Region-level traffic prediction method based on depth map convolutional network
By constructing an adaptive adjacency matrix and combining DAGCN and GRU models, the problem of existing traffic flow prediction methods ignoring global correlation is solved, and the accuracy of traffic flow prediction and the robustness of the model are improved.
Patent Information
- Application Number
- CN202510049554.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-13
- Publication Date
- 2025-05-16
AI Technical Summary
When the existing traffic flow prediction method uses graph convolution network (GCN), it mainly focuses on the information between neighboring nodes, ignores global correlation, resulting in limited prediction accuracy.
A regional-level traffic prediction method based on a depth map convolution network is proposed. By constructing an adaptive adjacency matrix A, fusion of spatial and semantic information, and combining DAGCN and GRU models for traffic feature prediction.
Through the dynamically adjusted adaptive adjacency matrix, the model can better understand the deep relationship between nodes, improve prediction accuracy, and balance local and global information through an adaptive adjustment mechanism to generate more distinctive representations.
Smart Images
Figure CN120012982A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of traffic prediction, and more specifically, the present invention relates to a regional level traffic prediction method based on a deep graph convolutional network. Background Art
[0002] With the acceleration of urbanization and the continuous growth of population, urban traffic congestion is becoming increasingly serious, causing great inconvenience to residents' daily travel. Accurately predicting and analyzing traffic flow on urban road networks can effectively assist transportation departments in deploying and guiding traffic flow in advance, thereby improving the efficiency of road network operation and alleviating traffic congestion.
[0003] Traffic flow prediction estimates the traffic status in the future by analyzing historical traffic status data (including traffic flow, speed, congestion level and other related information). Existing traffic flow prediction methods use graph convolutional networks (GCN) to efficiently capture non-Euclidean correlations, and construct an adjacency matrix by using the distance between sensors as the weight of the edge to model spatial correlation.
[0004] Graph Convolutional Network (GCN) learns node feature representation by iteratively aggregating features of adjacent nodes. It mainly focuses on the information between neighboring nodes, thus ignoring global correlation, which in turn limits the prediction accuracy of the model. Summary of the invention
[0005] The present invention provides a regional traffic prediction method based on deep graph convolutional network, aiming to improve the above problems.
[0006] The present invention is implemented as follows: a regional traffic prediction method based on a deep graph convolutional network, the method is specifically as follows:
[0007] (1) Construct an undirected graph G within the prediction area and generate the spatial adjacency matrix A of the prediction area based on the undirected graph G. sp ;
[0008] (2) Read the historical feature sequences of each node in the undirected graph G, calculate the similarity between any two historical feature sequences based on the spatiotemporal association, and then form the semantic adjacency matrix A sp ;
[0009] (3) The spatial adjacency matrix A sp With the semantic adjacency matrix A se Perform fusion to form an adaptive adjacency matrix A;
[0010] (4) The adaptive adjacency matrix A and the historical feature sequences of all nodes are input into the traffic feature prediction model, and the traffic feature prediction model outputs the traffic characteristics of the corresponding intersections in the future within a set time period.
[0011] Furthermore, the spatial adjacency matrix A sp The element in row i and column j of The calculation is based on the following formula:
[0012]
[0013] Among them, d ij represents the actual distance between the intersections corresponding to the i-th node and the j-th node in the road network, and σ and α are the control space adjacency matrix A sp Parameters for sparsity settings.
[0014] Furthermore, the historical feature sequence X i With the historical feature sequence X j The similarity calculation formula is as follows:
[0015]
[0016] Among them, W mn Represents the historical feature sequence X i The mth historical feature With the historical feature sequence X j The nth historical feature of The weight between them, r(m,n) is the element in the mth row and nth column of the distance matrix R, which represents the historical feature sequence X i The mth historical feature and historical feature sequence X j The distance between the nth historical features, p is the optimal distance set.
[0017] Furthermore, the weight W mn The calculation formula is as follows:
[0018]
[0019] Among them, M c is the midpoint of the historical feature sequence, and g is a random number between 0 and 1.
[0020] Furthermore, the semantic adjacency matrix A se The element in row i and column j of Represents the historical feature sequence X of the i-th node i and the historical feature sequence X of the jth node j Are the elements similar? The calculation formula is as follows:
[0021]
[0022] Among them, β is the control space adjacency matrix A seParameters for sparsity settings.
[0023] Furthermore, the calculation formula of the adaptive adjacency matrix A is as follows:
[0024]
[0025] Among them, E A ∈R N×D is a random matrix, the matrix element value is between 0 and 1, N is the number of nodes, D is the traffic feature dimension of the node, Softmax is a normalized exponential function, E A T Represents the matrix E A The transpose of .
[0026] Furthermore, the traffic feature prediction model includes a DAGCN model and a GRU model connected in sequence, the DAGCN model is formed by connecting multiple DAGCN units in sequence, and the GRU model is formed by connecting multiple GRU units in sequence.
[0027] Furthermore, the feature extraction process of the DAGCN unit is as follows:
[0028] Z=MLP(X in );
[0029]
[0030] H=Stack(Z,h1,...,h k );
[0031]
[0032] Among them, MLP is a multi-layer perceptron, and the parameter k represents the hyperparameter of the propagation depth. where s k =h k s, s∈R D×1 is a trainable vector, σ(·) is the sigmoid activation function; Stack, Reshape and Squeeze represent stacking, reshaping and squeezing operations respectively; Softmax is the normalization function, and the input feature X of the first DAGCN unit in The historical feature sequences of all nodes form a historical feature matrix X, and the input feature X of the subsequent DAGCN unit in is the output X of the previous DAGCN unit out .
[0033] Furthermore, the undirected graph G = (V, E), where V represents a set of nodes, the nodes are intersections in the road network in the prediction area, and E represents an edge set, recording the connectivity between nodes.
[0034] The regional traffic prediction method based on deep graph convolutional network provided by the present invention has the following beneficial technical effects:
[0035] (1) Traffic characteristics of nodes are predicted based on the adaptive adjacency matrix A, which integrates the dynamic attributes of historical traffic data and the semantic spatial relationship between nodes. This enables the model to better understand the deep relationship between nodes. For example, nodes that are geographically far away may also have similar traffic patterns. This dynamically adjusted adjacency matrix can reflect the spatiotemporal changes in the actual traffic network, thereby improving the accuracy of the model;
[0036] (2) A deep graph convolutional network model (DAGCN model) is designed. After the DAGCN model decouples transformation and propagation, a deeper graph convolutional network can be used to learn the representation of graph nodes over a larger receptive field. Secondly, an adaptive adjustment mechanism is used to adaptively balance the local and global information of each node, which helps to generate more discriminative representations. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 A flowchart of a regional traffic prediction method based on a deep graph convolutional network provided by an embodiment of the present invention;
[0038] Figure 2 A schematic diagram of a feature extraction process of a DAGCN unit provided in an embodiment of the present invention;
[0039] Figure 3 A comparison chart of experimental results of different configurations on the SZ_TAXI dataset provided by an embodiment of the present invention;
[0040] Figure 4 A comparison chart of experimental results of different configurations on the NYC_TAXI dataset provided by an embodiment of the present invention;
[0041] Figure 5 Perturbation analysis on the SZ_TAXI data set provided by an embodiment of the present invention;
[0042] Figure 6 Perturbation analysis on the NYC_TAXI dataset provided by an embodiment of the present invention;
[0043] Figure 7 The prediction level of 15min at node 17 in SZ_TAXI provided by the embodiment of the present invention;
[0044] Figure 8 The prediction level of node 99 in NYC_TAXI for 1h is provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0045] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.
[0046] Figure 1 A flowchart of a regional traffic prediction method based on a deep graph convolutional network provided by an embodiment of the present invention, the method is specifically as follows:
[0047] (1) Construct an undirected graph G within the prediction area and generate the spatial adjacency matrix A of the prediction area based on the undirected graph G. sp ;
[0048] In an embodiment of the present invention, an undirected graph G is constructed for the prediction area where traffic flow prediction is required, where G = (V, E), wherein V represents a set of nodes, |V| = N, the nodes are intersections in the road network in the prediction area, and E represents an edge set for recording the connectivity between nodes.
[0049] Spatial adjacency matrix A sp The element in row i and column j of Represents the connectivity between the i-th node and the j-th node, the element The calculation is based on the following formula:
[0050]
[0051] Among them, d ij represents the actual distance between the intersections corresponding to the i-th node and the j-th node in the road network, and σ and α are the control space adjacency matrix A sp Parameters for sparsity settings.
[0052] (2) Read the historical feature sequences of each node in the undirected graph G, calculate the similarity between any two historical feature sequences based on the spatiotemporal association, and then form the semantic adjacency matrix A sp ;
[0053] In the embodiment of the present invention, the computational time complexity of generating a similarity matrix using the DTW algorithm based on the similarity of time series is relatively high, and is difficult to apply in practice; moreover, the DTW algorithm calculates the distances of all points between two time series to achieve equal weights for each point in the sequence. The present invention believes that adjacent points between two time series are more important than other points, and the relative importance of the phase difference between the points should be taken into account. Therefore, the present invention proposes a fast weighted dynamic warping algorithm (FWDTW, Fast Weighted Dynamic Time Warping) to calculate the similarity of two time series, which is convenient for capturing traffic characteristics that change over time, can better understand the dynamic behavior of traffic information, and reveal important patterns hidden in the time series, thereby accurately capturing spatiotemporal semantic associations.
[0054] The historical feature sequence of the i-th node in the undirected graph G And the historical feature sequence of the jth node in, They represent the traffic characteristics collected by the i-th node and the j-th node at the current step length t, They respectively represent the traffic characteristics collected at the i-th node and the j-th node in the historical step tn. The traffic characteristics include: traffic flow and average speed.
[0055] In the embodiment of the present invention, the historical feature sequence X i With the historical feature sequence X j The similarity calculation formula is as follows:
[0056]
[0057] Among them, W mn Represents the historical feature sequence X i The mth historical feature With the historical feature sequence X j The nth historical feature of The weight between them, r(m,n) is the element in the mth row and nth column of the distance matrix R, which represents the historical feature sequence X i The mth historical feature and historical feature sequence X j The distance between the nth historical features reflects the similarity of time series more effectively than the Euclidean distance, and p is the optimal distance set.
[0058] In order to assign weights according to the phase difference between two historical features, the weight W mn Defined as:
[0059]
[0060] Among them, M cis the midpoint of the historical feature sequence. If there are 11 historical features in the historical feature sequence, then M c =5, g is a random number between 0 and 1.
[0061] Semantic adjacency matrix A se The element in row i and column j of Represents the historical feature sequence X of the i-th node i and the historical feature sequence X of the jth node j Are the elements similar? The calculation is based on the following formula:
[0062]
[0063] Among them, β is the control space adjacency matrix A se Parameters for sparsity settings.
[0064] (3) The spatial adjacency matrix A sp With the semantic adjacency matrix A se Perform fusion to form an adaptive adjacency matrix A;
[0065] In the embodiment of the present invention, the adaptive adjacency matrix A is a spatial adjacency matrix A sp and the semantic adjacency matrix A se The adaptive adjacency matrix A is specifically defined as:
[0066]
[0067] Among them, E A ∈R N×D is a random matrix, the matrix element value is between 0 and 1, N is the number of nodes, and D is the traffic feature dimension of the node. Softmax is a normalized exponential function, and ReLU is selected as the nonlinear activation function of the network because compared with sigmoid and tanh activation functions, ReLU function can effectively prevent gradient explosion, E A T Represents the matrix E A The transpose of A∈R N×N is the adaptive adjacency matrix.
[0068] (4) The adaptive adjacency matrix A and the historical feature sequences of all nodes are input into the traffic feature prediction model, and the traffic feature prediction model outputs the traffic characteristics of the corresponding intersections in the future within a set time period.
[0069] In traffic status prediction, identifying complex spatial correlations between regions is a core challenge. Urban road networks present irregular non-Euclidean spatial characteristics, which makes traditional CNNs incapable of dealing with such problems. In contrast, GCN shows significant advantages in this regard and provides an effective solution for feature extraction in irregular areas. In recent years, the use of GCN to capture spatial features from irregular road networks has achieved remarkable results. Among them, the convolution operation of the graph is a key step in achieving neighborhood aggregation and plays a decisive role in extracting spatial features. However, traditional neighborhood aggregation methods generally only consider one layer of direct neighbors. When expanding the receptive domain, it will cause the representation transformation and propagation to become entangled with the receptive domain parameters in the propagation, which ultimately leads to performance degradation. This problem is called the over-smoothing problem. The DAGCN model proposed in the present invention represents the graph convolution operation as a decoupling of transformation and propagation, which improves the performance degradation problem caused by the expansion of the receptive domain. The feature extraction process of the DAGCN unit is as follows: Figure 2 As shown in Figure 1, the DAGCN model consists of three core steps: conversion, propagation, and adaptive adjustment. Through the adaptive adjustment mechanism, the DAGCN model can achieve the balance and effective integration of local and global information of each node.
[0070] In an embodiment of the present invention, the traffic feature prediction model includes a DAGCN model and a GRU model connected in sequence. The DAGCN model is formed by connecting multiple DAGCN units in sequence, and the GRU model is formed by connecting multiple GRU units in sequence. The feature extraction process of the DAGCN unit is as follows:
[0071] Z=MLP(X in );
[0072]
[0073] H=Stack(Z,h1,...,h k );
[0074]
[0075] Among them, MLP is a multi-layer perceptron, and the parameter k represents the hyperparameter of the propagation depth. where s k =h k s, s∈R D×1 is a trainable vector, σ(·) is the sigmoid activation function; Stack, Reshape and Squeeze represent stacking, reshaping and squeezing operations respectively, in order to maintain the consistency of size during the calculation process; Softmax is the normalization function, the input feature X of the first DAGCN unit inThe historical feature sequences of all nodes form a historical feature matrix X, and the input feature X of the subsequent DAGCN unit in is the output X of the previous DAGCN unit out .
[0076] The feature X in Input the multi-layer perceptron MLP for feature conversion to obtain the feature matrix Z∈R B×N×D , B is the set Batch batch size. In the process of propagation and conversion, a symmetric neighborhood aggregation propagation mechanism is adopted: in, I is the identity matrix, For the matrix The degree matrix of For the matrix The element in the i-th row and j-th column, h l ∈R B×N×D Represents the information collected by the node in the subtree with height l as the root. By stacking different receptive domains and the node itself, H∈R B×N×(k+1)×D , with the increase of depth, h l = contains more global information, a smaller l may not capture sufficient and necessary local information, while a larger l may introduce more global information, thereby diluting relatively important local information. In order to find the best receptive field for each node, a trainable projection vector s is used to generate a retention score to represent information from different regions. The retention score s is used to measure how much information from different propagation layers each node should retain. After using reshape and squeeze operations to maintain the consistency of the calculation scale, X is obtained through the Softmax function. out ∈R B×N×D , which enables the DAGCN model to adaptively balance the local and global information of each node.
[0077] Since GRU has fewer parameters and improves the convergence speed while maintaining prediction accuracy, the GRU model can be used to extract the dynamic characteristics of traffic status from time series data in traffic status prediction.
[0078] The regional traffic prediction method based on deep graph convolutional network provided by the present invention has the following beneficial technical effects:
[0079] (1) Traffic characteristics of nodes are predicted based on the adaptive adjacency matrix A, which integrates the dynamic attributes of historical traffic data and the semantic spatial relationship between nodes. This enables the model to better understand the deep relationship between nodes. For example, nodes that are geographically far away may also have similar traffic patterns. This dynamically adjusted adjacency matrix can reflect the spatiotemporal changes in the actual traffic network, thereby improving the accuracy of the model;
[0080] (2) A deep graph convolutional network model (DAGCN model) is designed. After the DAGCN model decouples transformation and propagation, a deeper graph convolutional network can be used to learn the representation of graph nodes over a larger receptive field. Secondly, an adaptive adjustment mechanism is used to adaptively balance the local and global information of each node, which helps to generate more discriminative representations.
[0081] The present invention conducts experiments on the Shenzhen taxi dataset (SZ_TAXI) and the New York taxi dataset (NYC_TAXI) to evaluate the prediction performance of the regional-level traffic prediction method based on deep graph convolutional network (ST-DAGCN model for short) proposed in the present invention.
[0082] The SZ_TAXI dataset is derived from taxi GPS data collected in Shenzhen, China in January 2019. During preprocessing, weekend and holiday data were first removed, leaving only data from 22 working days, covering 78 areas in Shenzhen city divided by administrative postal codes. The NYC_TAXI dataset is provided by the New York City Taxi and Limousine Commission, covering data from January 1, 2020 to March 30, 2020, involving a total of 263 areas. The relevant information of the two datasets is shown in Table 1.
[0083] Table 1 Dataset information
[0084]
[0085] The present invention uses the following benchmark models to evaluate the performance of the ST-DAGCN model. HA uses the average value of historical traffic flow information as the prediction result; ARIMA is a parameter model widely used in various traffic prediction studies; GRU is a deep learning model based on RNN, which is widely used in traffic prediction; ASTGCN uses the spatiotemporal attention mechanism to model spatiotemporal dependencies respectively; T-GCN captures temporal and spatial dependencies to predict short-term traffic flow; Tms-GCN captures temporal and multi-spatial dependencies to predict regional-level traffic flow.
[0086] The performance of the ST-DAGCN model is compared with the baseline model using the following three commonly used metrics:
[0087] (a) Mean absolute error (MAE):
[0088] Among them, Y i =(y1,...,y N ) is the target value; is the predicted value; i It is expressed as the target value of the ith time interval (ith step); is the predicted value of the i-th time interval; P is the number of predicted time intervals.
[0089] (b) Mean absolute percentage error (MAPE):
[0090] (c) Root mean square error (RMSE):
[0091] The experiments of the present invention are compared from four aspects: prediction performance comparison, ablation experiment, model robustness experiment and model interpretability.
[0092] In order to verify the effectiveness of the ST-DAGCN model in traffic prediction tasks, the prediction accuracy of the present invention is compared with that of the classic model. Tables 2 and 3 respectively give the results of the tests on the SZ_TAXI and NYC_TAXI datasets. By analyzing the experimental results, the following conclusions can be drawn: (1) Compared with mathematical statistical models such as HA and ARIMA, the deep learning model exhibits higher prediction accuracy. (2) The ASTGCN and T-GCN models are superior to the GRU model that only considers time correlation by constructing spatiotemporal correlations. The models that consider multiple spatial correlations and TmS-GCN are superior to the AST-GCN model and the T-GCN model. (3) The ST-DAGCN model that constructs an adaptive matrix is superior to spatiotemporal models such as T-GCN and TmS-GCN, that is, the use of autonomously generated adaptive adjacency matrices and DAGCN to adaptively capture global and local spatiotemporal correlations can improve the prediction performance.
[0093] Table 2 Comparison of prediction performance on the SZ_TAXI dataset
[0094]
[0095] Note: The name of the paper corresponding to the HA model is: Passenger flow prediction of subway transfer stations based on nonparametric regression model; the name of the paper corresponding to the ARIMA model is: A Comparison of ARIMA and LSTM in Forecasting Time Series; the name of the paper corresponding to the GRU model is: Using LSTM and GRU neural networks methods for traffic flow prediction; the name of the paper corresponding to the AST-GCN model is: Attention based spatial-temporal graph convolutional networks for traffic flow forecasting; the name of the paper corresponding to the T-GCN model is: T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction; the name of the paper corresponding to the Tms-GCN model is: Region-Level Traffic Prediction Based on Temporal Multi-Spatial Dependence Graph Convolutional Network from GPS Data.
[0096] Table 3 Comparison of prediction performance on the NYC_TAXI dataset
[0097]
[0098] Through ablation experiments, this paper verifies the impact of hidden spatiotemporal information on traffic prediction. We develop two variants based on the ST-DAGCN model. Figure 3 and Figure 4 As shown in Figure 2, by comparing ST-DAGCN and two variants based on this model on the SZ_TAXI and NYC_TAXI datasets, we can analyze their differences and evaluate their performance. The following are the characteristics of these four variants.
[0099] (1) EA: Only use learnable parameters adjacency E A The generated dynamic matrix is used as the parameter matrix.
[0100] (2) SP: A dynamic matrix generated by a predefined adjacency matrix is used as a parameter matrix.
[0101] (3)AM: A dynamic matrix generated by the semantic adjacency matrix is used as the parameter matrix.
[0102] (4) ST-DAGCNa: Eliminate the DAGCN model and use the GCN model as a replacement.
[0103] according to Figure 3 and Figure 4 , ST-DAGCN achieved the lowest prediction results of MAE, RMSE and MAPE on both datasets. The dynamic adjacency matrix generated by directly learnable parameters in EA cannot fully reflect the real spatial relationship, so the prediction performance is poor compared with the dynamic adjacency matrix generated by geographic spatial relationship in SP. The dynamic adjacency matrix generated by the similarity of time series in AM mines deeper spatial patterns and takes into account the dynamic properties of nodes, making the prediction results better. In ST-DAGCNa, only GCN is used to capture the spatial information of the neighborhood, and the expansion of the receptive field leads to performance degradation, and fails to effectively capture local and global spatial features. These results show that the autonomously generated adaptive adjacency matrix can capture the spatial features in the traffic area more effectively than the predefined adjacency matrix, and the prediction accuracy of ST-DAGCN is higher than that of ST-DAGCNd, which shows that DAGCN can effectively capture local and global spatial features and improve the prediction performance of the model.
[0104] Gaussian noise is added to the data to verify the robustness of the model. The noise follows a Gaussian distribution N∈(0,σ 2 ), (σ∈0.2, 0.4, 0.6, 0.8, 1.0), the experimental results are as follows Figure 5 and Figure 6 As shown, it can be seen that the change in the value of the evaluation index as the noise increases can be ignored, which verifies the robustness of the model.
[0105] To illustrate the prediction ability of the model, the experiment visually compares the true value of the test set with the prediction results of the ST-DAGCN model. To further study the effectiveness of the adaptive adjacency matrix independently constructed based on node attributes and adjacency matrix, we visualize the results of the ablation experiment.
[0106] The self-generated adaptive adjacency matrix improves the model's perception of peaks and inflection points. Figure 7 and Figure 8 It can be seen that at the inflection point and peak, the model prediction results of the self-generated dynamic graph are closer to the true value than the model prediction results of the static graph.
[0107] The present invention has been described exemplarily. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.
Claims
1. A regional traffic prediction method based on deep graph convolutional network, characterized in that: The method is specifically as follows: (1) Construct an undirected graph G within the prediction area and generate the spatial adjacency matrix A of the prediction area based on the undirected graph G. sp ; (2) Read the historical feature sequences of each node in the undirected graph G, calculate the similarity between any two historical feature sequences based on the spatiotemporal association, and then form the semantic adjacency matrix A sp ; (3) The spatial adjacency matrix A sp With the semantic adjacency matrix A se Perform fusion to form an adaptive adjacency matrix A; (4) The adaptive adjacency matrix A and the historical feature sequences of all nodes are input into the traffic feature prediction model, and the traffic feature prediction model outputs the traffic characteristics of the corresponding intersections in the future within a set time period.
2. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 1, characterized in that: Spatial adjacency matrix A sp The element in row i and column j of The calculation is based on the following formula: Among them, d ij represents the actual distance between the intersections corresponding to the i-th node and the j-th node in the road network, σ and α are the control space adjacency matrix A sp Parameters for sparsity settings.
3. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 1, characterized in that: Historical feature sequence X i With the historical feature sequence X j The similarity calculation formula is as follows: Among them, W mn Represents the historical feature sequence X i The mth historical feature With the historical feature sequence X j The nth historical feature of The weight between them, r(m,n) is the element in the mth row and nth column of the distance matrix R, which represents the historical feature sequence X i The mth historical feature and historical feature sequence X j The distance between the nth historical features, p is the optimal distance set.
4. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 3, characterized in that: Weight W mn The calculation formula is as follows: Among them, M c is the midpoint of the historical feature sequence, and g is a random number between 0 and 1.
5. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 3, characterized in that: Semantic adjacency matrix A se The element in row i and column j of Represents the historical feature sequence X of the i-th node i and the historical feature sequence X of the jth node j Are the elements similar? The calculation formula is as follows: Among them, β is the control space adjacency matrix A se Parameters for sparsity settings.
6. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 1, characterized in that: The calculation formula of the adaptive adjacency matrix A is as follows: Among them, E A ∈R N×D is a random matrix, the matrix element value is between 0 and 1, N is the number of nodes, D is the traffic feature dimension of the node, Softmax is a normalized exponential function, E A T Represents the matrix E A The transpose of .
7. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 1, characterized in that: The traffic feature prediction model includes a DAGCN model and a GRU model connected in sequence. The DAGCN model is composed of multiple DAGCN units connected in sequence, and the GRU model is composed of multiple GRU units connected in sequence.
8. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 7, characterized in that: The feature extraction process of the DAGCN unit is as follows: Z=MLP(X in ); H=Stack(Z,h1,...,h k ); Among them, MLP is a multi-layer perceptron, and the parameter k represents the hyperparameter of the propagation depth. where s k =h k s, s∈R D×1 is a trainable vector, σ(·) is the sigmoid activation function; Stack, Reshape and Squeeze represent stacking, reshaping and squeezing operations respectively, Softmax is the normalization function, and the input feature X of the first DAGCN unit in The historical feature sequences of all nodes form a historical feature matrix X, and the input feature X of the subsequent DAGCN unit in is the output X of the previous DAGCN unit out .
9. The regional traffic prediction method based on deep graph convolutional network as claimed in claim 1, characterized in that: Undirected graph G = (V, E), where V represents a set of nodes, nodes are intersections in the road network in the prediction area, and E represents an edge set, recording the connectivity between nodes.