Traffic flow prediction method based on improved space-time diagram convolutional neural network

Abstract

The invention relates to a traffic flow prediction method based on an improved space-time diagram convolutional neural network, and the method comprises the following steps: 1), carrying out the graph learning through employing a gating circulation unit and an attention mechanism, and obtaining traffic flow data represented by an adjacent matrix; 2) converting the data into a spectral domain for learning by using operations of graph Fourier transform and discrete Fourier transform; and 3) using a graph convolutional neural network and a gating linear unit to carry out learning prediction on data in a spectral domain, and providing data support for intelligent traffic management. Compared with the prior art, the method has the advantages that the prediction precision of the model in the face of a complex road topological structure is improved, and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic prediction, and in particular to a traffic flow prediction method based on an improved spatiotemporal graph convolutional neural network. Background Art

[0002] Traffic forecasting, as an important part of intelligent transportation systems (ITSs), can provide a scientific basis for the management and planning of urban transportation networks. Transportation departments can deploy and guide traffic flow in advance according to expected traffic conditions, improve driving conditions, and alleviate traffic congestion. Accurate traffic flow forecasting based on historical time series data is a daunting challenge because it often requires the collection of complex spatial and temporal relationships, and requires modeling trends within time series and correlations between series at the same time.

[0003] With the development of mathematical theory and modern technology, further improving the accuracy of traffic prediction means considerable economic and safety benefits for road management. These potential benefits have promoted the continuous improvement of traffic prediction models. Among them, real-time traffic prediction can provide reasonable traffic guidance for intelligent transportation systems while reducing prediction delays, and reduce the operating pressure of vehicle commuting, which are all issues of concern to traffic management departments. Therefore, intelligent transportation systems need a new method that can reflect traffic status in real time and provide an intuitive reference range.

[0004] Graph Neural Networks (GNNs) have achieved considerable success in processing traffic data due to their superiority in processing graph data. Because GNNs pass information throughout the structure, each node in the graph can access information about its nearby environment. Modeling multivariate time series data using graph neural networks is a possible approach that can depict the variables in the sequence as nodes in a graph connected by hidden dependencies, while capturing the interdependencies and retaining the temporal trajectory of the sequence. Among the existing complex models, graph convolutional networks (GCNs) are often used to describe the spatial correlation of graph data. For traffic prediction, T-GCN and STGCN directly stack GCN with temporal modules (such as LSTM, GRU), which specifically record temporal patterns in the time domain and increase the model's perception of the spatiotemporal characteristics of the data. However, existing methods often rely on pre-set road topology structures, otherwise they cannot perform in practice. Summary of the invention

[0005] The purpose of the present invention is to get rid of the dependence on the pre-set road topology structure and improve the prediction accuracy of the model when facing complex road topology structures, and to provide a traffic flow prediction method based on an improved spatiotemporal graph convolutional neural network.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A traffic flow prediction method based on an improved spatiotemporal graph convolutional neural network comprises the following steps:

[0008] Step S1, input the traffic flow data X into the preprocessing module, filter and obtain historical traffic data, and transmit it to the graph learning module of the model, wherein the historical traffic data includes traffic information at each moment;

[0009] Step S2, the gated recurrent unit in the graph learning module calculates the hidden state of each time step t in turn according to the historical traffic data. For the current node, the gated recurrent unit uses the hidden state at time t-1 output by the previous node and the traffic information of the current node as the input of the current node, and obtains the hidden state at time t and the output of the current node. The gated recurrent unit obtains the gating signal of the control gate and update gate of the current node according to the hidden state passed down by the previous node and the traffic information of the current node;

[0010] Step S3, after obtaining the gate signal, use the reset gate to obtain the hidden state after reset, then splice the hidden state after reset and the traffic information of the current node, and then use the activation function to scale the spliced ​​data;

[0011] Step S4, using the update gate to memorize the scaled data, and forget the hidden state at time t-1, to obtain the updated hidden state of the current node, and the updated hidden state of the current node is input into the next node, and so on;

[0012] Step S5, using the self-attention mechanism to calculate the adjacency matrix A according to the hidden state output by the last node to represent the complete traffic data;

[0013] Step S6, after obtaining the adjacency matrix A, the graph learning module processes the adjacency matrix A into a unidirectional link form W;

[0014] Step S7, input the one-way link form adjacency matrix W obtained by the graph learning module and the historical traffic data into the prediction module, perform graph Fourier transform through the graph Laplace matrix, and input the result of the graph Fourier transform into the graph convolution;

[0015] Step S8, multiple graph convolutions are superimposed through the first-order approximation of graph Laplace to obtain a hierarchical linear formula to restore the deep spatial information and the historical traffic data after graph convolution;

[0016] Step S9, decomposing the historical traffic data after graph convolution into a frequency basis through discrete Fourier transform, and inputting it into a gated linear unit, learning the characteristics of the historical traffic data based on the frequency basis, and making predictions;

[0017] Step S10, performing inverse discrete Fourier transform and inverse image Fourier transform on the prediction result output by the gated linear unit;

[0018] Step S11, evaluate the transformed prediction result. If it does not meet the requirements of the loss function, return to step S7 until the loss function meets the requirements, and use the prediction result at this time as the final prediction result.

[0019] Furthermore, the signals of the control gate and the update gate are expressed as:

[0020] r=σ(W r )·[h t-1 ,x t ]

[0021] z=σ(W z )·[h t-1 ,x t ]

[0022] Among them, x t is the traffic information of the current node, h t-1 is the hidden state passed down from the previous node, σ is the activation function, W r is the weight of the control gate, W z is the weight of the update gate, r is the control gate signal, and z is the signal of the update gate.

[0023] Furthermore, the activation function used in S3 is the tanh function.

[0024] Furthermore, the hidden state after reset is:

[0025] h′=tanh(W h′ [h t-1 ⊙r,x t ])

[0026] Among them, h′ is the hidden state after reset, W h′ is the weight of the hidden state after reset, h t-1 is the hidden state passed down from the previous node, x t It is the traffic information of the current node.

[0027] Furthermore, the hidden state of the current node after update in S4 is:

[0028] h″=h t-1 ⊙z+h′⊙(1-z)

[0029] Among them, h″ is the hidden state of the current node after update, h′ is the hidden state after reset, z is the gate signal, that is, the signal of the update gate, h t-1 It is the hidden state passed down from the previous node.

[0030] Furthermore, the adjacency matrix A is:

[0031]

[0032] Q=h f 'A Q

[0033] K=h f 'A K

[0034] Among them, Q and K represent the query and key representation respectively, A Q and A K is the learning parameter of the self-attention mechanism, d is the hidden dimension size of Q and K, and h f ′ is the last hidden state obtained.

[0035] Furthermore, the one-way link form is:

[0036]

[0037] Among them, ReLU and tanh are activation functions, A1 and A2 represent the adjacency matrices of randomly initialized node embeddings learned during training, and α is a hyperparameter activation function that controls the node saturation rate.

[0038] Furthermore, graph convolution is expressed as:

[0039] Θ(L)x=Θ(UΛU T )x=UΘ(Λ)U T x

[0040] Among them, Θ is the convolution kernel, U is the graph Fourier basis, Λ is the eigenvalue diagonal matrix of the normalized graph Laplace, L is the graph Laplace matrix, and the graph Laplace matrix is:

[0041] L=UΛU T

[0042] Where U is the graph Fourier basis and Λ is the diagonal matrix of eigenvalues ​​of the normalized graph Laplace.

[0043] Furthermore, in S9, the historical traffic data after graph convolution is decomposed into a frequency basis through discrete Fourier transform. The frequency basis has a real part P and an imaginary part Q. The real part P and the imaginary part Q are respectively the inputs of the gates in the gated linear unit.

[0044] Furthermore, the loss function is:

[0045]

[0046] Where W θis a trainable parameter in the model; x t+1 is the true value, y represents the predicted value of the model, and x t is the traffic information of the current node, M is the length of the temporal convolution input of each node, and L is the loss function.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention adopts gated recurrent unit (GRU) and attention mechanism for graph learning, obtains traffic flow data represented by adjacency matrix, obtains hidden state of each node, forgets unimportant data, adds information of current node, and then calculates adjacency matrix by self-attention mechanism. It can automatically learn graph structure from traffic flow data, does not rely on predefined road topology, and converts data form by spectrum transformation, fully learns spatiotemporal features, and improves prediction accuracy of the model when facing complex road topology. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0050] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0051] The present invention proposes a traffic flow prediction method based on an improved spatiotemporal graph convolutional neural network. The flow chart of the method is as follows: Figure 1 As shown. The method of the present invention can automatically learn the graph structure from the traffic flow data, input the graph structure into the prediction module in the form of an adjacency matrix, and the prediction module converts the data into the spectral domain through spectral transformation before learning and predicting, further obtaining the spatiotemporal features therein, and then obtaining a more reasonable prediction result. The method of the present invention mainly includes three parts: 1) using the gated recurrent unit (GRU) and the attention mechanism to perform graph learning to obtain traffic flow data represented by an adjacency matrix; 2) using the graph Fourier transform (GFT) and discrete Fourier transform (DFT) operations to convert the data into the spectral domain for learning; 3) using the graph convolutional neural network (GCN) and the gated linear unit (GLU) to learn and predict the data in the spectral domain, providing data support for intelligent traffic management.

[0052] The more specific steps of the present invention are as follows:

[0053] Step S1, input the traffic flow data X into the preprocessing module, filter out the historical traffic data, and transmit it to the graph learning module of the model;

[0054] Step S2, the GRU model in the graph learning module calculates the hidden state of each time step t in turn according to the input X. Based on the hidden state at time t-1 and the current traffic information as input, the traffic feature R at time t is obtained. The parameters of the GRU model are as follows:

[0055] r=σ(W r )·[h t-1 ,x t ]

[0056] z=σ(W z )·[h t-1 ,x t ]

[0057] Among them, x t is the current input, h t-1 It is the hidden state passed down from the previous node. This hidden state contains the relevant information of the previous node. GRU combined with x t and h t-1 You can get x' t and h t , the former is the output of the current node, and the latter is the state that needs to be passed to the next node. GRU passes the hidden state h from the previous node t-1 and the input x of the current node t To obtain two gated states, r represents the control gate and z represents the update gate;

[0058] Step S3, after obtaining the gate control signal, use the reset gate to obtain the reset data h′ t-1 =h t-1 ⊙r, then h′ t-1 With x t Concatenate them together, and then use the activation function tanh to scale the data to the range of -1 to 1 to obtain h′. The formula is as follows:

[0059] h′=tanh(W h′ [h t-1 ⊙r,x t ])

[0060] Where ⊙ represents the matrix element multiplication. h′ It mainly includes the current input x t , purposefully add h′ to the hidden state at the current moment to complete the memory of the current state;

[0061] Step S4 uses both forgetting and remembering steps at the same time, using the previously obtained update gate z, and the formula is as follows:

[0062] h″=h t-1 ⊙z+h′⊙(1-z)

[0063] The range of the gate signal z is between 0 and 1. The closer z is to 0, the more forgotten, and the closer z is to 1, the more remembered. t-1 h′⊙(1-z) means to selectively remember h′, which contains the node information at the current moment. Similar to the previous example, 1-z here will also forget some unimportant information in the h′ dimension. That is, the operation of this step is to forget the h′ that has been passed down. t-1 Some dimensional information in the , and add some dimensional information of the current node input;

[0064] Step S5, based on the last hidden state h f ′ is used to represent the complete traffic data, and the adjacency matrix A is calculated using the self-attention mechanism. The formula is as follows;

[0065]

[0066] Among them, Q and K represent the representation of queries and keys, respectively, which can be obtained by using the attention mechanism with learnable parameters A Q and A K The linear projection of is generated, d is the hidden dimension size of Q and K;

[0067] Step S6: After obtaining the adjacency matrix A, the graph learning module processes it into a one-way link form, and the formula is as follows:

[0068]

[0069] Where A1 and A2 represent the adjacency matrix of randomly initialized node embeddings learned during training, and α is a hyperparameter activation function that controls the node saturation rate. The ReLU activation function and the subtraction term regularize the adjacency matrix so that when w ij When it is positive, its diagonal corresponding term w ji will be zero. While making the adjacency matrix sparse, it reduces the computational cost of subsequent graph convolution;

[0070] Step S7: Input the adjacency matrix W and traffic data X obtained by the graph learning module into the prediction module. T Perform a graph Fourier transform on it and input it into the graph convolution at the same time. The formula is as follows:

[0071] Θ(L)x=Θ(UΛU T )x=UΘ(Λ)U T x

[0072] Among them, Θ is the convolution kernel, and the concept of spectral convolution is the input The product of the convolution kernel Θ. Figure Fourier basis is the normalized graph Laplacian is a diagonal matrix of eigenvalues ​​of L, and the filter Θ(Λ) is also a diagonal matrix.

[0073] In step S8, multiple graph convolution modules are superimposed through the first-order approximation of graph Laplacian to obtain a hierarchical linear formula to restore the deep spatial information. The formula is as follows:

[0074]

[0075] Among them, I n is the identity matrix, D ii =Σ j W ij The diagonal matrix of θ is the parameter of the convolution kernel, and W and D are respectively and To be reorganized.

[0076] Step S9, decompose the traffic data X after the graph convolution module into frequency basis through discrete Fourier transform, and input it into the gated linear unit to learn their feature representation. The formula is as follows:

[0077]

[0078] The output of the DFT has a real part P and an imaginary part Q, which are processed in parallel by the same operator with different parameters. For each node in the graph, GLU explores K of the input elements without padding. t neighborhood, which results in shortening the sequence length by K each time t-1 . Therefore, the temporal convolution input of each node can be regarded as a sequence of length M. This design uses the entire convolution structure on the time axis to capture the temporal dynamic behavior of traffic flow, allowing a hierarchical representation to be formed through a multi-layer convolution structure, thereby achieving a parallel and controllable training process. At the same time, P and Q are the inputs of the gates in GLU respectively; ⊙ represents the Hadamard product calculated element by element. The input P of the current state is crucial to defining the composition structure of the time series, while the dynamic fluctuations of the time series are controlled by the sigmoid gate σ(Q).

[0079] Step S10, performing inverse discrete Fourier transform and inverse image Fourier transform on the result obtained by GLU;

[0080] Step S11, finally, use L2 loss and fully connected layer as output layer to evaluate the effectiveness of the model, the formula is as follows:

[0081]

[0082] Where W θis a trainable parameter in the model; x t+1 is the true value, and y is the predicted value of the model.

[0083] Step S12, the prediction result is calculated again using the evaluation index, and the prediction result and the evaluation index result are output for analysis.

[0084] The method of the present invention can automatically learn graph structures from traffic flow data without relying on predefined road topology maps. It also combines GCN and GLU to further obtain spatiotemporal features in the spectral domain, and performs approximate calculations in this process. While improving prediction accuracy, it effectively reduces the computational overhead of the model, providing data support for coordinated management of intelligent transportation systems.

[0085] The following experiment verifies the method of the present invention:

[0086] 1. Experimental Setup

[0087] The network is trained on a 64-bit PC using Pytorch with 24GB RAM and NVIDIA RTX2080Ti GPU. We compare the proposed EST-GCN model with the following state-of-the-art methods, including: TCN, LSTNet, DeepState, DCRNN, STGCN, T-GCN, Graph WaveNet. Using grid search, we fine-tune the hyperparameters on the EST-GCN validation data. Finally, the channel size of each graph convolution layer is set to 64, and the kernel size of 1D convolution is set to 3. Next, we choose the RMSprop optimizer and decide to set 50 training epochs. The learning rate starts from 0.001 and decays at a rate of 0.7 every 5 epochs. The input data of the experiment is normalized to the range [0,1]. In addition, we split the dataset into three parts in a ratio of 7:2:1 for training, validation, and testing. We predict the traffic flow for the next 15 minutes, 30 minutes, and 45 minutes.

[0088] 2. Dataset

[0089] PeMSD7: It is collected in real time by the Caltrans Performance Measurement System (PeMS) in major areas of the California State Freeway System, and contains more than 39,000 sensor stations. The dataset aggregates data samples every 30 seconds into 5-minute intervals. We selected 228 stations in District 7 of California and marked them as PeMSD7. The time range of the PeMSD7 dataset is weekdays in May and June 2012.

[0090] 3. Experimental results analysis:

[0091] The experimental results are shown in Table 1.

[0092] Table 1 Prediction results of different methods

[0093]

[0094] In terms of baseline models, TCN, LSTNet, and DeepState mainly perform time domain estimation for time input of sequence modeling. They cannot clearly capture the relationship between different time series, which is a common defect that hinders their expression in traffic prediction. Therefore, they perform poorly in a single experiment, especially as the prediction time length increases. On the other hand, GNN-based models, such as DCRNN, STGCN, and Graph WaveNet, can simulate spatial and temporal correlations. Among them, STGCN combines graph convolution with gated temporal convolution and achieves good results in spatiotemporal convolution blocks. However, in order to obtain the best results, these models must be used with the predefined Figure 1 Without a specified graph, these techniques cannot effectively transfer data. EST-GCN can automatically learn the graph structure of the data. Experiments show that this learning method can further improve the learning of data features and obtain more accurate experimental results.

[0095] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A traffic flow prediction method based on an improved spatiotemporal graph convolutional neural network, characterized in that: The method comprises the following steps: Step S1, input the traffic flow data X into the preprocessing module, filter and obtain historical traffic data, and transmit it to the graph learning module of the model, wherein the historical traffic data includes traffic information at each moment; Step S2, the gated recurrent unit in the graph learning module calculates the hidden state of each time step t in turn according to the historical traffic data. For the current node, the gated recurrent unit uses the hidden state at time t-1 output by the previous node and the traffic information of the current node as the input of the current node, and obtains the hidden state at time t and the output of the current node. The gated recurrent unit obtains the gating signal of the control gate and update gate of the current node according to the hidden state passed down by the previous node and the traffic information of the current node; Step S3, after obtaining the gate signal, use the reset gate to obtain the hidden state after reset, then splice the hidden state after reset and the traffic information of the current node, and then use the activation function to scale the spliced ​​data; Step S4, using the update gate to memorize the scaled data, and forget the hidden state at time t-1, to obtain the updated hidden state of the current node, and the updated hidden state of the current node is input into the next node, and so on; Step S5, using the self-attention mechanism to calculate the adjacency matrix A according to the hidden state output by the last node to represent the complete traffic data; Step S6, after obtaining the adjacency matrix A, the graph learning module processes the adjacency matrix A into a unidirectional link form W; Step S7, input the one-way link form adjacency matrix W obtained by the graph learning module and the historical traffic data into the prediction module, perform graph Fourier transform through the graph Laplace matrix, and input the result of the graph Fourier transform into the graph convolution; Step S8, multiple graph convolutions are superimposed through the first-order approximation of graph Laplace to obtain a hierarchical linear formula to restore the deep spatial information and the historical traffic data after graph convolution; Step S9, decomposing the historical traffic data after graph convolution into a frequency basis through discrete Fourier transform, and inputting it into a gated linear unit, learning the characteristics of the historical traffic data based on the frequency basis, and making predictions; Step S10, performing inverse discrete Fourier transform and inverse image Fourier transform on the prediction result output by the gated linear unit; Step S11, evaluate the transformed prediction result. If it does not meet the requirements of the loss function, return to step S7 until the loss function meets the requirements, and use the prediction result at this time as the final prediction result.

2. According to claim 1, a traffic flow prediction method based on an improved spatiotemporal graph convolutional neural network is characterized in that: The signals of the control gate and update gate are expressed as: r=σ(W r )·[h t-1 ,x t ] z=σ(W z )·[h t-1 ,x t ] Among them, x t is the traffic information of the current node, h t-1 is the hidden state passed down from the previous node, σ is the activation function, W r is the weight of the control gate, W z is the weight of the update gate, r is the control gate signal, and z is the signal of the update gate.

3. According to claim 2, a traffic flow prediction method based on an improved spatiotemporal graph convolutional neural network is characterized in that: The activation function used in S3 is the tanh function.

4. The traffic flow prediction method based on improved spatiotemporal graph convolutional neural network according to claim 3 is characterized in that: The hidden state after reset is: h′=tanh(W h′ [h t-1 ⊙r,x t ]) Among them, h′ is the hidden state after reset, W h′ is the weight of the hidden state after reset, h t-1 is the hidden state passed down from the previous node, x t It is the traffic information of the current node.

5. The traffic flow prediction method based on improved spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The hidden state of the current node after update in S4 is: h″=h t-1 ⊙z+h′⊙(1-z) Among them, h″ is the hidden state of the current node after update, h′ is the hidden state after reset, z is the signal of the update gate, and h t-1 It is the hidden state passed down from the previous node.

6. The traffic flow prediction method based on improved spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The adjacency matrix A is: Q=h f ′A Q K=h f ′A K Among them, Q and K represent the query and key representation respectively, A Q and A K is the learning parameter of the self-attention mechanism, d is the hidden dimension size of Q and K, and h f ′ is the last hidden state obtained.

7. The traffic flow prediction method based on improved spatiotemporal graph convolutional neural network according to claim 6 is characterized in that: The one-way link form is: Among them, ReLU and tanh are activation functions, A1 and A2 represent the adjacency matrices of randomly initialized node embeddings learned during training, and α is a hyperparameter activation function that controls the node saturation rate.

8. The traffic flow prediction method based on improved spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: Graph convolution is expressed as: Θ(L)x=Θ(UΛU T )x=UΘ(Λ)U T x Among them, Θ is the convolution kernel, U is the graph Fourier basis, Λ is the eigenvalue diagonal matrix of the normalized graph Laplace, L is the graph Laplace matrix, and the graph Laplace matrix is: L=UΛU T Where U is the graph Fourier basis and Λ is the diagonal matrix of eigenvalues ​​of the normalized graph Laplace.

9. The traffic flow prediction method based on improved spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: In S9, the historical traffic data after graph convolution is decomposed into a frequency basis through discrete Fourier transform. The frequency basis has a real part P and an imaginary part Q. The real part P and the imaginary part Q are the inputs of the gates in the gated linear unit respectively.

10. The traffic flow prediction method based on improved spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The loss function is: Where W θ is a trainable parameter in the model; x t+1 is the true value, y represents the predicted value of the model, and x t is the traffic information of the current node, M is the length of the temporal convolution input of each node, and L is the loss function.

Images