Urban road network dynamic OD estimation method based on graph convolutional neural network

Abstract

The invention discloses an urban road network dynamic OD estimation method based on a graph convolutional neural network. The method mainly comprises the steps of obtaining AVI detection data and mobile crowdsourcing trajectory data in a research road network, performing road network matching, counting the number of vehicles passing through each detection point in different time periods, and generating an AVI detection OD matrix and a mobile crowdsourcing OD matrix; modeling into graph structure data according to real dynamic OD characteristics of a road network, and constructing a daily dynamic OD relation graph based on a mobile crowdsourcing OD matrix; fusing the AVI detection OD matrix and the mobile crowdsourcing OD matrix by using a projection method, and generating a priori OD matrix as feature representation of nodes in a daily dynamic OD relation graph; the daily dynamic OD relation graph and the corresponding feature representation serve as input, a GCN model serves as a convolutional layer to construct an encoder, decoding is carried out through a full connection layer, and a dynamic OD estimation model based on a graph convolutional neural network is constructed; and calculating the total loss of the model and adjusting model parameters according to the total loss of the model to obtain a trained dynamic OD estimation model.

Description

Technical Field

[0001] The present application relates to the field of traffic information, and more specifically, to a method for estimating the dynamic OD (Origin-Destination) of an urban road network based on a graph convolutional neural network. Background Art

[0002] With the rapid advancement of urbanization and motorization in my country, urban traffic congestion is becoming increasingly serious. Using advanced road detection equipment to obtain real-time dynamic traffic basic data and scientifically analyze the causes of congestion is the key to managing urban traffic congestion. Real-time and complete road network vehicle origin-destination (OD) information is crucial for traffic management and control of urban road networks.

[0003] Traditional OD estimation methods involve multi-objective optimization modeling under many constraints, and the modeling and solution complexity is high. At the same time, due to the limitations of traditional traffic flow detection equipment, it is usually difficult to accurately calculate the OD matrix in actual applications, and it is difficult to meet the needs of intelligent traffic control in real urban road networks. With the development of emerging technologies such as electronic identification and intelligent connected vehicles, the application of emerging data sources such as mobile crowdsourcing trajectories and automatic vehicle identification (AVI) detection data has provided new possibilities for OD research. AVI equipment can record the license plate information of vehicles passing through the section, and the vehicle driving path can be restored by matching the information at different points. However, the coverage of AVI equipment is limited, and the accuracy of video recognition is limited by image processing technology, making it difficult for the path reconstruction method to work; mobile crowdsourcing data contains accurate trajectory information and a wide sampling range, but poor stability and low penetration rate of sampled vehicles, and cannot provide complete traffic spatiotemporal information. Therefore, how to effectively mine the information required for intelligent traffic control from mobile crowdsourcing trajectory data and AVI detection data, and how to use this information, is the key to optimizing urban traffic control strategies at this stage. Summary of the invention

[0004] 1. Purpose of the Invention

[0005] The present invention aims to integrate the spatiotemporal complementary characteristics of AVI detection data and mobile crowdsourcing trajectories, and proposes a dynamic OD estimation method for urban road networks based on graph convolutional neural networks. It takes advantage of multi-source data, fully mines the travel behavior characteristics and trajectory information contained in the two types of data, and comprehensively and accurately reproduces the spatiotemporal distribution of road network traffic status, providing a decision-making basis for refined traffic management and control, which has important practical significance for urban congestion relief and network optimization.

[0006] 2. Technical solution adopted by the present invention

[0007] The present invention proposes a method for estimating the dynamic OD of an urban road network based on a graph convolutional neural network. The specific steps of the method are as follows:

[0008] Step 1: Obtain AVI detection data and mobile crowdsourcing trajectory data in the research road network, perform road network matching, count the number of vehicles passing through each detection point in different time periods, and generate AVI detection OD matrix and mobile crowdsourcing OD matrix;

[0009] Step 2: Model the road network into graph structure data according to its real dynamic OD characteristics, and construct a daily dynamic OD relationship graph based on the mobile crowdsourcing OD matrix;

[0010] Step 3: Use the projection method to fuse the AVI detection OD matrix and the mobile crowdsourcing OD matrix to generate a priori OD matrix as the feature representation of the nodes in the daily dynamic OD relationship graph;

[0011] Step 4: Take the daily dynamic OD relationship graph and the corresponding feature representation as input, use the GCN (Graph Convolutional Network) model as the convolutional layer to build the encoder, decode it through the fully connected layer, and build a dynamic OD estimation model based on the graph convolutional neural network;

[0012] Step 5: Calculate the total model loss and adjust the model parameters according to the total model loss to obtain the trained dynamic OD estimation model.

[0013] Furthermore, the specific process of step 1 is as follows:

[0014] Step 1.1, obtain AVI detection data and mobile crowdsourcing trajectory data in the research road network, take the intersections in the research road network as detection point units, and perform road network matching based on the geographic location information of the AVI detection equipment and the GPS positioning information in the mobile crowdsourcing data;

[0015] Step 1.2: Count the traffic flow at each intersection at certain time intervals and time periods. Based on the AVI detection data, count the vehicle starting and ending point information of each detection point to generate the AVI detection OD matrix. Based on the mobile crowdsourcing data, extract the vehicle's travel starting and ending point information to generate the mobile crowdsourcing OD matrix.

[0016] Furthermore, the specific process of step 2 is as follows:

[0017] Step 2.1: Assuming that there are T study periods in a day, the intersections in the study road network are abstracted as nodes, the number is n, and the real OD flow is abstracted as weighted directed edges. The dynamic OD flow graph of the road network can be defined as To describe the urban road network OD, where V = {v i , i∈[1, 2, ..., n]}, t∈[1, 2, ..., T];

[0018] Step 2.2: Based on step 2.1, define the daily dynamic OD relationship diagram To describe the spatial topological structure of the road network OD, stored in the form of an adjacency matrix;

[0019] Step 2.3, aggregate the mobile crowdsourcing OD matrices obtained for multiple days by time period, divide them by the corresponding number of days, and obtain the daily average mobile crowdsourcing OD matrices for T time periods. Similarly, the daily average AVI detection OD matrices for T time periods can be calculated;

[0020] Step 2.4: Based on the T time-segment daily average mobile crowdsourcing OD matrix, use the KNN algorithm to construct the adjacency matrix. For each node vi∈V, select the node with v i The first k destination nodes with the largest OD flow from the starting point, at v i Add edges between the selected k nodes:

[0021]

[0022] Among them, KNN(v i ) returns the value of v i The first k destination nodes with the largest OD flow from the starting point are repeated for all nodes to obtain the entire edge set:

[0023]

[0024] So as to obtain the basic adjacency matrix:

[0025]

[0026] Step 2.5: Symmetrize the basic adjacency matrix and add the basic adjacency matrix to its transpose:

[0027]

[0028] In the new matrix A, let

[0029]

[0030] Step 2.6: Considering the daily cycle time characteristics of the road network OD change, repeat steps 2.3-2.5 for the T time period daily average moving crowdsourcing OD matrix to obtain the daily dynamic OD relationship diagram and its corresponding adjacency matrix representation.

[0031] Furthermore, the specific process of step 3 is as follows:

[0032] Step 3.1: Under the condition of limited data, assuming that the permeability distribution is uniform, the OD matrix is ​​detected based on the daily average AVI of different time periods. Calculate the sum of AVI detection OD matrix flow in the research road network:

[0033]

[0034] Wherein, N0 represents the set of intersections where AVI detection equipment is deployed;

[0035] Step 3.2: Crowdsourcing OD matrix based on daily average moving time periods Calculate the sum of the flow rates at the corresponding locations:

[0036]

[0037] Step 3.3: Calculate the average permeability p in different time periods t :

[0038]

[0039] Step 3.4, obtain the prior OD matrix based on the linear projection method:

[0040]

[0041] in, represents the AVI detection OD matrix of the t-th study period on a specific day, and N represents the set of all intersections in the study road network;

[0042] Step 3.5: Under the condition of sufficient data, assuming that the permeability distribution is uneven, the corresponding time-based daily average permeability matrix can be calculated based on the historical time-based daily average real OD matrix and the time-based daily average moving crowdsourcing OD matrix:

[0043]

[0044] At this time, the prior OD matrix obtained based on the linear projection method is:

[0045]

[0046] Step 3.6: Assume a graph G has N nodes, each node v i If there are P attribute features, then a feature matrix X∈R can be constructed N×P , where each row corresponds to a node in the graph G, representing the P-dimensional feature vector of the node; the prior OD matrix obtained As the graph constructed in step 2 The node feature matrix is ​​denoted as X t , then each node v iThe meaning of the j-th attribute feature is that the node v i As the starting point, with node v j is the OD flow value at the end point.

[0047] Furthermore, the specific process of step 4 is as follows:

[0048] Step 4.1: During a certain research period, the adjacency matrix A∈R obtained in steps 2 and 3 is N×N and the feature matrix X∈R N×N Input to the first graph convolution layer of the encoder, in which the graph convolution operation is performed to obtain the first layer hidden representation X (1) :

[0049]

[0050] in, I N is the unit matrix, N represents the number of nodes, for The degree matrix, W (0) is the convolution kernel weight matrix of the first layer, σ(·) is the nonlinear activation function, and the Softplus function is selected as the activation function. The specific formula is:

[0051] Softplus(x)=log(1+e x ) (13)

[0052] Each subsequent graph convolutional layer performs a convolution operation on the hidden representation of the previous layer:

[0053]

[0054] The encoder contains a total of L graph convolution layers. Dropout layers are used between layers to alleviate overfitting. The last layer only performs graph convolution operations. The output X (L) That is the learned node embedding vector;

[0055] Step 4.2: Embed the node into vector X (L) The input is sent to the decoder consisting of a fully connected layer, and after being transformed by the L′ layer of fully connected layers, the output features are generated:

[0056] X′ (l) =σ(X′ (l-1) W (l) +b (l) ) (15)

[0057] Where W (l) and b (l)are the weight and bias parameters of the first layer, respectively, and σ(·) is a nonlinear activation function; the last fully connected layer uses the Softplus activation function and sets the diagonal elements to 0 to prevent self-connection, and outputs the target vector C is the dimension of the prediction target. According to the meaning of the node feature matrix in step 3.6, we know that C = N. That is the estimated dynamic OD matrix of the urban road network.

[0058] Furthermore, the specific process of step 5 is as follows:

[0059] According to the constructed dynamic OD estimation model, the processed research road network data is divided into a training set and a validation set to train the model, and the node feature matrix output is calculated. The difference with the real road network OD matrix defines the objective loss function:

[0060]

[0061] in, Represents the reconstructed node feature matrix The value of an element in , yi represents the value of the OD flow at the corresponding position in the real road network OD matrix, N represents the number of nodes, and the model parameters are adjusted according to the total loss of the model to obtain the trained dynamic OD estimation model.

[0062] Compared with the prior art, the dynamic OD estimation method based on graph convolutional neural network in the embodiment of the present invention has the following beneficial effects:

[0063] 1. Utilizing the spatiotemporal complementary characteristics of AVI detection data and crowdsourced trajectory data, multi-source data fusion is performed to solve the limitation of single data source in traditional OD estimation and significantly improve the accuracy and real-time performance of dynamic OD estimation;

[0064] 2. The dynamic OD estimation model based on graph convolutional neural network can capture the complex spatiotemporal relationship of the road network structure. Even when the data is limited, the model can still achieve high-precision estimation;

[0065] 3. The present invention can be applied to urban road networks of different sizes and can be extended to different traffic detection equipment data sources, and has strong practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present application and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention:

[0067] Figure 1 is a flow chart of the method of the present invention;

[0068] Figure 2It is a flow chart of a specific method of the present invention;

[0069] Figure 3 A schematic diagram of the road network space range covered by the verification data according to a specific embodiment of the present invention;

[0070] Figure 4 It is a schematic diagram of the structure of a dynamic OD estimation model based on a graph convolutional neural network according to a specific embodiment of the present invention;

[0071] Figure 5 The mean absolute error and root mean square error of the model under different AVI device coverage rates described in the specific implementation mode of the present invention;

[0072] Figure 6 The mean absolute percentage error and mean square percentage error of the model under different AVI device coverage described in the specific embodiment of the present invention; DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0075] Unless otherwise specifically stated, the relative arrangement of the parts and steps described in these embodiments, numerical expressions and numerical values ​​do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to the actual proportional relationship. The technology, methods and equipment known to ordinary technicians in the relevant field may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be regarded as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0076] like Figure 1 As shown, the present invention proposes a dynamic OD estimation method based on graph convolutional neural network. Figure 2 Another specific flow chart of an embodiment of the present invention is shown. Figure 1 A dynamic OD estimation method based on a graph convolutional neural network is shown, which may specifically include the following steps:

[0077] Step 1: Obtain AVI detection data and mobile crowdsourcing trajectory data in the research road network, perform road network matching, count the number of vehicles passing through each detection point in different time periods, and generate AVI detection OD matrix and mobile crowdsourcing OD matrix.

[0078] As an example, the area within the second ring road of Baoding City, Hebei Province is selected as an example verification road network, and 36 major intersections in the road network are selected as the starting and ending points of vehicle travel. The intersections are numbered as follows: Figure 3 As shown. Among them, AVI detection data comes from the data collected by AVI equipment deployed at 36 intersections in the research road network. Due to damage to AVI equipment and other reasons, some data are lost. The effective collection period is from 9 am to 1 pm every day; crowdsourced trajectory data comes from crowdsourced vehicle data in the Baoding road network collected by shared travel and path navigation software, covering all intersections in the research road network. The specific steps are as follows:

[0079] Step 1.1, taking the intersections in the research road network as detection point units, and performing road network matching based on the geographic location information of the AVI detection equipment and the GPS positioning information in the mobile crowdsourcing data;

[0080] Step 1.2: Count the traffic flow at each intersection in time periods according to the 30-min time interval. Based on the AVI detection data, count the vehicle starting and ending point information of each detection point, and obtain the AVI detection OD matrix of 14 study periods every day. Based on the mobile crowdsourcing data, extract the vehicle's travel starting and ending point information to generate the mobile crowdsourcing OD matrix. Generate the mobile crowdsourcing OD matrix corresponding to the 14 study periods every day.

[0081] Step 2: According to the real dynamic OD characteristics of the road network, it is modeled as graph structure data, and a daily dynamic OD relationship graph is constructed based on the mobile crowdsourcing OD matrix.

[0082] As an example, this step may include the following steps:

[0083] Step 2.1: Subject to the limitation of the effective daily collection period of AVI data in the embodiment, there are 14 effective research periods in a day. The 36 intersections in the research road network are abstracted as nodes, and the real OD flow is abstracted as weighted directed edges. The dynamic OD flow graph of the road network can be defined. To describe the urban road network OD, where V = {v i ,i∈[1,2,…,36]}, t∈[1,2,…,14];

[0084] Step 2.2: Based on step 2.1, define the daily dynamic OD relationship diagram To describe the spatial topological structure of the road network OD, stored in the form of an adjacency matrix;

[0085] Step 2.3, aggregate the mobile crowdsourcing OD matrices obtained for multiple days by time period, divide them by the corresponding number of days, and obtain the daily average mobile crowdsourcing OD matrices for 14 time periods. Similarly, the daily average AVI detection OD matrices for 14 time periods can be calculated;

[0086] Step 2.4: Based on the daily average mobile crowdsourcing OD matrix of 14 time periods, the KNN algorithm is used to obtain the basic adjacency matrix through equations (1)-(3);

[0087] Step 2.5, symmetrize the basic adjacency matrix and obtain the adjacency matrix through equations (4)-(5);

[0088] Step 2.6: Considering the daily cycle time characteristics of the road network OD change, repeat steps 2.3-2.5 for the daily average moving crowdsourcing OD matrix of 14 time periods to obtain the daily dynamic OD relationship diagram and its corresponding adjacency matrix representation.

[0089] Step 3: Use the projection method to fuse the AVI detection OD matrix and the mobile crowdsourcing OD matrix to generate a priori OD matrix as the feature representation of the nodes in the daily dynamic relationship graph.

[0090] As an example, this step may include the following steps:

[0091] Step 3.1: Under the condition of limited data, assuming that the permeability distribution is uniform, the OD matrix is ​​detected based on the daily average AVI of different time periods. The sum of the AVI detection OD matrix flow in the research road network is calculated by formula (6);

[0092] Step 3.2: Crowdsourcing OD matrix based on daily average moving time periods The sum of the mobile crowdsourcing OD matrix flow at the corresponding position is calculated by formula (7);

[0093] Step 3.3: Calculate the average permeability p by time period using formula (8): t ;

[0094] Step 3.4: Obtain the prior OD matrix based on the linear projection method through formula (9):

[0095] Step 3.5: Under the condition of sufficient data, assuming that the permeability distribution is uneven, the corresponding time-divided daily average permeability matrix is ​​obtained based on the historical time-divided daily average real OD matrix and the time-divided daily average mobile crowdsourcing OD matrix through equations (10)-(11): and the prior OD matrix

[0096] Step 3.6: Obtain the prior OD matrix As the graph constructed in step 2 The node feature matrix is ​​denoted as X t , then each node v i The meaning of the j-th attribute feature is that the node v i As the starting point, with node v j is the OD flow value at the end point.

[0097] Step 4: Take the daily dynamic OD relationship graph and the corresponding feature representation as input, use the GCN model as the convolutional layer to build the encoder, decode it through the fully connected layer, and build a dynamic OD estimation model based on graph convolutional neural network.

[0098] The dynamic OD estimation model based on graph convolutional neural network proposed in this embodiment is as follows: Figure 4 shown.

[0099] As an example, this step may include the following steps:

[0100] Step 4.1: During a certain research period, the adjacency matrix A∈R obtained in steps 2 and 3 is 36×36 and the feature matrix X∈R 36×36Input into the encoder, perform graph convolution operation through equations (12)-(14) to obtain the node embedding vector X (L) ;

[0101] Step 4.2: Embed the node into vector X (L) Input to the decoder composed of fully connected layers, transform the L′ layer of fully connected layers through formula (15), and output the target vector C is the dimension of the prediction target. According to the meaning of the node feature matrix in this embodiment, we know that C=36. That is the estimated dynamic OD matrix of the urban road network.

[0102] Step 5: Calculate the total model loss and adjust the model parameters according to the total model loss to obtain the trained dynamic OD estimation model.

[0103] As an example, the processed research road network data is divided into a training set and a validation set to train the model. The model parameters are adjusted by the loss function defined in formula (16) to obtain a trained dynamic OD estimation model.

[0104] In order to verify the effectiveness of the present invention, the following experiments are conducted, which may specifically include the following aspects:

[0105] First, regarding the evaluation indicators, in order to comprehensively evaluate the dynamic OD estimation model of the present invention, four indicators, namely, mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and mean squared percentage error (MSPE), are used to evaluate the difference between the actual OD flow of the road network and the estimated value.

[0106] Secondly, regarding the experimental environment, this experiment was conducted under the Windows 10 operating system. Python was selected as the programming language, version 3.9, and the deep learning framework was implemented based on the Pytorch framework, version 1.13.1.

[0107] Thirdly, regarding the comparative experiment, this experiment verifies the optimization effect of the dynamic OD estimation model proposed in the present invention on the prior OD obtained by projection under the conditions of uniform permeability distribution and non-uniform permeability distribution. The model evaluation results are shown in Table 1. In addition, this experiment also explores the estimation performance of the model under different AVI coverage rates. The results are shown in Table 1. Figure 5 , 6 shown.

[0108] Table 1

[0109]

[0110] It can be seen from the data in Table 1 that the prior OD obtained by the projection method has a large error with the true OD, which is difficult to meet the needs of actual applications. After optimization by the dynamic OD estimation method proposed in the present invention, the accuracy is significantly improved, which can meet the needs of dynamic OD estimation of road networks. Experimental results show that under the assumption of uniform permeability, the accuracy is significantly improved after the OD estimation method proposed in the present invention is optimized. Among them, MAE decreased by 72.4%, RMSE decreased by 70.7%, MAPE decreased by 51.2%, and MSPE decreased by 73.4%, and the optimization effect was significant; at the same time, the MAE value was 15.94veh / 30min, and the RMSE value was 33.37veh / 30min, which can meet the accuracy requirements of dynamic OD estimation of road networks. Under the assumption of uneven permeability distribution, the error between the prior OD obtained by the projection method and the true OD is lower than that under uniform permeability conditions, with MAE of 23.99veh / 30min, but RMSE, MAPE, and MSPE are 44.79veh / 30min, 45.85%, and 61.06%, respectively, indicating that the estimation effect is relatively poor for the case where the true OD flow value is large, and there may be outliers. After the OD estimation method proposed in the present invention is optimized, the estimation accuracy is significantly improved. Obviously, the assumption of uneven permeability distribution is closer to the real road scene, and adding the prior information of the permeability distribution matrix can improve the estimation accuracy to a certain extent.

Claims

1. A method for estimating dynamic OD of urban road networks based on graph convolutional neural network, characterized in that: The following steps are involved: Step 1: Obtain AVI detection data and mobile crowdsourcing trajectory data in the research road network, perform road network matching, count the number of vehicles passing through each detection point in different time periods, and generate AVI detection OD matrix and mobile crowdsourcing OD matrix; Step 2: Model the road network into graph structure data according to its real dynamic OD characteristics, and construct a daily dynamic OD relationship graph based on the mobile crowdsourcing OD matrix; Step 3: Use the projection method to fuse the AVI detection OD matrix and the mobile crowdsourcing OD matrix to generate a priori OD matrix as the feature representation of the nodes in the daily dynamic OD relationship graph; Step 4: Take the daily dynamic OD relationship graph and the corresponding feature representation as input, use the GCN model as the convolutional layer to build the encoder, decode it through the fully connected layer, and build a dynamic OD estimation model based on graph convolutional neural network. Step 5: Calculate the total model loss and adjust the model parameters according to the total model loss to obtain the trained dynamic OD estimation model.

2. According to the method for estimating dynamic OD of urban road network based on graph convolutional neural network in claim 1, in step 1, the process of extracting OD flow from AVI detection data and mobile crowdsourcing trajectory data comprises: (1) Obtain AVI detection data and mobile crowdsourcing trajectory data in the study road network, take the intersections in the study road network as detection point units, and perform road network matching based on the geographic location information of the AVI detection equipment and the GPS positioning information in the mobile crowdsourcing data; (2) The traffic flow at each intersection is counted in time periods according to certain time intervals. Based on the AVI detection data, the vehicle starting and ending point information of each detection point is counted to generate the AVI detection OD matrix. Based on the mobile crowdsourcing data, the vehicle travel starting and ending point information is extracted to generate the mobile crowdsourcing OD matrix.

3. According to the method for estimating dynamic OD of urban road network based on graph convolutional neural network in claim 1, in step 2, the process of constructing a daily dynamic OD relationship diagram based on the mobile crowdsourcing OD matrix comprises: (1) The intersections in the research road network are abstracted as nodes, the OD flow is abstracted as weighted directed edges, and a dynamic OD flow graph of the road network is constructed; (2) Define the daily dynamic OD relationship graph and store the spatial topological structure of the road network OD in the form of an adjacency matrix; (3) The mobile crowdsourcing OD matrix obtained for multiple days is aggregated by time period, divided by the corresponding number of days, and the daily average mobile crowdsourcing OD matrix and the daily average AVI detection OD matrix of each time period are calculated; (4) Based on the KNN algorithm, an adjacency matrix is ​​constructed, and the first k neighbor nodes with the largest OD flow of each node are selected to generate a daily dynamic OD relationship graph.

4. According to the method for estimating dynamic OD of urban road network based on graph convolutional neural network in claim 1, in said step 3, obtaining the feature representation of nodes in the daily dynamic OD relationship graph of the prior OD matrix comprises: (1) Under the condition of limited data, assuming that the permeability distribution is uniform, the sum of the AVI detection OD flow in the study network is calculated based on the daily average AVI detection OD matrix in different time periods; (2) Calculate the sum of OD flows at the corresponding locations based on the daily average mobile crowdsourcing OD matrix in different time periods; (3) Calculate the average permeability p in different time periods t , and use the linear projection method to generate the prior OD matrix; (4) Under the condition of sufficient data, assuming that the permeability distribution is uneven, the corresponding time-based daily average permeability matrix can be calculated based on the historical time-based daily average real OD matrix and the time-based daily average mobile crowdsourcing OD matrix. Thus, the prior OD matrix is ​​generated based on the linear projection method.

5. According to the method for estimating dynamic OD of urban road network based on graph convolutional neural network according to claim 1, in step 4, building a dynamic OD estimation model based on graph convolutional neural network comprises: (1) Inputting the adjacency matrix and node feature matrix of the daily dynamic OD relationship graph into the encoder, performing graph convolution operation, aggregating node features, and generating node embedding vectors; (2) The node embedding vector is decoded through the fully connected layer and the target OD matrix is ​​output.

6. According to the method for estimating dynamic OD of urban road network based on graph convolutional neural network according to claim 1, in step 5, it is characterized in that: The formula for calculating the total loss of the model is: in, Represents the reconstructed node feature matrix The value of an element in , y i It represents the value of OD flow at the corresponding position in the real road network OD matrix, N represents the number of nodes, and the model parameters are adjusted according to the total loss of the model to obtain the trained dynamic OD estimation model.

Images