A dynamic traffic bottleneck identification method based on a space-time diagram attention network

Abstract

The application discloses a dynamic traffic bottleneck identification method based on a space-time graph attention network and belongs to the field of intelligent traffic, and the method comprises the following steps: acquiring road network data and trajectory data, pre-processing the trajectory data based on the road network data to obtain pre-processed trajectory data; constructing a space-time graph attention network, inputting the road network data and the pre-processed trajectory data into the space-time graph attention network model to obtain node representation with time sequence dynamics; and processing the node representation with time sequence dynamics by using a dynamic traffic bottleneck identification algorithm to identify a dynamic traffic bottleneck. The application can effectively extract the spatial proximity of road network nodes, accurately capture the space-time dynamics of traffic flow propagation, and the identification algorithm has a high running speed and meets the real-time requirement.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation, and particularly relates to a dynamic traffic bottleneck identification method based on a space-time graph attention network. BACKGROUND

[0002] Existing traffic bottleneck identification methods mainly explore the relationship between traffic bottlenecks and traffic demand and road network service level from two aspects. Among them, the method based on the percolation theory is widely studied. This method evaluates the reliability of the traffic network based on the percolation theory, and identifies bottlenecks by analyzing the overall influence of each road section on the flow of travel demand in the road network. This kind of method is suitable for identifying fixed traffic bottlenecks caused by road network planning, relatively stable excessive traffic demand, etc., and has stability and predictability. However, the existing research on dynamic traffic bottlenecks mainly uses probability models and statistical methods to identify traffic bottlenecks, which cannot guarantee the accuracy of bottleneck identification. The reason is that dynamic traffic bottlenecks have concurrency, instantaneity and propagation, and it is necessary to study the mutual relationship between road network nodes under real-time congestion state, and the existing indicators are difficult to quantitatively analyze the dynamic space-time characteristics. Therefore, in order to obtain rich road network features, graph representation learning maps network data (i.e. graph data) into a low-dimensional feature vector, which can retain most of the network information such as the topology of the graph and the relationship between nodes, and is widely used in downstream tasks such as node classification and traffic flow prediction. In order to quantify the traffic correlation between road sections, Road2Vec and Seg2Vec learn the feature representation of road sections using vehicle trajectories, and have outstanding performance in traffic flow prediction and traffic congestion relief, respectively. However, the above models are based on specified time period division for model training, which has certain limitations in obtaining dynamic space-time characteristics of traffic flow.

[0003] Dynamic traffic bottlenecks have concurrency, instantaneity and propagation, and it is necessary to study the mutual relationship between road network nodes under real-time congestion state, and the existing indicators are difficult to quantitatively analyze the dynamic space-time characteristics, which seriously affects the accuracy of bottleneck identification and cannot support the formulation of effective traffic congestion control strategies. Therefore, the application proposes a dynamic traffic bottleneck identification method based on a space-time graph attention network. SUMMARY

[0004] To solve the above technical problems, the application proposes a dynamic traffic bottleneck identification method based on a space-time graph attention network to solve the problems existing in the prior art.

[0005] To achieve the above purpose, the application provides a dynamic traffic bottleneck identification method based on a space-time graph attention network, which comprises:

[0006] obtain road network data and trajectory data, preprocess the trajectory data based on the road network data to obtain preprocessed trajectory data;

[0007] construct a spatio-temporal graph attention network, input the road network data and the preprocessed trajectory data into the spatio-temporal graph attention network model to obtain node representation with time sequence dynamics;

[0008] adopt a dynamic traffic bottleneck identification algorithm to process the node representation with time sequence dynamics to identify dynamic traffic bottlenecks.

[0009] Optionally, the process of preprocessing the trajectory data based on the road network data to obtain preprocessed trajectory data comprises:

[0010] perform geospatial coordinate conversion on the trajectory data to obtain coordinate system converted trajectory data; wherein the coordinate system converted trajectory data and the road network data are data in the same coordinate system;

[0011] perform data cleaning on the coordinate system converted trajectory data to obtain cleaned trajectory data;

[0012] map the cleaned trajectory data to the road network by adopting an ST-Matching algorithm to generate preprocessed trajectory data.

[0013] Optionally, the process of inputting the road network data and the preprocessed trajectory data into the spatio-temporal graph attention network to obtain node representation with time sequence dynamics comprises:

[0014] the graph attention enhancement module of the spatio-temporal graph attention network extracts spatial correlation and traffic flow propagation relationship based on the road network data and the preprocessed trajectory data;

[0015] calculate attention coefficients based on a multi-head graph attention mechanism, and model the spatial correlation and traffic flow propagation relationship based on the attention coefficients to obtain single time slice node representation;

[0016] process a plurality of the single time slice node representation based on a long short-term memory model to obtain node representation with time sequence dynamics.

[0017] Optionally, the graph attention enhancement module comprises a spatial enhancement GAT and a traffic flow enhancement GAT;

[0018] the process of extracting spatial correlation and traffic flow propagation relationship based on the graph attention enhancement module comprises:

[0019] the spatial enhancement GAT processes the road network data based on node spatial coordinates and road segment length to obtain spatial correlation;

[0020] The traffic flow enhancement GAT processes the pretreated trajectory data to obtain a traffic flow propagation relationship.

[0021] Optionally, a calculation expression of the attention coefficient is:

[0022] ;

[0023] wherein, represents importance of node j to node i, A is an attention network, and are feature vectors of node i and node j, is a feature vector of an edge, and and are weight matrices of nodes and edges respectively.

[0024] Optionally, the process of processing a plurality of single-time-slice node representations based on the long short-term memory model to obtain a node representation with timing dynamics includes:

[0025] grouping a plurality of single-time-slice node representations into a time sequence according to a time granularity;

[0026] inputting the time sequence into the long short-term memory model, and outputting a timing-enhanced node representation through an input gate, a forget gate and an output gate;

[0027] obtaining a node representation with timing dynamics based on the timing-enhanced node representation.

[0028] Optionally, the process of processing the node representation with timing dynamics to identify a dynamic traffic bottleneck by using a dynamic traffic bottleneck identification algorithm includes:

[0029] calculating congestion road segments based on roads, road speeds and free flow speeds, and constructing a congestion road segment set based on a plurality of congestion road segments;

[0030] calculating a congestion propagation influence of each road segment based on the congestion road segment set;

[0031] arranging the congestion propagation influences of the road segments in descending order to identify a dynamic traffic bottleneck.

[0032] Optionally, a calculation expression of the congestion propagation influence of each road segment is:

[0033] ;

[0034] wherein, is a road segment, is a local congestion cluster, represents an n-order neighbor set of a certain node, and Sim() is a traffic correlation measurement formula, , , are eigenvectors of nodes u, i, j respectively, is a congestion link set, (·) is a congestion propagation influence of a link, is a link between nodes u, i, is a link between nodes j, u.

[0035] Compared with the prior art, the present application has the following advantages and technical effects:

[0036] The present application proposes a dynamic traffic bottleneck identification method based on a space-time graph attention network. By obtaining road network data and trajectory data, and preprocessing the trajectory data to match the road network data, accurate data basis is provided for subsequent processing. Then, a space-time graph attention network model is constructed, the road network data and the preprocessed trajectory data are input into the model, and the node representation with time sequence dynamics is obtained. The model combines the graph attention mechanism and the long short-term memory network architecture, which can effectively depict the coupling relationship between the spatial topological structure of the road network and the space-time characteristics of the traffic flow, and accurately measure the traffic correlation between the nodes of the road network. Then, the dynamic traffic bottleneck identification algorithm is used to process the node representation with time sequence dynamics, and the dynamic traffic bottleneck is efficiently and accurately identified. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application and their

[0038] Figure 1 is an example of a road network directed graph of an embodiment of the present application;

[0039] Figure 2 is a space-time graph attention network of an embodiment of the present application;

[0040] Figure 3 is a cosine similarity of 3n-order neighbors of an embodiment of the present application, wherein figure (a) is an early morning peak, and figure (b) is a late morning peak;

[0041] Figure 4 is a dynamic cosine similarity of an embodiment of the present application;

[0042] Figure 5 is the running efficiency of an embodiment of the present application, wherein figure (a) is the online processing efficiency, and figure (b) is the model training time;

[0043] Figure 6 is a congestion propagation influence visualization of a local congestion cluster near the west side of the square from 8:55 to 9:00 on October 5, 2016 of an embodiment of the present application;

[0044] Figure 7 Fig. (a) is traffic speed improvement, and Fig. (b) is congestion state duration;

[0045] Figure 8 Fig. (a) is traffic congestion state before control, and Fig. (b) is traffic state after control of road section 1;

[0046] Figure 9 Fig. is a flowchart of a dynamic traffic bottleneck identification method based on a spatio-temporal graph attention network according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0049] The present application aims to build a graph representation learning model that integrates the spatial topology of the road network and the dynamic spatio-temporal characteristics of the traffic flow, and to quantitatively analyze the spatio-temporal propagation characteristics of the traffic bottleneck. In order to achieve this technical goal, the present application proposes a road network representation model based on spatio-temporal graph attention network ST-GAT (Spatio-Temporal Graph Attention Networks), which captures the topological structure correlation between road nodes through spatial enhanced GAT, and models the dynamic influence of traffic flow propagation through traffic flow enhanced GAT, and uses a multi-head attention mechanism to adaptively learn different spatial correlation patterns and highlight key node relationships. This model innovatively combines the graph attention mechanism with the long short-term memory model LSTM (Long Short-Term Memory), and through the dual modeling of spatial topology constraints and dynamic propagation of traffic flow, it breaks through the limitations of traditional methods in representing complex road networks. In addition, a local congestion cluster is defined for the frequently occurring traffic congestion condition, a congestion propagation influence quantification model based on the correlation between nodes is constructed, and a dynamic traffic bottleneck identification algorithm is proposed.

[0050] As a specific embodiment of the present embodiment, the following definitions are made:

[0051] (1) Road network (Road network): Road network is defined as a directed graph where V is the set of nodes, representing intersections in the road network, and E is the set of directed edges, each edge e = (u, v) represents a road segment from node u to node v. is determined by a source node and a target node .

[0052] (2) n-hop neighborhood: Given a directed (or undirected) graph , the n-hop neighborhood of a node is denoted as , which is defined as the set of nodes in graph G with a shortest distance to node u no greater than n. As shown in Figure 1 , the 1-hop neighborhood of node Figure 1 in is , and the 2-hop neighborhood is .

[0053] (3) Node representation: Given a road network graph , node representation aims to learn a feature representation for each node, mapping each node to a d-dimensional feature vector.

[0054] (4) Road speed: Given a road e and a confidence threshold min_sup, the road speed is defined as the average of the average speeds of all vehicles passing through road e in a specified time interval t, as shown in equation (1). (k is the number of vehicles).

[0055] (1)

[0056] (5) Congested road: Given a road e and its road speed v e,t , the free-flow speed v f . When v e,t is less than or equal to C times v f , road e is defined as a congested road. Where C is a default parameter, .

[0057] (6) Local congestion cluster: Given a set of congested roads E t in a certain time interval t, a local congestion cluster c is defined as a graph structure composed of a subset of E t , i.e. , such that each road segment in satisfies at least one of the following conditions:

[0058] 1) is a congested link: all n-order neighbors of nodes i and j belong to ;

[0059] 2) is not a congested link: at least one n-order neighbor of nodes i or j belongs to .

[0060] (7) Congestion propagation influence: Given the set of congested links E t in a time interval t and the local congestion cluster , the congestion propagation influence of node and link e ij in E c is calculated by formula (2):

[0061] (2)

[0062] where Sim() is a traffic correlation measure formula, such as cosine similarity. is a link, is a local congestion cluster, denotes the n-order neighbor set of a node, , , are the feature vectors of nodes u, i, and j, respectively, is the set of congested links, (·) is the congestion propagation influence of a link, is the link between nodes u and i, is the link between nodes j and u. The above definitions take into account the directionality of traffic congestion propagation influence: when e ij is a congested link, the link will only affect the traffic conditions of its upstream links; when e ij is a non-congested link, the traffic flow on the link only affects its downstream links.

[0063] (8) Dynamic traffic bottleneck identification: Given the local congestion cluster , and a positive integer K (K ), the dynamic traffic bottleneck is defined as a subset of E c such that the sum of the congestion propagation influences of the subset is maximum, and =K.

[0064] A dynamic traffic bottleneck identification method based on a space-time graph attention network is provided in this embodiment, which includes the following steps:

[0065] Problem description: Traffic congestion usually causes the speed of vehicles on a particular road to slow down, and this situation lasts for a short time but often occurs. The present invention defines a congestion section by comparing the road speed and the free flow speed. The free flow speed refers to the average speed that a driver will travel under non-congestion or other adverse conditions. The present invention uses the Fth percentile of all valid speeds on a section as its free flow speed, with the default value of F being 85. By monitoring the speed of vehicles, traffic congestion can be quickly and accurately detected.

[0066] Space-time graph attention network (ST-GAT): The space-time graph attention network proposed by the present invention is as shown in Figure 2 The model first extracts the spatial correlation and traffic flow propagation relationship of each node and its neighbor nodes based on the graph attention enhancement module, i.e., the spatial enhancement GAT and the traffic flow enhancement GAT. Based on the multi-head graph attention mechanism, the attention coefficient is calculated, and each set of attention mechanism can learn different node associations and feature weights, thereby highlighting the complex spatial structure of the road network and injecting graph structure information into the model through masked attention, i.e., the calculation of the attention coefficient only considers the n-order neighbors of the node, such as the 2-order neighbors. Then, the output node representation vector is grouped into a time series according to the specified time granularity (such as 5 minutes), and the feature vector of each time interval represents the road network structure characteristics and traffic flow propagation relationship in that time period. Since the traffic flow changes dynamically over time, the present invention extracts the space-time variation characteristics of the traffic flow based on the LSTM neural network, thereby measuring the dynamic traffic correlation between nodes. Finally, the prediction result, i.e., the traffic flow at the next time, is output through the fully connected layer.

[0067] Further, the spatial feature enhancement GAT, which is based on the node spatial coordinates and the road segment length, extracts the spatial correlation between the nodes of the road network. The attention coefficient represents the importance of node j to node i, and the attention coefficient is calculated by equation (3). By performing linear transformation on the features of node i and node j, as well as the features of the edge, followed by feature concatenation and mapping to a real number, i.e., the attention coefficient.

[0068] (3)

[0069] In the formula, represents the importance of node j to node i, A is the attention network, and are the feature vectors of node i and node j, is the feature vector of the edge , and ​ weight matrix of nodes and edges, respectively.

[0070] Finally, the normalized attention coefficient is obtained by using Softmax for normalization, as shown in equation (4):

[0071] (4)

[0072] Further, the traffic flow feature enhances the GAT, and the input node feature is the traffic flow passing through the node in a certain time interval. The edge feature includes the number of vehicles from node i to node j and the number of vehicles from node j to node i in a certain time interval t, that is The attention coefficient between the center node and its neighbor nodes is affected by the traffic flow between the nodes. Since the traffic flow has directionality, the importance of the neighbor node to the center node depends on the traffic flow from the center node to the neighbor node. This means that when the neighbor node has traffic congestion or other abnormalities, the traffic flow from the center node will be affected.

[0073] Further, the traffic flow spatiotemporal feature is extracted, and the application adopts an LSTM network as the core architecture of time series modeling. The LSTM module is composed of a plurality of interconnected memory cells, each of which includes an input gate (Input Gate), a forget gate (Forget Gate) and an output gate (Output Gate) three gate control structures, which effectively solve the gradient disappearance problem in traditional recurrent neural networks through parameterized control of information filtering and state updating process. In specific implementation, the node representation vector generated by the graph attention enhancement module is constructed into an ordered data sequence in the time dimension according to the specified time granularity (such as 5 minutes), which is used as the time series input of the LSTM network.

[0074] This architecture design enables the model to effectively capture the nonlinear variation of traffic flow in consecutive time slots, especially the traffic mutation characteristics of the morning and evening peak periods. Through the multi-level feature fusion mechanism, the final output node representation vector not only integrates the spatial correlation characteristics of the road network topology, but also realizes dynamic modeling of the traffic bottleneck propagation effect through the time series memory function of the gated recurrent unit.

[0075] Dynamic traffic bottleneck identification algorithm: for the vehicle GPS trajectory data continuously collected by the real-time traffic monitoring system, the application proposes a dynamic traffic bottleneck identification method (as shown in algorithm 1). This method screens the key nodes with cascading effect from the detected congestion road segments through a three-stage processing flow:

[0076] (1) Local congestion cluster generation: Based on the road network graph structure, the spatially related congestion cluster within the n-order neighborhood range is constructed for each congestion link (line 1). This operation is implemented by a breadth-first search algorithm to ensure that the links within each cluster have direct or indirect connectivity.

[0077] (2) Dynamic propagation influence calculation: The congestion propagation influence of each link in the local congestion cluster is calculated based on equation (2) in definition 7 (line 6).

[0078] (3) Key congestion link generation: The links in the local congestion cluster are sorted in descending order according to the congestion propagation influence value (line 9), and the top K links are selected as the key congestion links, i.e., traffic bottlenecks (line 10).

[0079] Algorithm 1: Traffic Bottleneck Identification Algorithm Input: congestion road segment set E t , road network node representation vector f, positive integer n, K. Output: Top-K key congestion road segments 1. G c ←GenerateCC(E t , n)2. R← 3. ForEach G c do4. r c ← 5. ForEach c do6. PI(e)=ComputePI(e,c,f)7. Append (e, PI(e)) to r c 8. EndFor9. Sort r c by PI(e) in descending order10. Select the top-K critical roads in r c and add to resultR11. EndFor12. Return R]]>

[0080] Further, as another specific embodiment of the present embodiment, as shown in Figure 9 , an open-source real data set is used for evaluation tests.

[0081] The road network data and trajectory data are obtained, and the object data set is an open-source real data set, including OSM road network data and DiDi-provided taxi trajectory data set, as shown in Table 1.

[0082] Table 1

[0083] Data Name Data Source Data Volume Road Network Data OpenStreetMap 1 ]]> 3151 nodes, 7336 edges Trajectory Data DiDi chuxing 2 ]]> 36.3 million trajectory records, 4167 million location points

[0084] The effectiveness and running efficiency of the proposed method are evaluated through comparative experiments and case analysis. The effectiveness of the proposed ST-GAT model is evaluated by learning the representation vectors of road network nodes, calculating the cosine similarity between neighbor nodes, and comparing with Road2Vec and Seg2Vec. The efficiency of the algorithm is evaluated by the execution time of the offline mining and online processing procedures. In addition, the effectiveness of key congestion road identification is evaluated through case analysis based on the SUMO (Simulation of Urban MObility) traffic simulation software. The running environment of the experiment is a 64-bit server of Ubuntu 20.04.4 (operating system) with Intel Xeon Gold 6226R CPU@2.90GHz x 32 and 256GB RAM, and the GPU is GeForce RTX3090 (24G).

[0085] The trajectory data is preprocessed based on the road network data to obtain preprocessed trajectory data. Since the original trajectory data is unordered and full of noise, the present embodiment constructs a flow-based preprocessing operation to improve data quality. The preprocessing procedure includes the following three key links:

[0086] (1) Geospatial coordinate system conversion: Since the geospatial coordinate system of the collected taxi trajectory data is the Mars coordinate system, which is inconsistent with the WGS84 coordinate system of the OSM road network data, the trajectory data is converted to the position points in the WGS84 coordinate system, thereby supporting subsequent map matching operations, and obtaining the trajectory data after coordinate system conversion.

[0087] (2) Data cleaning: The trajectory after coordinate system conversion is subjected to data cleaning operations such as de-duplication, deletion of small trajectories and abnormal points, etc. to obtain cleaned trajectory data. The abnormal points include trajectory points with speed far exceeding the legal speed limit, trajectory points with long time interval, etc. The processing method for abnormal points is to disconnect the trajectory from the abnormal point and divide it into two trajectories.

[0088] (3) Map matching: Due to errors caused during data collection, there is a certain error distance between the spatial position of the trajectory point and the road network. Therefore, the present application adopts a high-efficiency map matching algorithm, namely the ST-Matching algorithm, to map the trajectory data onto the road network to obtain pre-processed trajectory data and generate trajectory data composed of road network nodes.

[0089] After the data preprocessing stage, the original trajectory data is converted into 3.68 million trajectory records sorted by time, including about 657 million position points.

[0090] The space-time graph attention network is constructed, the road network data and the pre-processed trajectory data are input into the space-time graph attention network model to obtain node representations with time sequence dynamics; the process includes: the graph attention enhancement module of the space-time graph attention network extracts spatial correlation and traffic flow propagation relationship based on the road network data and the pre-processed trajectory data; attention coefficients are calculated based on a multi-head graph attention mechanism, and the spatial correlation and traffic flow propagation relationship are modeled based on the attention coefficients to obtain single-time slice node representations; a long short-term memory model is used to process a plurality of single-time slice node representations to obtain node representations with time sequence dynamics.

[0091] Further, the graph attention enhancement module includes a spatial enhancement GAT and a traffic flow enhancement GAT; the process of extracting spatial correlation and traffic flow propagation relationship based on the graph attention enhancement module includes: the spatial enhancement GAT processes the road network data based on node spatial coordinates and road segment length to obtain spatial correlation; the traffic flow enhancement GAT processes the pre-processed trajectory data to obtain traffic flow propagation relationship.

[0092] Further, the process of processing the single time slice node representation based on the long short-term memory model to obtain a node representation with time sequence dynamics comprises: grouping the single time slice node representations according to a time granularity to form a time sequence; inputting the time sequence into the long short-term memory model to output a time sequence enhanced node representation through an input gate, a forget gate and an output gate; and obtaining the node representation with time sequence dynamics based on the time sequence enhanced node representation.

[0093] The node representation with time sequence dynamics is processed by using a dynamic traffic bottleneck identification algorithm to identify a dynamic traffic bottleneck. The process comprises: calculating congestion road segments based on a road, a road speed and a free flow speed, and constructing a congestion road segment set based on the congestion road segments; calculating a congestion propagation influence of each road segment based on the congestion road segment set; and arranging the congestion propagation influences of the road segments in descending order to identify the dynamic traffic bottleneck.

[0094] Based on the above technical solution, a comparative experiment is performed, and results of the comparative experiment are as follows.

[0095] (1) Spatial adjacency between road network nodes: Based on the representation vector of the road network node, the cosine similarity between the n-order neighbor nodes is calculated to measure the spatial adjacency between the nodes. The early morning peak and the evening peak time intervals are selected, and Road2Vec, Seg2Vec and ST-GAT are trained to obtain the representation vector of the road network node, and the average cosine similarity of the n-order neighbor is calculated. It can be seen from Figure 3 that the cosine similarity of the three models decreases with the increase of the neighbor order, which indicates that the farther the spatial topological distance of the node is, the lower the cosine similarity is. Among them, the decline trend of ST-GAT within 3-order neighbors is more obvious, which indicates that the model can better reflect the change of spatial adjacency. In addition, compared with the early morning peak, the cosine similarity between the nodes in the evening peak changes more obviously. The reason is that the early morning peak is mainly for commuting, the path is relatively fixed (such as residential area→workplace), and the traffic flow has strong regularity.

[0096] While the purpose of travel in the evening peak is more scattered (such as workplace→residential area, commercial area, school, etc.), vehicles frequently switch routes, resulting in more unstable neighbor relationships of the nodes in the road network.

[0097] (2) Dynamic propagation characteristics of traffic flow: The ST-GAT model proposed in the application is based on a graph attention enhancement module, which extracts the spatial correlation between road network nodes and the traffic flow propagation relationship, and the node feature vector output in each time interval represents the road network structure characteristics and traffic flow propagation relationship in the time period. The cosine similarity of Road2Vec, Seg2Vec and ST-GAT in the 8:00-9:00 time period is compared by selecting a local node and its 1-order and 2-order neighbors. As shown in Figure 4As shown in the figure, the cosine similarity of Road2Vec and Seg2Vec does not change over time, because the node feature vectors obtained by training are static and unchangeable. The cosine similarity between nodes obtained by the ST-GAT model proposed in the application fluctuates up and down over time, and the fluctuation rule has a certain correlation with the change of traffic flow. The green line in the figure is the traffic flow value in different time periods. In the first half of the time interval, the traffic flow is in the growth period, and the correlation between the nodes of the road network is not stable. In the later time interval, the traffic flow reaches a certain order of magnitude, and the correlation between the nodes also presents a similar rising and falling trend with the rise and fall of the traffic flow. Therefore, the ST-GAT model proposed can reflect the dynamic correlation of traffic flow between road network nodes. In addition, through model parameter tuning, the following parameters have the best model training effect: learning rate is 0.001, head number is 8, and the number of hidden layers is half of the output number, i.e. 64.

[0098] (3) Running efficiency: In terms of running efficiency, Figure 5 Fig. (a) shows the online processing running time of the traffic bottleneck identification method proposed in the application, including data cleaning, map matching, congestion detection and bottleneck identification. From Figure 5 Fig. (a) shows that when the number of trajectory points increases from 10,000 to 300,000, the running time of the online processing process is basically within 1 minute. The results show that the traffic bottleneck identification method proposed in the application has high efficiency and is suitable for real-time traffic monitoring scenarios.

[0099] In addition, Figure 5 Fig. (b) is the comparative experimental results of model training time. Among them, Seg2Vec-r-30 and Seg2Vec-r-50 respectively represent the training time when the random walk trajectory length r is 30 and 50. When r increases, the training time will increase accordingly. For example, Figure 5 As shown in Fig. (b), the training time of the proposed spatio-temporal graph attention model consists of three stages of S-GAT, T-GAT and LSTM. S-GAT is a spatial feature enhanced GAT, and its running time is the shortest; T-GAT is a traffic flow feature enhanced GAT, and its training time is similar to that of Seg2Vec-r-50; the training time of the LSTM module is shorter than that of Seg2Vec-r-50; and the efficiency of the three stages of the model is better than that of Road2Vec. Therefore, the proposed model has certain scalability in training efficiency.

[0100] Further, as another specific embodiment of the present embodiment, in order to further verify the effectiveness of the proposed model, SUMO traffic simulation software is used to simulate the traffic state after traffic control based on key congestion sections, so as to verify the effect of traffic bottleneck identification on traffic control.

[0101] The local congestion road section in Chengdu City on October 05, 2016, during the morning rush hour (8:55-9:00) is selected, the node representation vector of the next time interval is predicted based on the proposed ST-GAT model, and the congestion propagation influence (PI value) is calculated based on formula (2), and the visualization result of the PI value ranking (such as Table 2) is as shown in Figure 6 The road section is represented by a node, and the larger the PI value of the road section, the larger the node display.

[0102] Based on the SUMO traffic simulator, the traffic control of each road section is selected, and the traffic speed improvement after 20 minutes of control is as shown in Figure 7 (a), the Top-5 control road sections are road section-1, road section-4, road section-2, road section-5 and road section-13, which are basically consistent with the PI value ranking. In addition, the congestion state duration results of the Top-5 control road sections are as shown in Figure 7 (b), the control effect of road section 1 is the best, which shortens the congestion duration by more than 50%, followed by road section-4 and road section-2, which are consistent with the PI value ranking. It is worth noting that when road section-1 and road section-4 are controlled at the same time, the congestion duration does not decrease. The reason is that the more the number of controlled road sections, the more vehicles may be caused to detour, thereby increasing the travel time, so the effect is not as good as the scheme with fewer controlled road sections. Figure 8 The SUMO visualization interface shows the traffic conditions before and after controlling road section 1, wherein the numbers in the figure represent the road section numbers, which correspond to the road sections in Figure 6 The traffic conditions of originally congested road section-9 and road section-10 have improved significantly.

[0103] Table 2

[0104] Ranking Node Number Traffic Congestion Propagation Impact 1 1 0.97 2 4 0.92 3 2 0.85 4 5 0.77 5 13 0.64 6 14 0.57 7 9 0.56 8 3 0.51 9 12 0.49 10 10 0.40

[0105] Real-time management of urban traffic congestion is a key challenge to improve the efficiency of urban operation, and the core is to accurately identify the root cause of congestion-dynamic traffic bottleneck. In view of this urgent need, the spatiotemporal propagation characteristics of traffic congestion are deeply explored, and an innovative dynamic traffic bottleneck identification method is proposed. The core contribution of the present application is to build a spatiotemporal graph attention network (ST-GAT) model, which creatively combines the graph attention mechanism with the long short-term memory (LSTM) network architecture, effectively describes the complex spatial topological structure of the road network and its coupling relationship with the dynamic spatiotemporal characteristics of traffic flow, and accurately measures the traffic correlation between road network nodes. Based on this, the concept of local congestion cluster is further defined, and a model for quantifying congestion propagation influence is constructed, and finally the efficient and accurate identification of dynamic traffic bottleneck is realized.

[0106] The experimental verification fully proves the superiority of the method: the ST-GAT model can not only effectively extract the spatial proximity of the nodes of the road network, but also accurately capture the complex spatio-temporal dynamics implied in the propagation of traffic flow; the developed dynamic bottleneck identification algorithm has high running speed and meets the real-time requirement. More importantly, through case analysis of the typical traffic scene of Chengdu City on the SUMO simulation platform, real-time identification of dynamic traffic bottlenecks is successfully realized. The targeted traffic control strategy based on the identification result significantly shortens the duration of congestion state (more than 50%), which strongly verifies the effectiveness and great potential of the method in the practical application of traffic control.

[0107] In summary, the dynamic traffic bottleneck identification method based on the spatio-temporal graph attention network provided by the present application provides a powerful theoretical model and practical tool for in-depth understanding of the traffic congestion propagation mechanism and accurate positioning of the congestion source. The efficient identification capability and significant application effect provide important scientific basis and technical support for real-time and accurate traffic control decision-making of urban traffic managers, and have important significance for promoting the intelligent and fine development of urban traffic congestion governance.

[0108] In view of the real-time governance requirement of urban road traffic congestion, a quantitative analysis model of traffic congestion propagation characteristics is studied to find the root cause of traffic congestion, i.e. traffic bottleneck. The present application proposes a spatio-temporal graph attention network ST-GAT. The model innovatively combines the graph attention mechanism with the long short-term memory (LSTM) architecture to build a graph representation learning model that integrates the spatial topology of the road network and the spatio-temporal characteristics of traffic flow, thereby measuring the traffic correlation between nodes. A local congestion cluster under the condition of frequent traffic congestion is defined, and a congestion propagation influence quantification model based on the traffic correlation between nodes is constructed to identify dynamic traffic bottlenecks. Experimental results show that the proposed spatio-temporal graph attention network can not only extract the spatial proximity of the nodes of the road network, but also express the spatio-temporal dynamics of traffic flow propagation. At the same time, the dynamic traffic bottleneck identification algorithm has high running speed. In addition, through case analysis of the traffic conditions of Chengdu City based on the SUMO traffic simulation software, real-time identification of dynamic traffic bottlenecks is realized, and the traffic control strategy formulated based on the identified traffic bottlenecks can shorten the duration of congestion state by more than 50%, which demonstrates the effectiveness of the present application for formulating traffic control measures and provides decision support for real-time governance of traffic congestion.

[0109] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A dynamic traffic bottleneck identification method based on a spatio-temporal graph attention network, characterized in that, The method comprises the following steps: obtaining road network data and trajectory data, and preprocessing the trajectory data based on the road network data to obtain preprocessed trajectory data; constructing a spatio-temporal graph attention network, inputting the road network data and the preprocessed trajectory data into the spatio-temporal graph attention network model to obtain node representation with time sequence dynamics, which comprises: a graph attention enhancement module of the spatio-temporal graph attention network extracts spatial correlation and traffic flow propagation relationship based on the road network data and the preprocessed trajectory data; attention coefficients are calculated based on a multi-head graph attention mechanism, and the spatial correlation and the traffic flow propagation relationship are modeled based on the attention coefficients to obtain single-time slice node representation; a long short-term memory model is used to process a plurality of single-time slice node representations to obtain node representation with time sequence dynamics; a dynamic traffic bottleneck identification algorithm is used to process the node representation with time sequence dynamics to identify a dynamic traffic bottleneck.

2. The dynamic traffic bottleneck identification method based on the spatio-temporal graph attention network according to claim 1, characterized in that, The preprocessing of the trajectory data based on the road network data to obtain preprocessed trajectory data comprises: performing geographic spatial coordinate conversion on the trajectory data to obtain coordinate system converted trajectory data; wherein the coordinate system converted trajectory data and the road network data are data in the same coordinate system; performing data cleaning on the coordinate system converted trajectory data to obtain cleaned trajectory data; using an ST-Matching algorithm to map the cleaned trajectory data to the road network to generate preprocessed trajectory data.

3. The dynamic traffic bottleneck identification method based on the spatio-temporal graph attention network according to claim 1, characterized in that, The graph attention enhancement module comprises a spatial enhancement GAT and a traffic flow enhancement GAT. The process of extracting spatial correlation and traffic flow propagation relationship based on the graph attention enhancement module comprises: the spatial enhancement GAT processes the road network data based on node spatial coordinates and road segment length to obtain spatial correlation; the traffic flow enhancement GAT processes the preprocessed trajectory data to obtain traffic flow propagation relationship.

4. The dynamic traffic bottleneck identification method based on the spatio-temporal graph attention network according to claim 1, characterized in that, The calculation expression of the attention coefficient is: ； wherein, represents the importance of node j to node i, A is an attention network, and are feature vectors of node i and node j, is a feature vector of edge , and are weight matrices of nodes and edges, respectively.

5. The dynamic traffic bottleneck identification method based on the spatio-temporal graph attention network according to claim 1, characterized in that, The process of processing a plurality of single-time slice node representations based on the long short-term memory model to obtain node representation with time sequence dynamics comprises: a plurality of single-time slice node representations are grouped into time series according to time granularity; the time series are input into the long short-term memory model, and time sequence enhanced node representation is output through an input gate, a forget gate and an output gate; node representation with time sequence dynamics is obtained based on the time sequence enhanced node representation.

6. The dynamic traffic bottleneck identification method based on the spatio-temporal graph attention network according to claim 1, characterized in that, The process of processing the node representation with time sequence dynamics using a dynamic traffic bottleneck identification algorithm to identify a dynamic traffic bottleneck comprises: congestion road segments are calculated based on roads, road speeds and free flow speeds, and a congestion road segment set is constructed based on a plurality of congestion road segments; congestion propagation influence of each road segment is calculated based on the congestion road segment set; dynamic traffic bottlenecks are identified by arranging congestion propagation influences of various road segments in descending order.

7. The method of claim 6, wherein, The expression for calculating congestion propagation influence of each road segment is: ； wherein, is a road segment, is a local congestion cluster, denotes the n-order neighbor set of a certain node, Sim() is a traffic correlation measure formula, , , are the feature vectors of nodes u, i, j, respectively, is a set of congested road segments, (·) is the congestion propagation influence of a road segment, is a road segment between nodes u, i, is a road segment between nodes j, u.

Images