusion devices.

[0019] Preferably, in step 4, when Z is used as input, the l-th layer bilinear spatiotemporal fusion model can be expressed as:

[0020]

[0021] Preferably, in step 5, the method for obtaining the prediction result by using the bilinear spatiotemporal fusion model is:

[0022] F, T, G for and G back As input, features are fused through forward spatiotemporal fusion and reverse spatiotemporal fusion respectively to obtain bidirectional hidden representation, and then the two are fused through overall fusion, and finally a layer of FC is used to obtain the result of this layer.

[0023] Preferably, step 5 comprises the following sub-steps:

[0024] Step 5.1: In the first layer, the forward spatiotemporal fusion and reverse spatiotemporal fusion parts are fused separately to obtain bidirectional hidden representations, and then the two are fused through overall fusion. Finally, a layer of FC is used to obtain the result of this layer, which can be expressed as:

[0025] H1=FC(RMLPs(RMLPs(F||G for ||T1)|| RMLPs(F||G back ||T1)))

[0026] Step 5.2: The result of the upper bilinear spatiotemporal fusion model is used as the new feature of the next bilinear spatiotemporal fusion model. Therefore, the lth layer can be expressed as:

[0027] H l+1 =FC(RMLPs(RMLPs(F||G for ||T l ||H l )||RMLPs(F||G back ||T l ||H l )))

[0028] Assuming that N layers of bidirectional fusion models are stacked, multi-layer output results can be obtained. Let H = concat(H1, ..., H N ), where H i ∈R N×F×P , combining the results of each layer to generate the final prediction

[0029]

[0030] Among them, R represents a real number, N represents the number of layers of the bidirectional fusion model, F represents the number of features, P represents the length of the past time series, and S represents the length of the predicted future time series.

[0031] In the above calculation scheme, the traditional graph convolution method is abandoned through the bilinear spatiotemporal fusion model, and a bilinear spatiotemporal fusion model based on residual multilayer perceptron is proposed to effectively capture the spatiotemporal dependency.

[0032] The present invention has the following characteristics and beneficial effects:

[0033] (1) Considering the impact of non-periodic factors in time series characteristics on traffic flow prediction, a time series prediction method based on noise enhancement is proposed. By decoupling the periodic characteristics in traffic flow data from noise and performing time modeling, the impact of non-periodic peak flow on the prediction results is alleviated.

[0034] (2) In response to the complex spatial dependency, a diffusion feedback spatial model is proposed for spatial modeling. The explicit association characteristics between nodes in the predetermined graph are retained, and the transmission and feedback between upstream and downstream flows are used to describe the characteristics of the road network and capture spatial dependencies.

[0035] (3) Abandoning the commonly used graph convolution method, a more concise spatiotemporal prediction network is proposed to effectively capture spatiotemporal dependencies. The spatial and temporal features of the intersection are associated through a multi-layer bidirectional feature fusion model, which improves the prediction performance and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be 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 labor.

[0037] Figure 1 Schematic diagram of the noise enhancement timing model in the present invention.

[0038] Figure 2 It is a schematic diagram of the bidirectional diffusion feedback space model in the present invention.

[0039] Figure 3 It is a schematic diagram of the bilinear spatiotemporal fusion model in the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0041] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings.

[0042] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0043] The present invention provides a traffic flow prediction method based on a decoupled fusion network. First, a traffic flow prediction model based on a decoupled fusion network is constructed. The traffic flow prediction model includes three parts:

[0044] Part 1, Noise Enhanced Time Series Model: Current prediction methods mainly focus on learning periodic time series features, but have weak adaptability in dealing with the impact of noise. In actual traffic scenarios, traffic time series data can be regarded as a collection of periodic signals and noise. Noise includes random fluctuations and outliers in traffic data, which are usually not covered by periodic patterns. It not only affects the performance of traffic flow prediction, but is also one of the indispensable features in traffic flow prediction. Therefore, the present invention provides a time series prediction method based on noise enhancement, which decouples the periodic features in traffic flow data from noise to alleviate the impact of non-periodic peak traffic on the prediction results.

[0045] Part II, Bidirectional Diffusion Feedback Spatial Model: Using graph convolutional neural networks or their attention mechanisms for spatial modeling is a common approach, but the prediction network becomes more complex and the required computational time becomes larger through operations such as synchronous spatiotemporal local graph convolution, dynamic updating, and multi-scale attention mechanisms, but the improvement is limited. In addition, due to the problems of inaccurate and missing information in the predefined graph itself, some works have proposed adaptive graph learning or dynamic graph learning to overcome these problems, but the designed methods often require the introduction of additional modules, which increases the complexity of the model and the difficulty of capturing spatial dependencies. Therefore, the present invention provides a spatial modeling method that retains the explicit correlation features between nodes in the predefined graph and uses the transmission and feedback between upstream and downstream traffic to describe the characteristics of the road network.

[0046] The third part, the bilinear spatiotemporal fusion model, abandons the traditional graph convolution method and proposes a bilinear spatiotemporal fusion model based on residual multi-layer perceptron to effectively capture spatiotemporal dependencies.

[0047] Then, obtain the open source traffic flow dataset and train the constructed traffic flow prediction model.

[0048] Specifically, in this embodiment, open source data sets PEMS03, PEMS03, PEMS07, and PEMS08 are used, and the data are all collected from the Caltrans Performance Measurement System (PeMS).

[0049] During training, the past 12 consecutive time steps are used to predict the next 12 consecutive time steps, that is, P = S = 12. In the above four data sets, the sampling frequency is the same, and the number of samples in one day is the same, that is, N D =288. The model is trained using Adamoptimizer with an initial learning rate of 0.001. The batch size is set to 32. The spatial feature matrix embedding dimension M = 64, and the temporal feature matrix embedding dimension Q = 32.

[0050] Among them, the loss function

[0051] Where x = x1, ..., x n represents the true value, represents the predicted value, and Ω represents the index of the observed sample. In this embodiment, |Ω|=P=S.

[0052] Finally, the trained traffic flow prediction model is used to predict traffic flow.

[0053] The method for predicting traffic flow of the present invention comprises the following steps:

[0054] Step 1: Obtain traffic flow data and obtain time series features through noise enhancement time series model.

[0055] Taking traffic flow data X as input, the periodic component F is extracted through the noise enhancement time series model Per , then the traffic flow data X and the periodic component F Per The difference between Noi , and finally the periodic component F Per and the noise component F Noi Concatenate to obtain the time series feature F = concat(F Noi , F Per ).

[0056] It can be understood that due to the randomness of noise, the periodic component F can be extracted from the traffic flow data X. Per , and then the noise component F can be obtained by the difference between the two Noi , that is, F Noi =XF Per .

[0057] Specifically, Figure 1As shown in the figure, due to the characteristics of noise, the data of a certain time step containing noise can easily appear "different" in the time series, that is, it shows a weak correlation with the data of other time steps. In comparison, the periodic component shows a more regular or stable trend, and its time series correlation is stronger. This embodiment uses a multi-head attention mechanism to calculate the time series correlation, and obtains the required time series features, that is, the periodic component, based on this correlation.

[0058] Further, based on the above analysis, this embodiment designs a two-stage time series feature processing. Figure 1 As shown in Figure 1, at the beginning of both stages, X captures periodic information through the self-attention module. In order to prevent homogeneity, a residual multi-layer perceptron (RMLP) is added in the first stage. In the second stage, the results of the first stage are used as input and the results of the self-attention module in the second stage are cross-attentioned to further obtain periodic information. Finally, the periodic component F is obtained by sorting it out through RMLP. Per .

[0059] By enhancing the time series model with noise, it is possible to distinguish and extract periodic features rather than simply fitting noise. This mechanism enables the model to identify and strengthen periodic signals while suppressing non-periodic noise components. This step also ensures that the model focuses on periodic traffic patterns during the learning process, thereby improving the accuracy of predictions and the generalization ability of the model. Further, the time series feature F = concat(F Noi , F Per ).

[0060] Step 2: construct a bidirectional diffusion feedback space model consisting of a forward diffusion feedback space model and a reverse diffusion feedback space model. The forward diffusion feedback space model and the reverse diffusion feedback space model have the same structure, both including a two-layer fully connected network and a residual connection.

[0061] The model retains the explicit correlation features between nodes to reflect the changing trend of the associated time series data. The model embeds the predefined graph and learns it, which only contains two layers of fully connected networks and one residual connection. It efficiently and dynamically adjusts the structure of the predefined graph and extracts effective spatial features from it. At the same time, this embodiment takes into account the diffusion process in the reverse direction. The bidirectional spatial features can more effectively and accurately describe the characteristics of the road network and can help the model learn the transmission and feedback of upstream and downstream nodes.

[0062] Step 3: Get the known road network adjacency matrix A and the transpose A of the road network adjacency matrix A T , according to A and A respectively T Get an A for and A back The predefined graph Afor As input, the forward graph G is obtained through the forward diffusion feedback space model. for , the predefined graph A back As input, the reverse graph G is obtained by back-diffusion feedback space model. back .

[0063] Among them, the input is the adjacency matrix representation of the road network, A is the adjacency matrix of the road network, and A T is the transpose of A. A represents: if there is an edge from vertex i to vertex j, then Aij is 1 (or the weight of the edge). If there is no edge from i to j, then Aij is 0.

[0064] Furthermore, A for A is normalized. back Yes T After normalization, A T It is the transpose of A. Because the road network has directionality, transposition means the reverse direction. for is the normalized adjacency matrix, A back The adjacency matrix is ​​transposed and then normalized.

[0065] Specifically, Figure 2 As shown, the predefined graph is the normalized weighted adjacency matrix. The predefined graph is A for =A / rowsum(A), A back =A T / rowsum(A T ). The forward and reverse diffusion feedback spatial models are shown in the figure.

[0066] Specifically, the forward graph G for and the reverse graph G back Method for obtaining: Construct a bidirectional diffusion feedback space model: This model consists of forward and reverse diffusion feedback space models, both of which have the same structure, including two layers of fully connected networks and residual connections.

[0067] Forward diffusion: During the forward diffusion process, the model uses a predefined graph A for As input, feature extraction is performed through the forward diffusion feedback space model. Predefined Figure A for It is a normalized weighted adjacency matrix, which represents the explicit association characteristics between nodes (such as intersections or road sections) in the transportation network.

[0068] Backward Diffusion: Similar to forward diffusion, the backward diffusion process uses a predefined graph A back (Figure A for The transpose of ) is used as input and feature extraction is performed through a back-diffusion feedback space model. This allows the model to capture flow transfer and feedback from downstream to upstream.

[0069] Feature extraction: During forward and backward diffusion, the model extracts spatial features through two layers of fully connected networks and residual connections. Residual connections help the model retain the original information while allowing the network to learn more complex patterns.

[0070] Output features: After being processed by the diffusion feedback space model, the model output G for and G back , representing the spatial features extracted during the forward and backward diffusion processes, respectively.

[0071] Feature fusion: The obtained G for and G back It will be used together with other features in the subsequent bilinear spatiotemporal fusion model for deeper feature fusion and prediction.

[0072] Step 4: construct a bilinear spatiotemporal fusion model, which includes three networks: forward spatiotemporal fusion, reverse spatiotemporal fusion, and overall fusion. Among them, all three networks use multi-layer RMLP as feature fuser.

[0073] Specifically, Figure 3 As shown, when Z is taken as input, its l-th layer bilinear spatiotemporal fusion model can be expressed as:

[0074]

[0075] After the above-mentioned spatiotemporal modeling, the obtained F, T, G for and G back All of them are equally and uniformly regarded as features in the traffic flow prediction task, and a simple but effective bilinear spatiotemporal fusion model is designed to fuse the features.

[0076] Step 5: F, T, G for and G back All the prediction results are equally and uniformly regarded as the features in the traffic flow prediction task, and N prediction results are output through the N-layer bilinear spatiotemporal fusion model. All the prediction results are spliced, and then the final prediction result is obtained through the feature fuser.

[0077] Where T is the time embedding matrix to capture the dependencies of time series data. Let ND be the number of sensor samples in a day, N W =7 is the number of days in a week. Time embedding matrix E D and E W Defined in and In the space, Q represents the embedding dimension. In this embodiment, these embedding matrices are indexed by time identifiers (time points in a day, days in a week) to obtain the corresponding time-dependent representations. Finally, the daily embedding and weekly embedding are concatenated to form a complete time embedding matrix T = concat (T D , T W ).

[0078] Specifically, in the first layer, the forward spatiotemporal fusion and the reverse spatiotemporal fusion parts are respectively fused to obtain bidirectional hidden representations, and then the two are fused through overall fusion. Finally, a layer of FC is used to obtain the result of this layer, which can be expressed as:

[0079] H1=FC(RMLPs(RMLPs(F||G for ||T1)||RMLPs(F||G back ||T1)))

[0080] The result of the upper bilinear spatiotemporal fusion model is used as the new feature of the next bilinear spatiotemporal fusion model, so the lth layer can be expressed as:

[0081] H l+1 =FC(RMLPs(RMLPs(F||G for ||T l ||H l )||RMLPs(F||G bcck ||T l ||H l )))

[0082] Assuming that N layers of bidirectional fusion models are stacked, multi-layer output results can be obtained. Let H = concat(H1, ..., H N ), where H i ∈R N×F×P , combining the results of each layer to generate the final prediction

[0083]

[0084] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A traffic flow prediction method based on decoupled fusion network, It is characterized in that The steps include: Step 1: Obtain traffic flow data and obtain time series features through noise enhancement time series model; Step 2: construct a bidirectional diffusion feedback space model consisting of a forward diffusion feedback space model and a reverse diffusion feedback space model, wherein the forward diffusion feedback space model and the reverse diffusion feedback space model have the same structure, both including a two-layer fully connected network and a residual connection; Step 3: Get the known road network adjacency matrix A and the transpose A of the road network adjacency matrix A T , according to A and A respectively T Get an A for and A back The predefined graph A for As input, the forward graph G is obtained through the forward diffusion feedback space model. for , the predefined graph A back As input, the reverse graph G is obtained through the forward diffusion feedback space model. back ; Step 4: construct a bilinear space-time fusion model, which includes three networks: forward space-time fusion, reverse space-time fusion and overall fusion; Step 5: F, T, G for and G back All the prediction results are equally and uniformly regarded as the features in the traffic flow prediction task, and N prediction results are output through the N-layer bilinear spatiotemporal fusion model. All the prediction results are spliced, and then the final prediction result is obtained through the feature fuser.

2. According to claim 1, a traffic flow prediction method based on a decoupled fusion network, It is characterized in that In step 1, the traffic flow data x is used as input, and the periodic component F is extracted through the noise enhancement time series model. Per , then the traffic flow data x and the periodic component F Per The difference between Noi , and finally the periodic component F Per and the noise component F Noi Concatenate to obtain the time series feature F = concat(F Noi , F Per ).

3. According to claim 1, a traffic flow prediction method based on a decoupled fusion network, It is characterized in that In step 1, the method of extracting the periodic component by using the noise enhanced time series model is: The first stage: input traffic flow data x, and then pass it through a self-attention module and a residual multi-layer perceptron at a time; The second stage: The result of the first stage is used as input and the result of the second stage self-attention module is used for multi-head cross attention. Finally, the periodic component F is obtained through RMLP. Per .

4. According to claim 1, a traffic flow prediction method based on a decoupled fusion network, It is characterized in that In step 3, the predefined graph is a normalized weighted adjacency matrix. for =A / rowsum(A), A back =A r / rowsum(A T ).

5. According to claim 1, a traffic flow prediction method based on a decoupled fusion network, It is characterized in that In step 4, the three networks all use multi-layer RMLP as feature fusion.

6. A traffic flow prediction method based on a decoupled fusion network according to claim 1, It is characterized in that In step 4, when Z is used as input, the l-th layer bilinear spatiotemporal fusion model can be expressed as:

7. A traffic flow prediction method based on a decoupled fusion network according to claim 6, It is characterized in that In step 5, the method for obtaining the prediction result through the bilinear spatiotemporal fusion model is: F, T, G for and G back As input, features are fused through forward spatiotemporal fusion and reverse spatiotemporal fusion respectively to obtain bidirectional hidden representation, and then the two are fused through overall fusion, and finally a layer of FC is used to obtain the result of this layer.

8. According to claim 1, a traffic flow prediction method based on decoupled fusion network, It is characterized in that The step 5 includes the following sub-steps: Step 5.1: In the first layer, the forward spatiotemporal fusion and reverse spatiotemporal fusion parts are fused separately to obtain bidirectional hidden representations, and then the two are fused through overall fusion. Finally, a layer of FC is used to obtain the result of this layer, which can be expressed as: H 1 =FC(RMLPs(RMLPs(F||G for ||T 1 )||RMLPs(F|G back ||T 1 ))) Step 5.2: The result of the upper bilinear spatiotemporal fusion model is used as the new feature of the next bilinear spatiotemporal fusion model. Therefore, the lth layer can be expressed as: H l+1 =FC(RMLPs(RMLPs(F||G for ||T l ||H l )||RMLPs(F||G bcck ||T l ||H l ))) Assuming that N layers of bidirectional fusion models are stacked, multi-layer output results can be obtained. Let H = concat(H 1 , …, H N ), where H i ∈R N×F×P , combining the results of each layer to generate the final prediction Among them, R represents a real number, N represents the number of layers of the bidirectional fusion model, F represents the number of features, P represents the length of the past time series, and S represents the length of the predicted future time series.

Citation Information

Cited By

  • Traffic flow prediction method and system based on diffusion model

    CN119889026A

  • Traffic flow prediction method and system based on graph neural network and diffusion model

    CN120216864A

  • Traffic flow prediction method and system based on graph neural networks and diffusion models

    CN120216864B