Traffic flow prediction method and system based on decoupling graph convolution circulation network

Abstract

The invention discloses a traffic flow prediction method and system based on a decoupling graph convolution circulation network, a multi-scale time embedded encoder, a sensor specific graph convolution network and a signal decoupling mechanism are used, and the method comprises the following steps: S1, inputting historical traffic data, including flow and speed, and carrying out standardized preprocessing; s2, performing adaptive fusion on minute-level, daily-period and weekly-period features of the input historical traffic data through a multi-scale time embedded encoder MTEE to generate a multi-scale spatial-temporal feature matrix; according to the method, the defects of a traditional model in time multi-scale modeling, node heterogeneity adaptation and sudden fluctuation processing are overcome. The prediction effect of the method is obviously superior to that of the prior art, the MAE index is obviously reduced, and the method is suitable for intelligent traffic management, navigation optimization and emergency decision support.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of intelligent transportation systems and deep learning, and in particular to a traffic flow prediction method and system based on a decoupled graph convolutional recurrent network. Background Art

[0002] The existing intelligent transportation system has the following problems: 1. Insufficient modeling of multi-scale time dependencies: Traffic data has minute-level fluctuations (such as sudden congestion), daily cycles (morning and evening rush hours), and weekly cycles (differences between weekdays and weekends). Traditional methods (such as LSTM long short-term memory networks and TCN temporal convolutional networks) require preset fixed time windows or explicit separation of time scales, resulting in insufficient modeling of long-term trends and short-term fluctuations. For example, the STGCN (Spatiotemporal Graph Convolutional Network) model only captures local time patterns through one-dimensional convolution and cannot adaptively fuse multi-scale features.

[0003] 2. Sensor behavior heterogeneity is ignored: Traffic sensors present diverse patterns due to geographical locations (such as highways, commercial areas, residential areas) and environmental factors (weather, events). Existing graph convolution methods (such as GCN graph convolutional neural networks and GraphSAGE) use global shared parameters and assume that all sensor behaviors are homogeneous, resulting in insufficient modeling capabilities for node-level heterogeneity. For example, the DCRNN (Diffused Convolutional Recurrent Network) model introduces diffuse convolution to model spatial dependencies, but relies on a predefined static graph structure; it relies on a static adjacency matrix and cannot dynamically adjust the relationship between sensors.

[0004] 3. Mixed processing of steady-state and non-steady-state signals: Traffic signals include steady state (regular flow) and non-steady state (accidents, sudden fluctuations caused by weather). Traditional models (such as AGCRN adaptive graph convolutional network) do not explicitly separate the two types of signals, resulting in a significant increase in the prediction error of sudden events. Experiments show that in rainy and snowy weather scenarios, the RMSE index (root mean square error) of existing models increases by an average of 23.7%. AGCRN: Generates a dynamic adjacency matrix through node embedding, but lacks multi-scale time embedding and hierarchical signal decoupling mechanism.

[0005] That is, the existing prediction models have poor traffic prediction accuracy in complex traffic scenarios. Summary of the invention

[0006] To this end, it is necessary to provide a traffic flow prediction method and system based on a decoupled graph convolutional recurrent network to solve the problem of poor traffic prediction accuracy in existing complex traffic scenarios.

[0007] To achieve the above object, the present invention provides a traffic flow prediction method based on a decoupled graph convolutional recurrent network, comprising: Step S1: input historical traffic data, including flow rate and speed, and perform standardized preprocessing; Step S2: Adaptively fuse the minute-level, daily cycle and weekly cycle features of the input historical traffic data through the multi-scale time embedding encoder MTEE to generate a multi-scale spatiotemporal feature matrix; Step S3: Based on the sensor-specific graph convolutional network SS-GCN, an independent weight matrix and bias vector are generated for each traffic flow collection sensor, and an adjacency matrix between sensors is dynamically constructed, and spatiotemporal features are extracted according to the multi-scale spatiotemporal feature matrix; Step S4: Through the decoupled graph convolutional recurrent network DGCRN, the steady-state signal and the non-steady-state residual signal are separated layer by layer, and the steady-state signal of each layer and the spatiotemporal features are integrated to generate the final prediction result; Step S5: Output the traffic flow, speed or demand forecast value in the future time window.

[0008] Furthermore, the implementation of the multi-scale time embedding encoder in step S2 includes: Step S21: Set the time point and day of the week Mapping to a learnable high-dimensional embedding space to generate daily and weekly time embedding vectors and , the formula is: , , in, , Refers to the time point and day of the week The associated one-hot encoded cell vector; Refers to the time point and day of the week The associated learnable embedding matrix, Represents a set of real numbers, and represents different sets of data in different formulas; , The time points and day of the week The dimension of is the preset embedding dimension; Step S22: Map the original data to the same dimension as the time embedding vector through the convolution module and concatenate it with the time embedding vector to generate a fused multi-scale feature spatiotemporal matrix , the formula is: , in Indicates that multiple input tensors are concatenated on the specified axis. The daily and weekly time embedding vectors are fused into a unified time representation ， is the high-dimensional feature extracted by one-dimensional convolution, N refers to the number of sensors, C is the dimension after concatenation, and T refers to the number of time steps of the input historical traffic data.

[0009] Furthermore, the implementation of the sensor-specific graph convolutional network SS-GCN in step S3 includes: Step S31: Generate an independent weight matrix for each sensor node and the bias vector , the formula is: , , in It is the low-dimensional embedding matrix of the sensor, which is used to represent the personalized characteristics of different traffic flow collection sensors; , is a shared parameter; d is the feature dimension of each node, C and F represent the input and output dimensions of SS-GCN respectively; Step S32: Based on the sensor embedding matrix , dynamically construct the adjacency matrix , the formula is: , in, is the preset embedding dimension size of the sensor node, is an activation function, which is used to normalize a numerical vector into a probability distribution vector, and the sum of each probability is 1; is the activation function in the neural network; For the matrix The transposed matrix of Step S33: Update the spatiotemporal features of the nodes through dynamic graph convolution: , in is the identity matrix, which is used to preserve the node's own characteristics.

[0010] Furthermore, the implementation of the decoupled graph convolutional recurrent network in step S4 includes: Step S41: Use the sensor gating module to calculate the update gate and reset gate , and its gate calculation is: , , Where: t represents the tth time step, represents the hidden state of the previous time step, σ is the Sigmoid function; Step S42: Use reset gate To calculate the candidate hidden state : , in, is the hyperbolic tangent function, represents the input at time step t, and is a learnable parameter, represents the Hadamard product. To incorporate the residual connection, the intermediate state is calculated as: , Finally, use the update gate To update the hidden state , thereby maintaining a balance between the previous hidden state and the newly computed residual state: , Step S43: The signal decoupling unit SDU decouples the steady-state and non-steady-state signals from the output of the sensor-specific gating module SSGM, thereby allowing the DGCRN to model; the decoupling process begins by using the decoupling gate to decouple the hidden state at layer l Extracting steady-state signals , The layer position for the DGCRN layer is: , in is a transformation function that isolates the stable features of the traffic data; the residual gate computes the non-stationary components by subtracting the steady-state signal from the input: , Refers to the current layer, i.e. The input of the layer; Refers to the next layer, namely The input of the layer; Step S44: Layered prediction fusion: weighted sum of the steady-state signals of each layer to generate the final prediction result , which is the future road traffic data, where L represents the total number of DGCRN layers: .

[0011] The present invention also provides a traffic flow prediction system based on a decoupled graph convolutional recurrent network, comprising a memory and a processor, wherein a computer program is stored on the memory, and when the computer program is executed by the processor, the steps of any method described in the present invention are implemented.

[0012] Different from the existing technology, the above technical solution proposes a new decoupled graph convolutional recurrent network (GDGCRN, General Decoupled Graph Convolutional Recurrent Network), which adaptively fuses minute, day and week period features through a multi-scale time embedding encoder; then generates independent parameters for each sensor through a sensor-specific graph convolutional network to construct a dynamic adjacency matrix; finally, the decoupled graph convolutional recurrent network is used to hierarchically separate steady-state and non-steady-state signals, thereby improving the adaptability to emergencies and the accuracy of traffic prediction in complex traffic scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the method of the present invention; Figure 2 A schematic diagram of a multi-scale cycle of traffic flow according to the present invention; Figure 3 This is the overall architecture diagram of the decoupled graph convolutional recurrent network based on the present invention. DETAILED DESCRIPTION

[0014] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0015] See also Figures 1 to 3, the present invention provides a new decoupled graph convolutional recurrent network (GDGCRN for short), which includes three core modules: Multiscale Temporal Feature Encoder (MTEE): Adaptively integrate minute, day, and week period features. Sensor-Specific Graph Convolutional Network (SS-GCN): Generate independent parameters for each sensor and build a dynamic adjacency matrix. Decoupled Graph Convolutional Recurrent Network (DGCRN): Hierarchically separate steady-state and non-steady-state signals to improve adaptability to emergencies. Based on the above-mentioned new decoupled graph convolutional recurrent network, the present invention provides a traffic flow prediction method based on a decoupled graph convolutional recurrent network, including: step S1: input historical traffic data, including flow and speed, and perform standardized preprocessing; step S2: adaptively fuse the minute-level, daily cycle and weekly cycle features of the input historical traffic data through a multi-scale time embedding encoder MTEE to generate a multi-scale spatiotemporal feature matrix; step S3: based on a sensor-specific graph convolutional network SS-GCN, generate an independent weight matrix and bias vector for each traffic flow collection sensor, and dynamically construct an adjacency matrix between sensors, and extract spatiotemporal features according to the multi-scale spatiotemporal feature matrix; step S4: through a decoupled graph convolutional recurrent network DGCRN, separate the steady-state signal and the non-steady-state residual signal layer by layer, and generate the final prediction result after fusing the steady-state signal of each layer with the spatiotemporal features; step S5: output the traffic flow, speed or demand prediction value in the future time window.

[0016] This scheme uses a multi-scale time embedding encoder to adaptively fuse minute, day, and week periodic features; then generates independent parameters for each sensor through a sensor-specific graph convolutional network to construct a dynamic adjacency matrix; finally, a decoupled graph convolutional recurrent network is used to hierarchically separate steady-state and non-steady-state signals, thereby improving adaptability to emergencies and the accuracy of traffic prediction in complex traffic scenarios.

[0017] Traffic flow prediction diagram Figure 2The figure shows the change of traffic flow trends over time. The horizontal axis represents the days of the week, the vertical axis represents the traffic flow, and the background color represents different time periods, namely weekdays (light orange) and weekends (red). The daily traffic peaks and troughs are roughly the same, showing a regular fluctuation pattern. There are also significant differences in weekly traffic patterns, especially between weekdays and weekends. On weekdays (Monday to Friday, light orange background), traffic flow varies greatly, and peaks usually occur during the morning and afternoon rush hours. On weekends (Saturday and Sunday, red background), overall traffic flow is reduced, with lower peaks and less fluctuation.

[0018] In certain embodiments, such as Figure 3 As shown, it includes the multi-scale time embedding encoder MTEE in step S2, which inputs: the minute, hour, and week information of the timestamp t. Then step S21: time embedding mapping is performed, including daily cycle embedding and weekly cycle embedding. The daily cycle embedding is: merging the hour and minute into a 1440-dimensional daily time point ToD(t), which represents the current time point Relative position in the day, via a learnable embedding matrix Map it into a high-dimensional vector representation .in, For time point For example, it can represent the total number of time periods in a day (e.g., 288 time periods corresponding to every 5 minutes in a day), For the preset embedding dimension, the formula for daily embedding is: .

[0019] For the weekly cycle, embed the day-of-the-week feature DoW( ) represents the day of the week corresponding to the current time step. By embedding the matrix Map it to a high-dimensional representation : .

[0020] Then feature fusion is performed, and Add them together to get the fused time embedding table.

[0021] ,in, , Refers to the time point and day of the week The associated one-hot encoded cell vector; Refers to the time point and day of the week The associated learnable embedding matrix, Represents a set of real numbers, and represents different sets of data in different formulas; , The time points and day of the week The dimension of is the preset embedding dimension; Step S22: Map the original data to the same dimension as the time embedding vector through the convolution module and concatenate it with the time embedding vector to generate a fused multi-scale feature spatiotemporal matrix , the formula is: , in Indicates that multiple input tensors are concatenated on the specified axis. The daily and weekly time embedding vectors are fused into a unified time representation ， is the high-dimensional feature extracted by one-dimensional convolution, N refers to the number of sensors, C is the dimension after concatenation, and T refers to the number of time steps of the input historical traffic data.

[0022] Furthermore, if Figure 3 As shown, the implementation of the sensor-specific graph convolutional network SS-GCN in step S3 includes: node parameter independence, including step S31: generating an independent weight matrix for each sensor node and the bias vector , the formula is: , , in It is the low-dimensional embedding matrix of the sensor, which is used to represent the personalized characteristics of different traffic flow collection sensors; , is a shared parameter; d is the feature dimension of each node, C and F represent the input and output dimensions of SS-GCN respectively; although the parameters are shared, each node can still obtain unique weights and biases through the diversity of node embeddings, thereby maintaining the model's ability to express the characteristics of different nodes.

[0023] Then, a dynamic adjacency matrix is ​​constructed, including step S32: based on the sensor embedding matrix , dynamically construct the adjacency matrix , the formula is: , in, is the preset embedding dimension size of the sensor node, is an activation function, which is used to normalize a numerical vector into a probability distribution vector, and the sum of each probability is 1; is the activation function in the neural network; For the matrix The transposed matrix of .

[0024] Finally, in step S33, the spatiotemporal features of the nodes are updated through dynamic graph convolution: , in is the identity matrix, which is used to preserve the node's own characteristics.

[0025] In certain embodiments, such as Figure 3 As shown, the implementation of the decoupled graph convolutional recurrent network in step S4 includes: step S41: using the sensor gating module to calculate the update gate and reset gate , using dynamic graph convolution to replace the update gate and reset gate of the traditional GRU, its gating calculation is: , , Where: t represents the tth time step, represents the hidden state of the previous time step, σ is the Sigmoid function; Step S42: Use reset gate To calculate the candidate hidden state : , in, is the hyperbolic tangent function, represents the input at time step t, and is a learnable parameter, represents the Hadamard product. In order to incorporate the residual connection, the intermediate state is calculated as: , Finally, use the update gate To update the hidden state , thereby maintaining a balance between the previous hidden state and the newly computed residual state: , Step S43: The signal decoupling unit SDU decouples the steady-state and non-steady-state signals from the output of the sensor-specific gating module SSGM, thereby allowing the DGCRN to model; the decoupling process begins by using the decoupling gate to decouple the hidden state at layer l Extracting steady-state signals , The layer position for the DGCRN layer is: , in is a transformation function that isolates the stable features of the traffic data; the residual gate computes the non-stationary components by subtracting the steady-state signal from the input: , Refers to the current layer, i.e. The input of the layer; Refers to the next layer, namely The input of the layer; Step S44: Hierarchical prediction fusion: non-stationary component It is used as the input of the next layer of DGCRN. Finally, the weighted sum of the steady-state signals of each layer is performed to generate the final prediction result. , which is the future road traffic data, where L represents the total number of DGCRN layers: .

[0026] In order to comprehensively evaluate the performance of the proposed GDGCRN model, we compared it with three traditional statistical baselines and thirteen other deep learning models. These baselines and models cover a wide range of methods, from simple historical average-based methods to advanced graph-based spatiotemporal attention mechanisms, providing a comprehensive benchmark for our method.

[0027] Table 1 summarizes the prediction performance of the proposed GDGCRN model compared with 16 baseline models on four benchmark datasets (PeMS 04, PeMS 08, NYCBike, and PeMSD 7(M)). The table reports three widely used evaluation metrics: MAE, RMSE, and MAPE.

[0028] The GDGCRN model of the present invention consistently outperforms all baseline models on the four datasets and achieves the best results in all indicators.

[0029] Table 1 Prediction performance table

[0030] For example, on PeMS04, our GDGCRN model records a MAE of 18.35, which is 4.6% lower than the second best method (AFDGCN, MAE: 19.09). On NYCBike, our GDGCRN model achieves a MAE of 2.09, surpassing AGCRN (MAE: 2.16) and demonstrating its ability to capture fine-grained spatiotemporal dependencies. The average MAE, RMSE, and MAPE values ​​of our GDGCRN model on all datasets are 9.35, 15.80, and 20.36%, respectively. The corresponding values ​​for AFDGCN are 9.82 (+4.79%), 16.21 (+2.53%), and 21.90% (+7.03%), while those for SSGCRTN are 9.89 (+5.46%), 16.16 (+2.23%), and 21.82% (+6.68%). These results show that the performance of the GDGCRN model of the present invention is consistently better than that of AFDGCN and SSGCRTN, with improvements ranging from 2.23% to 7.03%, highlighting the advantages of the GDGCRN model of the present invention in simulating the spatial and temporal correlations of traffic data.

[0031] The present invention also provides a traffic flow prediction system based on a decoupled graph convolutional recurrent network, including a memory and a processor, wherein the storage medium stores a computer program, and the computer program implements the steps of the above method when executed by the processor. The storage medium of this embodiment can be a storage medium set in an electronic device, and the electronic device can read the content of the storage medium and achieve the effects of the present invention. The storage medium can also be a separate storage medium, and the storage medium is connected to an electronic device, so that the electronic device can read the content in the storage medium and implement the method steps of the present invention. The prediction system of the present invention can improve the accuracy of traffic prediction in complex traffic scenarios.

[0032] It should be noted that, although the above embodiments have been described in this article, the patent protection scope of the present invention is not limited thereby. Therefore, based on the innovative concept of the present invention, changes and modifications made to the embodiments described herein, or equivalent structures or equivalent process changes made using the contents of the present invention specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included in the patent protection scope of the present invention.

Claims

1. A traffic flow prediction method based on a decoupled graph convolutional recurrent network, characterized in that: include: Step S1: input historical traffic data, including flow rate and speed, and perform standardized preprocessing; Step S2: Adaptively fuse the minute-level, daily cycle and weekly cycle features of the input historical traffic data through the multi-scale time embedding encoder MTEE to generate a multi-scale spatiotemporal feature matrix; Step S3: Based on the sensor-specific graph convolutional network SS-GCN, an independent weight matrix and bias vector are generated for each traffic flow collection sensor, and an adjacency matrix between sensors is dynamically constructed, and spatiotemporal features are extracted according to the multi-scale spatiotemporal feature matrix; Step S4: Through the decoupled graph convolutional recurrent network DGCRN, the steady-state signal and the non-steady-state residual signal are separated layer by layer, and the steady-state signal of each layer and the spatiotemporal features are integrated to generate the final prediction result; Step S5: Output the traffic flow, speed or demand forecast value in the future time window.

2. The traffic flow prediction method based on decoupled graph convolutional recurrent network according to claim 1 is characterized by: The implementation of the multi-scale time embedding encoder in step S2 includes: Step S21: Set the time point and day of the week Mapping to a learnable high-dimensional embedding space to generate daily and weekly time embedding vectors and , the formula is: ， ， in, , Refers to the time point and day of the week The associated one-hot encoded cell vector; Refers to the time point and day of the week The associated learnable embedding matrix, Represents a set of real numbers, and represents different sets of data in different formulas; , The time points and day of the week The dimension of is the preset embedding dimension; Step S22: Map the original data to the same dimension as the time embedding vector through the convolution module and concatenate it with the time embedding vector to generate a fused multi-scale feature spatiotemporal matrix , the formula is: ， in Indicates that multiple input tensors are concatenated on the specified axis. The daily and weekly time embedding vectors are fused into a unified time representation ， is the high-dimensional feature extracted by one-dimensional convolution, N refers to the number of sensors, C is the dimension after concatenation, and T refers to the number of time steps of the input historical traffic data.

3. The traffic flow prediction method based on decoupled graph convolutional recurrent network according to claim 2 is characterized by: The implementation of the sensor-specific graph convolutional network SS-GCN in step S3 includes: Step S31: Generate an independent weight matrix for each sensor node and the bias vector , the formula is: ， ， in It is the low-dimensional embedding matrix of the sensor, which is used to represent the personalized characteristics of different traffic flow collection sensors; , is a shared parameter; d is the feature dimension of each node, C and F represent the input and output dimensions of SS-GCN respectively; Step S32: Based on the sensor embedding matrix , dynamically construct the adjacency matrix , the formula is: ， in, is the preset embedding dimension size of the sensor node, is an activation function, which is used to normalize a numerical vector into a probability distribution vector, and the sum of each probability is 1; is the activation function in the neural network; For the matrix The transposed matrix of Step S33: Update the spatiotemporal features of the nodes through dynamic graph convolution: ， in is the identity matrix, which is used to preserve the node's own characteristics.

4. The traffic flow prediction method based on decoupled graph convolutional recurrent network according to claim 3 is characterized by: The implementation of the decoupled graph convolutional recurrent network in step S4 includes: Step S41: Use the sensor gating module to calculate the update gate and reset gate , and its gating calculation is: ， ， Where: t represents the tth time step, represents the hidden state of the previous time step, σ is the Sigmoid function; Step S42: Use reset gate To calculate the candidate hidden state : ， in, is the hyperbolic tangent function, represents the input at time step t, and is a learnable parameter, represents the Hadamard product. To incorporate the residual connection, the intermediate state is calculated as: ， Finally, use the update gate To update the hidden state , thereby maintaining a balance between the previous hidden state and the newly computed residual state: ， Step S43: The signal decoupling unit SDU decouples the steady-state and non-steady-state signals from the output of the sensor-specific gating module SSGM, thereby allowing the DGCRN to model; the decoupling process begins by using the decoupling gate to decouple the hidden state at layer l Extracting steady-state signals , The layer position for the DGCRN layer is: ， in is a transformation function that isolates the stable features of the traffic data; the residual gate computes the non-stationary components by subtracting the steady-state signal from the input: ， Refers to the current layer, i.e. The input of the layer; Refers to the next layer, namely The input of the layer; Step S44: Layered prediction fusion: weighted sum of the steady-state signals of each layer to generate the final prediction result , which is the future road traffic data, where L represents the total number of DGCRN layers: 。 5. Traffic flow prediction system based on decoupled graph convolutional recurrent network, characterized by: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images