Network-level traffic operation state inference method for unobservable road section

Abstract

The invention relates to the technical field of intelligent traffic control, in particular to an unobservable road section-oriented network-level traffic operation state inference method, which comprises the following steps of: by taking checkpoint detection data as a data basis, completing data preprocessing based on checkpoint data characteristics, and extracting traffic flow characteristic parameters; constructing a model algorithm based on a Bayesian network and a traffic flow fundamental diagram, and realizing traffic flow characteristic parameter completion for unobservable road sections; spatial characteristics of a road network are introduced on the basis of constructing a double attention mechanism-directed graph long short-term memory network model. Therefore, the traffic management department can extract the traffic flow characteristic parameters by means of the checkpoint and complement the traffic flow characteristic parameters of the unobservable road section by means of the model algorithm so as to construct a complete road network traffic state inference model. Therefore, the problem that the real-time traffic state cannot be mastered due to the lack of corresponding detection equipment in part of road sections in a real urban road network is solved, and the traffic operation state of the whole road network can be deduced and mastered.

Description

Technical Field

[0001] The present invention relates to the field of intelligent traffic control technology, and in particular to a network-level traffic operation state inference method for unobservable road sections. Background Art

[0002] With the advancement of science and technology, the development of intelligent transportation systems has provided an important means for traffic management departments to grasp the operation status and changing trends of urban road traffic. Compared with infrared, induction coils, GPS and other detection equipment, the bayonet detection equipment based on visual perception technology has the advantages of wide deployment range, high data validity and good stability, and has become the most commonly used data source for traffic management departments. Although the bayonet detection equipment has the widest application range, there are still a considerable number of roads that have not yet been installed with traffic detection equipment, that is, for traffic managers, there are a large number of sections in urban roads where their traffic operation status cannot be effectively grasped. In response to this problem, researchers have sought solutions from the perspectives of traffic flow theory modeling and artificial intelligence modeling, but the method based on traffic flow theory modeling has poor universality, harsh applicable conditions, and low feasibility; the method based on artificial intelligence modeling requires a large amount of reliable data, and some sections are completely unobservable under realistic conditions, that is, the data is completely missing, and the task of traffic state estimation cannot be effectively completed.

[0003] The rapid development of deep learning has provided a new opportunity for inferring traffic operation status. It has a high degree of self-organizing and adaptive ability, information storage capacity and strong computing power. It can adapt to the diversity and complexity of input data and express arbitrary linear and nonlinear relationships. It can obtain a pattern from the training data and infer the results of the test data based on this pattern. However, the current inventions give little consideration to the ability of the road network itself to attract traffic flow, and traffic managers are more concerned about the changing trends of traffic operation status during peak hours in the morning and evening. Existing deep learning algorithms lack special considerations for peak hours in the morning and evening. Especially in the real data environment where some sections of the road are unobservable, there is a lack of effective means to infer network-level traffic operation status.

[0004] More importantly, the traffic network is a complex dynamic system, and its flow state is affected by both time series and spatial position changes. There are direct or indirect connections and influences between traffic flows in different regions, which makes it difficult to quickly and accurately integrate traffic data of different time and space dimensions, which is not conducive to the extraction of key information, and further increases the difficulty of inferring network-level traffic operation status. Therefore, technical personnel are urgently needed to solve the above problems. Summary of the invention

[0005] Therefore, in view of the above-mentioned existing problems and defects, the designers of the present invention collected relevant information, conducted multiple evaluations and considerations, and continued to experiment and modify the method after many years of R&D experience in this industry, which ultimately led to the emergence of a network-level traffic operation status inference method for unobservable sections.

[0006] In order to solve the above technical problems, the present invention relates to a network-level traffic operation state inference method for unobservable road sections, which comprises the following steps:

[0007] S1. A data preprocessing method based on the police checkpoint detection data is used to correct or delete abnormal data, and on this basis, characteristic parameters of traffic flow, such as flow rate, speed and density, are extracted;

[0008] S2. Construct a traffic flow characteristic parameter completion model based on Bayesian network and traffic flow basic graph to complete the traffic flow characteristic parameters of unobservable sections in the road network;

[0009] S3. Build a directed graph long short-term memory network model based on dual attention mechanism, introduce the attention mechanism of GAT model, and capture the factors affecting the traffic flow of upstream and downstream sections;

[0010] S4. Based on the space syntax theory, the topological value of each road section in the road network and the vehicle speed characteristics of each road section in the road network are calculated, wherein the topological value of each road section in the road network is used as a spatial dimension feature; the topological value of each road section in the road network is combined with the vehicle speed characteristics to construct a directed graph convolution operator;

[0011] S5. Reversely input the spatial dimension features and directed graph convolution operator obtained in step S4 into the dual attention mechanism-directed graph long short-term memory network model to evaluate the effectiveness of the model.

[0012] As a further improvement of the technical solution disclosed in the present invention, step S1 includes the following sub-steps:

[0013] S11, matching and comparing the traffic detection data of two adjacent public security checkpoints, and eliminating invalid license plate data that cannot be correctly read and unreasonable license plate data that appear within a fixed observation time window;

[0014] S12, dividing the signal cycle to which the current vehicle belongs according to the timestamp of the public security checkpoint detection data, deducting 4 seconds of starting loss from the time of the first vehicle passing the stop line in each cycle, and obtaining the green light start time of each cycle;

[0015] S13, comparing the green light on time with the actual green light on time to obtain an offset, performing normal distribution fitting on the distribution of the offset in the entire study area, selecting the offset with the highest normal distribution probability as the calibration value of the drift correction algorithm, and realizing correction of the offset error of the green light on time;

[0016] S14, selecting a fixed observation time window, counting the traffic volume, average travel speed and average vehicle density of observable sections in the road network, and eliminating abnormal data by calculating the statistical distribution characteristics of the average travel speed of vehicles.

[0017] As a further improvement of the technical solution disclosed in the present invention, step S2 includes the following sub-steps:

[0018] S21. Based on the Greenshields model, the functional relationship of the basic graph of traffic flow is fitted by the least squares method;

[0019] S22. Construct a Bayesian network graph structure for unobservable sections in a road network based on basic traffic flow theory;

[0020] S23. Determine the relationship between nodes in the Bayesian network model of S22 through the basic graph function relationship of the traffic flow constructed in S21, and complete the traffic flow characteristic parameters of the unobservable sections in the road network.

[0021] As a further improvement of the technical solution disclosed in the present invention, step S3 includes the following sub-steps:

[0022] S31, introduce the graph attention module to explore the influence of traffic flow parameters between adjacent road sections;

[0023] S32. Long short-term memory network is integrated with attention mechanism to optimize the weights of the function.

[0024] As a further improvement of the technical solution disclosed in the present invention, step S4 includes the following sub-steps:

[0025] S41. Based on the space syntax theory, the topological value of each road segment in the road network is calculated and used as the spatial dimension feature;

[0026] S42. By obtaining the average travel speed of the vehicle in each road section, a speed variance matrix of each road section is constructed as the vehicle running speed feature, and together with the spatial dimension feature constructed in S41, it constitutes the feature vector of the model input, and constructs a directed graph convolution operator.

[0027] As a further improvement of the technical solution disclosed in the present invention, step S5 includes the following sub-steps:

[0028] S51, selecting the mean absolute error, root mean square error and mean absolute error percentage in the neural network algorithm as the model effectiveness evaluation indicators;

[0029] S52. Input actual data and evaluate the effectiveness of the dual attention mechanism introduced into the model through model ablation analysis.

[0030] In practical applications, the network-level traffic operation state inference method for unobservable road sections disclosed in the present invention can achieve at least the following beneficial technical effects, specifically:

[0031] 1) Aiming at solving practical problems, the most widely used and easily accessible public security checkpoint detection data is used as the data source, and the Bayesian model and traffic flow basic graph model are used to complete the traffic flow characteristic parameters of unobservable sections, and a dual attention mechanism-directed graph long short-term memory network model is constructed to capture the traffic flow influencing factors of upstream and downstream sections while improving the balance of training data, so as to achieve accurate inference of the actual operation status of road network traffic;

[0032] 2) According to actual data verification, the dual attention mechanism-directed graph long short-term memory network model meets expectations in short-term traffic flow prediction and achieves excellent results. It provides theoretical support for predicting the traffic flow conditions of road sections and road networks in the short term in the future, and provides a basis for traffic management and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 This is an overall flow chart of the network-level traffic operation status inference method for unobservable road sections disclosed in the present invention.

[0035] Figure 2 The present invention discloses a layout diagram of a road network.

[0036] Figure 3 The schematic diagram of the architecture of the Bayesian network constructed in the network-level traffic operation status inference method for unobservable road sections disclosed in the present invention.

[0037] Figure 4 This is a schematic diagram of the architecture of the dual attention mechanism-directed graph long short-term memory network model constructed in the network-level traffic operation status inference method for unobservable road sections disclosed in the present invention.

[0038] Figure 5 This is a schematic diagram of the architecture of the attention mechanism introduced into the long short-term memory network of the network-level traffic operation status inference method for unobservable road sections disclosed in the present invention.

[0039] Figure 6 A schematic diagram of the layout of an example road network for verifying the effectiveness of the network-level traffic operation status inference method for unobservable road sections disclosed in the present invention.

[0040] Figure 7 This is a schematic diagram of the results after the traffic flow characteristic parameters of the unobservable road section are completed in the network-level traffic operation status inference method for the unobservable road section disclosed in the present invention.

[0041] Figure 8 This is a comparison chart of the MAE and RMSE results of the dual attention mechanism-directed graph long short-term memory network model compared with the baseline model in the network-level traffic operation status inference method for unobservable road sections disclosed in the present invention.

[0042] Fig. 9 The present invention discloses an ablation analysis result of a network-level traffic operation status inference method for unobservable road sections based on a dual attention mechanism-directed graph long short-term memory network model. DETAILED DESCRIPTION

[0043] In order to facilitate those skilled in the art to fully understand the technical solutions disclosed in the present invention, Figure 2 A schematic diagram of the layout of a road network disclosed in the present invention is shown. It can be clearly known that some sections lack checkpoint detection equipment or are affected by other difficult-to-overcome factors, which can easily lead to the inability to fully obtain traffic flow operation parameters, and thus the inability to infer the traffic operation status of the entire road network after a period of time. In view of this, the present invention proposes a network-level traffic operation status inference method for unobservable sections. The present invention is applied to the public security checkpoint detection data, which is the most extensive data source of the traffic management department, as the data source, and the data preprocessing process is completed based on the statistical distribution characteristics of the checkpoint detection data, and duplicate, missing, and abnormal data are deleted to effectively extract traffic flow characteristic parameters. The basic graph function of the traffic flow is fitted by the Greenshields model, and a Bayesian network model for under-observed sections is constructed. The relationship between the nodes of the Bayesian network model is determined by the constructed basic graph function of the traffic flow, and the solution is completed to obtain the traffic flow characteristic parameters of the under-observed sections. The road sections are abstracted as nodes, the spatial similarity matrix and traffic flow similarity matrix are defined, a directed graph convolution operator is built, and the dual attention mechanism-directed graph long short-term memory network model is constructed. The traffic flow characteristics and infrastructure characteristics are comprehensively considered to realize the inference of road network traffic operation status considering under-observed sections.

[0044] This invention is oriented to solving practical problems. It uses the most widely used and easily accessible police checkpoint detection data as the data source, and uses the Bayesian model and the basic traffic flow graph model to complete the traffic flow characteristic parameters of the under-observed sections, and constructs a dual attention mechanism-directed graph long short-term memory network model to capture the traffic flow influencing factors of the upstream and downstream sections while improving the balance of the training data, and realizing the inference of the traffic operation status of the road network. Figure 1 As shown in , the specific steps of the network-level traffic operation status inference method for unobservable road sections are as follows:

[0045] S1. Through the data preprocessing method based on the public security checkpoint detection data, the abnormal data is corrected or deleted, and on this basis, the characteristic parameters of traffic flow, speed and density are extracted;

[0046] S11, matching the vehicle license plate recognition data of the upstream and downstream adjacent intersections, and eliminating the invalid license plate data that cannot be correctly read and the unreasonable license plate data that appears within the fixed observation time window;

[0047] S12, dividing the signal cycle to which the current vehicle belongs according to the timestamp of the license plate recognition data, deducting 4 seconds of starting loss from the time of the first vehicle passing the stop line in each cycle, and obtaining the green light start time of each cycle;

[0048] S13, comparing the green light on time with the actual green light on time to obtain an offset, performing normal distribution fitting on the distribution of the offset in the entire study area, selecting the offset with the highest normal distribution probability as the calibration value of the drift correction algorithm, and realizing correction of the offset error of the green light on time;

[0049] S14, select a fixed observation time window, through the formula T i =(t k+1 ,-t k ) Statistically calculate the average travel speed of observable sections in the road network, obtain the average vehicle density in the time window by the entry and exit method, and calculate the statistical distribution characteristics of the average travel speed of vehicles by the following formula: i When -10%<α<200%, the data is considered valid, otherwise the data stamp is deleted, thereby reducing abnormal data caused by on-street parking, recognition errors, etc.;

[0050]

[0051] Among them, T i represents the travel time of the i-th vehicle passing through the road section; T k represents the time when the i-th vehicle passes through the k-th intersection; Indicates the average travel speed of vehicles in the time window;

[0052] S2. Construct a traffic state estimation model based on Bayesian network and traffic flow basic graph to estimate the traffic state of under-observed sections in the road network;

[0053] S21, based on the Greenshields model Taking the traffic flow characteristic parameters of the observable road section obtained in step S14 as the data basis, the traffic flow basic graph function relationship is fitted by the least squares method to obtain the V of the observable road section. f , K j ;

[0054] Where: V f Indicates the free flow speed of the road section; K j represents the blocking flow density;

[0055] S22. Based on the principle of conservation of vehicle flow in and out of the road section, construct Figure 3 The Bayesian network graph structure shown;

[0056] S23, through the traffic flow basic graph function relationship constructed in S21, determine the relationship between the nodes in the Bayesian network model in S22. Based on the traffic flow theory, the traffic flow characteristic parameters of the under-observed road section satisfy the following formula:

[0057] Q 欠-1 =Q 入-1 +Q 入-2 +Q 入-3

[0058] Q 欠-2 =Q 入-4 +Q 入-5 +Q 欠-1

[0059] ······

[0060] Q 出 =Q 欠-n +Q n-1 +Q n

[0061]

[0062] Where: Q 欠-1 represents the traffic volume in the time window of the first under-observed road section; Q 入-1 , Q 入-2 , Q 入-3 They represent the traffic volume flowing into the first under-observed road section time window, and other parameters can be obtained similarly; k 欠-1 , k 欠-2 They represent the traffic density of the under-observed road section respectively, and other parameters can be obtained in the same way. The traffic state estimation of the under-observed road section is realized;

[0063] S3. Build a dual attention mechanism-directed graph long short-term memory network model (ADALSTM);

[0064] S31, the model constructed in this link is as follows Figure 4 As shown in the figure, we first introduce the graph attention module into the long short-term memory network LSTM, abstract each road section in the road network into a directed graph, and Apply the attention sharing mechanism, through formula e ij =a(WX i ,WX j ) calculates the attention coefficient, where e ij Represents the importance of j to road section i. The softmax function is used for normalization so that the coefficients between different nodes can be quantitatively represented. The input feature V is operated with the weight matrix introduced with the attention mechanism to obtain the new feature V′, which can explore the influence of traffic flow parameters between adjacent road sections.

[0065] S32. Introduce the attention mechanism into the long short-term memory network model, such as Figure 5 As shown, the public notice is as follows. and is the road segment feature representation obtained by LSTM, tanh is the hyperbolic tangent activation function, is the parameter of the attention layer, nw is the number of time windows in a day, Calculate the relative distance between time windows i and j and take the remainder. i,j Convert the importance score of road segment wj to road segment wi into a probability value between 0 and 1. Set the data sampling period of the day to 7:00-19:00, so that the morning and evening peaks can receive more attention and improve the effectiveness of the model;

[0066]

[0067] S4. Based on the space syntax theory, the topological value of each road section in the road network is calculated as the spatial dimension feature; the vehicle speed feature of each road section in the road network is calculated, and the two are combined to construct a directed graph convolution operator;

[0068] S41. Based on the theory of space syntax, the integration degree of each road section in the road network is calculated. The integration degree matrix is ​​used to measure the influence of the spatial dimension characteristics of the road section. The integration degree can be calculated by the formula. Define the spatial similarity matrix of the road section I∈R N*N , I is calculated using the following formula;

[0069]

[0070]

[0071] in:

[0072] MD(xi): mean depth of element xi;

[0073] RA(xi): relative asymmetry of element xi;

[0074] n: total number of elements;

[0075] RRA(xi): contrast asymmetry of element xi;

[0076] RA of Diamond: relative asymmetry value of the diamond-shaped topology;

[0077] Integration(xi): integration degree of element xi;

[0078] jm=inti-intj is the difference between sections vi and vj in the mth dimension; T is the length of the time series;

[0079] S42. Construct a speed variance matrix to measure the impact of traffic flow operation. The elements vi,j in the matrix are calculated using the following formula:

[0080]

[0081] Among them: gm=zi-zj is the difference between sections vi and vj in the mth dimension; T is the length of the time series.

[0082] S5. Input data to effectively infer the traffic operation status of the entire road network under the background of unobservable road sections in the road network, set reasonable evaluation indicators, and evaluate the effectiveness of the model;

[0083] S51. Select the three evaluation indicators commonly used in neural network algorithms, namely mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) as the model effectiveness evaluation indicators;

[0084] S52. Input actual data and evaluate the effectiveness of the dual attention mechanism introduced into the model through model ablation analysis.

[0085] S53. Compare with four neural network algorithms: ANN, CNN, LSTM, and DGCLSTM to verify the effectiveness of the model in inferring network-level traffic operation status for unobservable sections.

[0086] The technical content of the present invention is explained below with reference to a specific example.

[0087] Take the example of two sections of the road network that are completely lacking in traffic flow detection equipment. Figure 6As shown in Figure 2, the traffic flow status changes of sections A and B cannot be inferred. Taking 15 minutes as the time window, the method proposed in the present invention first completes the traffic flow characteristic parameters of the under-observed sections. The results are as follows: Figure 7 As shown, the completed data reflects the time-varying characteristics of traffic flow and the data is highly stable, so the results are credible.

[0088] The completed traffic flow characteristic parameters of the entire road section are input into the dual attention mechanism-directed graph long short-term memory network model constructed in this study. The mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are selected as three evaluation indicators. The model inference results are compared with the inference results of neural network (ANN), recurrent neural network (RNN), long short-term memory network (LSTM), and directed graph convolution-long short-term memory neural network (DGC-LSTM). Table 1 records the average prediction performance of ADALSTM and baseline models at 15 minutes, 30 minutes, and 60 minutes. Figure 8 Describe the comparison of MAE and RMSE results with the baseline model. As shown in the table, the short-term prediction performance of the ADALSTM model is better than other comparison models. Specifically, CNN cannot reflect complex spatiotemporal relationships. RNN extracts the time-dependent characteristics of data based on CNN, so its effect is better than CNN, but the model is insufficient in extracting spatial relationships in traffic networks. The DGC-LSTM model takes into account the spatiotemporal characteristics of the road network and refines the traffic characteristics of the road section. Its prediction performance is significantly better than LSTM in all three time steps, but the model does not fully consider the temporal characteristics of road network traffic flow. ADALSTM adds a temporal attention module to improve the balance of data. Its prediction performance is significantly better than DGC-LSTM in the short term, proving the effectiveness of the model.

[0089] Table 1. Comparison of model performance

[0090]

[0091] We selected three days of data from the original dataset to perform ablation analysis on the ADALSTM model to verify the effectiveness of the model’s attention mechanism. Fig. 9Ablation experiments were conducted on the ADALSTM model on two graph structures. Among them, the nonA-DALSTM model represents the original model without the spatial attention mechanism of the road section, and the AD-nonALSTM model represents the original model without the temporal attention mechanism. The results show that the ADALSTM model has better performance. The traffic flow parameters of the two graph structures have different performances during peak hours in the morning and evening, but overall the ADALSTM can achieve good prediction results and has good fitting ability. Overall, the performance of the ADALSTM model is better than that of the other baseline models. It can be seen that the dual attention mechanism and graph convolution operator can effectively improve the model's ability to extract spatiotemporal information and effectively improve the model's effect.

[0092] In summary, using real traffic flow data as verification, compared with other algorithms, the inference model proposed in this invention is superior to other models in short-term traffic flow prediction. It can effectively judge the traffic flow conditions of road sections and road networks in the future within a short period of time, and provide a basis for traffic management and control.

[0093] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest range of convergences with the principles and novel features disclosed herein.

Claims

1. A network-level traffic operation state inference method for unobservable road sections, characterized in that: The following steps are involved: S1. A data preprocessing method based on the police checkpoint detection data is used to correct or delete abnormal data, and on this basis, characteristic parameters of traffic flow, such as flow rate, speed and density, are extracted; S2. Construct a traffic flow characteristic parameter completion model based on Bayesian network and traffic flow basic graph to complete the traffic flow characteristic parameters of unobservable sections in the road network; S3. Build a directed graph long short-term memory network model based on dual attention mechanism, introduce the attention mechanism of GAT model, and capture the factors affecting the traffic flow of upstream and downstream sections; S4. Based on the space syntax theory, the topological value of each road section in the road network and the vehicle speed characteristics of each road section in the road network are calculated, wherein the topological value of each road section in the road network is used as a spatial dimension feature; the topological value of each road section in the road network is combined with the vehicle speed characteristics to construct a directed graph convolution operator; S5. Reversely input the spatial dimension features and directed graph convolution operator obtained in step S4 into the dual attention mechanism-directed graph long short-term memory network model to evaluate the effectiveness of the model.

2. The network-level traffic operation state inference method for unobservable road sections according to claim 1 is characterized in that: Step S1 includes the following sub-steps: S11, matching and comparing the traffic detection data of two adjacent public security checkpoints, and eliminating invalid license plate data that cannot be correctly read and unreasonable license plate data that appear repeatedly within a fixed observation time window; S12, dividing the signal cycle to which the current vehicle belongs according to the timestamp of the public security checkpoint detection data, deducting 4 seconds of starting loss from the time of the first vehicle passing the stop line in each cycle, and obtaining the green light start time of each cycle; S13, comparing the green light on time with the actual green light on time to obtain an offset, performing normal distribution fitting on the distribution of the offset in the entire study area, selecting the offset with the highest normal distribution probability as the calibration value of the drift correction algorithm, and realizing correction of the offset error of the green light on time; S14, selecting a fixed observation time window, counting the traffic volume, average travel speed and average vehicle density of observable sections in the road network, and eliminating abnormal data by calculating the statistical distribution characteristics of the average travel speed of vehicles.

3. The network-level traffic operation state inference method for unobservable road sections according to claim 1 is characterized in that: Step S2 includes the following sub-steps: S21. Based on the Greenshields model, the functional relationship of the basic graph of traffic flow is fitted by the least squares method; S22. Construct a Bayesian network graph structure for unobservable sections in a road network based on basic traffic flow theory; S23. Determine the relationship between nodes in the Bayesian network model of S22 through the basic graph function relationship of the traffic flow constructed in S21, and complete the traffic flow characteristic parameters of the unobservable sections in the road network.

4. The network-level traffic operation state inference method for unobservable road sections according to claim 1 is characterized in that: Step S3 includes the following sub-steps: S31, introduce the graph attention module to explore the influence of traffic flow parameters between adjacent road sections; S32. Long short-term memory network is integrated with attention mechanism to optimize the weights of the function.

5. The network-level traffic operation state inference method for unobservable road sections according to claim 1 is characterized in that: Step S4 includes the following sub-steps: S41. Based on the space syntax theory, the topological value of each road segment in the road network is calculated and used as the spatial dimension feature; S42. By obtaining the average travel speed of the vehicle on each road section, a speed variance matrix of each road section is constructed as the vehicle running speed feature, which together with the spatial dimension feature constructed in S41 constitutes the feature vector of the model input, and a directed graph convolution operator is constructed.

6. The network-level traffic operation state inference method for unobservable road sections according to claim 1 is characterized in that: Step S5 includes the following sub-steps: S51, selecting the mean absolute error, root mean square error and mean absolute error percentage in the neural network algorithm as the model effectiveness evaluation indicators; S52. Input actual data and evaluate the effectiveness of the dual attention mechanism introduced into the model through model ablation analysis.

Images