Expressway traffic distribution method and system based on neural network

Abstract

The invention provides an expressway traffic distribution method and system based on a neural network. According to the method, a traffic distribution OD matrix is extracted by using vehicle toll collection information detected by an ETC portal of a highway network in a year and a path information database, and a deep learning model based on a deep neural network containing graph attention is constructed in combination with population data, economic development data and travel impedance data of a district-county-level traffic zone, so that the traffic distribution of the district-county-level traffic zone is obtained. And inputting the data into the model for training so as to optimize parameters of the gravity model and obtain a reasonable gravity model. And finally, realizing an expressway path selection model in combination with an XGBoost algorithm, and dividing traffic flow into a road network. According to the method, fusion of multi-source data, the deep neural network and a machine learning algorithm is realized, the precision of generation, prediction and distribution of the traffic distribution flow is greatly improved, and powerful support is provided for long-term planning management of the expressway.

Description

Technical Field

[0001] The invention relates to a highway traffic distribution method and system, in particular to a highway traffic distribution method and system based on a neural network. Background Art

[0002] Expressways are an important means of rapid regional inter-city transportation. The distribution of expressway traffic volume reflects the intensity of urban economic and social activities to a certain extent. Their construction has an increasingly profound impact on the process of urban modernization. Therefore, with the rapid changes and development of economic, population, land use and other factors, the development of expressways is constantly encountering new requirements.

[0003] Traffic distribution prediction models are divided into two categories according to the prediction range: one is the growth coefficient model that mainly predicts medium- and short-term traffic distribution, including the simple growth coefficient model, Fratar model, Furness model, Detroit model, etc.; the other is the structural model that predicts medium- and long-term traffic distribution, including the gravity model and entropy model.

[0004] The growth coefficient model assumes that the traffic generation and attraction of the traffic community will grow at a specific coefficient in the future. The calculation methods of the growth coefficient of different models are different. This type of model is simple and easy to understand, but it requires a complete basic matrix. At the same time, the model assumes that traffic resistance is fixed and does not consider the impact of reduced transportation costs, improved infrastructure, and traffic congestion. Once the land use and economic structure have undergone major changes within the study period, the prediction accuracy will be greatly reduced. Therefore, the growth coefficient model is suitable for short-term traffic distribution prediction in areas with relatively stable traffic growth conditions.

[0005] The gravity model was first used to predict traffic distribution in 1955. It is still the most popular and widely used model. However, according to the classical gravity model, when the travel cost is zero (or extremely low), the travel exchange volume will grow too fast and must be improved.

[0006] With the increasing deployment of highway node detectors (such as ETC toll detectors, radar detectors, etc.), monitoring data can more accurately reflect traffic flow characteristics, the overall status and mode of the highway network. The traditional OD data survey method has high labor costs, small sample sizes, bias, and complex travel scenarios. ETC data can obtain the starting and ending points of each vehicle in the highway network at a low cost to generate complete, timely, and accurate OD travel data, providing a sufficient and complete data foundation for further research on transportation planning. Summary of the invention

[0007] Aiming at the deficiencies of the prior art, the present invention provides a highway traffic distribution method and system based on a neural network.

[0008] In a first aspect, the present invention provides a highway traffic assignment method based on a neural network, specifically:

[0009] c1. Obtain the vehicle toll dataset and path information dataset of the district-level ETC gantry detection; divide the vehicle into three types according to the attributes of the whole sample, cars and trucks, and obtain different types of district-level traffic distribution OD matrices and OD distribution probability value matrices.

[0010] c2. Obtain district-level traffic community population data, traffic community economic development data, and traffic community travel impedance data, combine them with the OD distribution probability value matrix into a data set, and divide it into training, verification, and test sets.

[0011] c3. Build a deep learning gravity model, use the historical traffic distribution OD matrix, traffic community population data, economic development data, and travel impedance data as model input, use the obtained probability values ​​as the model output label set, perform supervised training, and iterate until the model effect is optimal.

[0012] c4. Test the model in the test set to obtain the latest district-level traffic distribution OD matrix status, as well as traffic community population data, economic development data, and travel impedance data, and predict the future traffic community traffic distribution OD matrix.

[0013] c5. Build a highway path selection model based on the XGBoost algorithm, distribute the predicted traffic distribution OD matrix to the road network, and realize the allocation of traffic demand.

[0014] In a second aspect, the present invention provides a highway traffic assignment system based on a neural network, comprising:

[0015] The data acquisition module is used to obtain the vehicle charging data set and path information data set of the district and county level ETC gantry detection; divide the vehicle into three types according to the attributes of the full sample, cars and trucks, and process different types of district and county level traffic distribution OD matrices and OD distribution probability value matrices.

[0016] The data combination module is used to obtain the district-level traffic community population data, traffic community economic development data, and traffic community travel impedance data, and combine them with the OD distribution probability value matrix into a data set, and divide it into training, verification and test sets.

[0017] The deep learning gravity model construction module is used to take the historical traffic distribution OD matrix, traffic community population data, economic development data, and travel impedance data as model inputs, and use the obtained probability values ​​as the model output label set for supervised training and iteration until the model effect is optimal.

[0018] The deep learning gravity model prediction module is used to test the model in the test set, obtain the latest district-level traffic distribution OD matrix status, as well as traffic community population data, economic development data, travel impedance data, and predict the future traffic community traffic distribution OD matrix.

[0019] The traffic demand allocation module is used to build a highway path selection model based on the XGBoost algorithm, allocate the predicted traffic distribution OD matrix to the road network, and realize the allocation of traffic demand.

[0020] The present invention has the following beneficial effects:

[0021] First, based on the traditional method, the present invention not only relies on the two variables of population and spatial distance, but also adds features such as land use, socio-economic attributes and transportation network to improve the prediction accuracy of actual transportation demand. Adding more features can better simulate the diverse travel behaviors in reality and improve the adaptability of the model to complex situations. At the same time, these features can reflect the dynamic changes of social and urban development, making the model more forward-looking.

[0022] Second, the present invention uses a deep learning algorithm to achieve the fusion of multi-source data features and the prediction of traffic distribution. It can obtain district and county characteristics through multi-source input data and explain the hidden economic radiation effects of districts and counties.

[0023] Third, the present invention performs highway path selection based on the XGBoost model and adopts a data-driven approach to find an adaptive and reasonable path allocation principle, which can more accurately reflect the operation of the traffic network and the distribution of traffic flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is an overall schematic diagram of the method in the embodiment of the present application;

[0025] Figure 2a It is the OD distribution map of traffic flow of all samples; Figure 2b It is the OD distribution map of car traffic flow; Figure 2c It is the OD distribution map of truck traffic distribution. DETAILED DESCRIPTION

[0026] The method of the present invention is further described below with reference to the accompanying drawings and specific examples.

[0027] As a traditional method of traffic distribution, the gravity model relies only on two variables: population and spatial distance. The lack of characteristics such as land use, socioeconomic attributes, and traffic networks limits the accurate prediction of actual traffic demand. In order to generate more realistic traffic demand, the method of the present invention proposes a traffic distribution model based on the gravity model and the graph attention deep neural network. The model integrates the gravity model with the deep neural network to predict traffic distribution, verifying that land use and economic factors are crucial to attracting traffic flow between regions. By introducing multi-source data and using deep neural network fitting, the generation and prediction accuracy of traffic distribution flow is greatly improved, aiming to provide strong support for the long-term planning and management of highways.

[0028] like Figure 1 As shown, the embodiment of the present application provides a highway traffic assignment method based on a neural network, the method comprising the following steps:

[0029] c1. Obtain the vehicle toll dataset and path information dataset of the district-level ETC gantry detection, divide them into three types according to vehicle attributes: full samples, cars, and trucks, and process them to obtain different types of district-level traffic distribution OD matrices, which are further processed into OD distribution probability value matrices.

[0030] c2. Obtain district-level traffic community population data, economic development data, and travel impedance data, combine them with the OD distribution probability value matrix into a data set, and divide it into training, validation, and test sets.

[0031] c3. Build a deep learning gravity model, use the historical traffic distribution OD matrix, traffic community population data, economic development data, and travel impedance data as model input, use the obtained probability value as the model output labeler, perform supervised training, and iterate until the model effect is optimal.

[0032] c4. Test the model in the test set to obtain the latest district-level traffic distribution OD matrix status, as well as traffic community population data, economic development data, and travel impedance data, and predict the future traffic community traffic distribution OD matrix;

[0033] c5. Build a highway path selection model based on XGBoost, distribute the predicted traffic distribution OD matrix to the road network, and realize the allocation of traffic demand.

[0034] In the above technical solution, the data acquired in step c1 includes:

[0035] (1) The vehicle toll collection dataset detected by the ETC entrance and exit toll gates at the district and county level, and the district and county level traffic distribution OD matrix are obtained according to the starting and ending toll stations;

[0036] (2) The vehicle path dataset detected by the toll gantries of ETC sections at the district and county level, and the vehicle travel path information dataset is obtained based on the location of the gantries;

[0037] (3) Vehicle ETC identification ID code to obtain vehicle category information.

[0038] Furthermore, in step c1, the travel path dataset is divided into paths by probability through a probabilistic operation, and the probability of a vehicle selecting a travel path under the same start and end conditions is obtained.

[0039] The data obtained in step c2 includes:

[0040] (1) Population data of the transportation community, including population size, population growth rate, and land area;

[0041] (2) Economic development data of the transportation community, including GDP, primary industry, secondary industry, and tertiary industry;

[0042] (3) Travel impedance data of transportation zones, including the average travel time and travel cost between transportation zones.

[0043] Land use density affects the intensity and direction of travel demand. For example, high-density commercial or office areas attract more external traffic. Socioeconomic attributes affect individual travel mode choices. For example, high-income groups may be more inclined to use private cars. Traffic network characteristics affect traffic flow and travel patterns.

[0044] Furthermore, in step c2, the data is normalized by the Min-Max normalization method. For the above three types of data X, the standardized input is obtained after normalization. As follows:

[0045]

[0046] Where X represents the input vector, X * represents the normalized input vector, represents the minimum value of the input vector, Indicates the maximum value of the input vector.

[0047] Then, it forms a data set together with the district-level traffic distribution OD matrix and is divided into training set, validation set and test set according to a certain ratio (6:2:2).

[0048] In a certain example, two indicators for evaluating the distribution prediction model are proposed: CPC index (Common Part of Commuters) and weighted mean absolute percentage error (WMAPE) as shown in the formula:

[0049]

[0050] in, is the actual flow rate, To predict traffic, For traffic area To the traffic area The weight of the traffic flow between them.

[0051] Step c3 includes:

[0052] c31, input and output construction

[0053] Construct the model input, including the processed traffic community population data, economic development data, and travel impedance data in c2, so as to obtain the list input. For the district-level traffic distribution OD matrix obtained in c1, by calculating the probability in the row direction and column direction, two probability distribution matrices are obtained and used as the output of the model.

[0054] c32. Building a deep learning model

[0055] For the historical traffic zone feature input, the fusion vector is obtained through the graph attention layer For the probability prediction stage input, the probability value of the target is obtained through multiple 1×1 one-dimensional convolution layers, graph attention convolution layers and two-dimensional softmax layers in sequence. The model calculation process is as follows:

[0056] In the historical traffic zone feature input stage, for any Has The characteristics of the traffic zone are , calculate any two points according to the formula The correlation score is . Using the adjacency matrix Constrain the feasible range of association scores and calculate the final attention weight parameters Finally, multi-head attention is used to calculate the new city feature matrix that integrates the feature information of surrounding cities. , the calculation formula is as follows.

[0057]

[0058] In the probability prediction stage, based on the updated city feature matrix Constructing any OD pair Input parameters , stack multiple one-dimensional convolution modules to calculate the association score at this time , then Perform graph attention convolution to obtain the attraction coefficient between OD pairs. Then use the two-dimensional Softmax operation to calculate the final multi-head attention probability distribution prediction output result. The row direction probability is , the sum of the probabilities of each row is 1, and the probability in the column direction is , the sum of the probabilities of each column is 1; the calculation formula is as follows:

[0059]

[0060] c33, model training and validation

[0061] The model is trained using the training set and verified using the validation set to improve the cross entropy loss function for back propagation calculation of model loss, and the model convergence is finally achieved through continuous iteration.

[0062]

[0063] In step c4, the model is tested in the test set to obtain the latest district-level traffic distribution OD matrix status and traffic community population data, economic development data, travel impedance data, and predict the future traffic community traffic distribution OD matrix.

[0064] In step c5, the vehicle travel path probability selection data is input to train the XGBoost algorithm to obtain a highway path selection model based on XGBoost, and the predicted traffic distribution OD matrix is ​​distributed to the road network to realize the distribution of traffic demand.

[0065] Based on the same concept of the above method embodiment, the present application also provides a highway traffic assignment system based on a neural network, including:

[0066] The data acquisition module is used to obtain the vehicle toll data set and path information data set detected by the district and county ETC gantry; divide the vehicle into three types according to the attributes of the whole sample, cars and trucks, and process different types of district and county traffic distribution OD matrices and OD distribution probability value matrices;

[0067] A data combination module is used to obtain the population data of the district-level traffic community, the economic development data of the traffic community, and the travel impedance data of the traffic community, and combine them with the OD distribution probability value matrix into a data set, and divide it into training, verification and test sets;

[0068] The deep learning gravity model building module is used to take the historical traffic distribution OD matrix, traffic community population data, economic development data, and travel impedance data as model inputs, and use the obtained probability values ​​as the model output label set for supervised training and iteration until the model effect is optimal;

[0069] The deep learning gravity model prediction module is used to test the model in the test set, obtain the latest district-level traffic distribution OD matrix status, as well as traffic community population data, economic development data, travel impedance data, and predict the future traffic community traffic distribution OD matrix;

[0070] The traffic demand allocation module is used to build a highway path selection model based on the XGBoost algorithm, allocate the predicted traffic distribution OD matrix to the road network, and realize the allocation of traffic demand.

[0071] Example: Taking the traffic flow OD data, travel route selection data and district and county attribute data collected by the ETC gantry as an example, highway OD demand prediction and allocation are performed:

[0072] 1. Extraction of OD matrix of real ETC data travel distribution

[0073] The vehicle toll information and route information detected by the ETC gantries of 87 county-level cities on the expressway network of a province within one year were used. The collected data were divided into three types of sample data: full sample, car and truck according to vehicle type. The traffic distribution OD matrix with a dimension of 87×87 was processed and the corresponding OD distribution probability matrix was calculated. The results are shown in the figure. Figure 2a , Figure 2b and Figure 2c shown.

[0074] 2. Traffic community data acquisition

[0075] The population size, population growth rate and land area of ​​the required districts and counties can be obtained through the administrative division information query platform; public and authoritative information such as GDP, primary industry, secondary industry and tertiary industry can be obtained through the national economic and social development bulletin; the average travel time and travel expenses between traffic zones can be calculated based on the charging system to obtain the most efficient and accurate traffic impedance data.

[0076] 3. Prediction model input and output construction

[0077] Traffic community population data, economic development data and travel impedance data are selected and used as input after Min-Max normalization processing. The traffic distribution OD matrix obtained by processing the ETC gantry data is used as the model output. The training set, validation set and test set are divided into training set, validation set and test set according to the ratio of 6:2:2. The model is trained with the training set data, and the model convergence and effect are verified through the validation set.

[0078] 4. Evaluation of prediction model performance

[0079] The trained model is used to predict the traffic distribution OD matrix between traffic zones. The prediction results of the model are compared with other models as shown in the table:

[0080]

[0081] It is obvious that TDP-MDFGAT has achieved the best performance on this OD dataset. The CPC of the traffic flow of the entire sample is 0.897, which is higher than other models. The improvement in W_MAPE is more obvious, with a comprehensive traffic flow of 20.64%, while other models have dropped significantly. The comprehensive traffic flow CPC of the traffic distribution predicted by the model reaches 0.897, which shows that the generated model can basically realize the simulation of traffic flow on county-level highways and roughly grasp the traffic flow trend. From W_MAPE=20.64%, it can also be seen that the prediction error of the overall traffic flow distribution is small, and the result is close to the true value. Compared with the traditional traffic distribution, the smaller division range of traffic cells makes the population flow more random, which makes the traffic distribution more difficult to predict. This application analyzes and predicts the traffic flow from a more macro perspective on the study of urban-level highway network travel, with a clearer object and higher prediction accuracy.

[0082] 5. Feature Importance Analysis

[0083] The SHAP model can be used to analyze the importance of different features, extract the importance of each feature in the model, and find that the distance feature has the highest importance.

[0084] 6. Highway route assignment model

[0085] The traffic distribution OD matrix obtained by the prediction model is used as the input of the allocation model to allocate travel demand to the highway network.

[0086] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A highway traffic assignment method based on neural network, It is characterized in that The method comprises the following steps: c1. Obtain the vehicle toll dataset and path information dataset detected by ETC gantry at the district and county level; divide the vehicle into three types: full sample, car and truck according to vehicle attributes, and obtain different types of district and county-level traffic distribution OD matrix and OD distribution probability value matrix; c2. Obtaining district-level traffic community population data, traffic community economic development data, and traffic community travel impedance data, and combining them with the OD distribution probability value matrix into a data set, and dividing them into training, validation, and test sets; c3. Build a deep learning gravity model, use the historical traffic distribution OD matrix, traffic community population data, economic development data, and travel impedance data as model input, use the obtained probability value as the model output label set, perform supervised training, and iterate until the model effect is optimal; c4. Test the model in the test set to obtain the latest district-level traffic distribution OD matrix status, as well as traffic community population data, economic development data, and travel impedance data, and predict the future traffic community traffic distribution OD matrix; c5. Build a highway path selection model based on the XGBoost algorithm, distribute the predicted traffic distribution OD matrix to the road network, and realize the allocation of traffic demand.

2. A highway traffic assignment method based on neural network according to claim 1, It is characterized in that The data obtained in step c1 includes: (1) The vehicle toll collection dataset detected by the ETC entrance and exit toll gates at the district and county level, and the district and county level traffic distribution OD matrix are obtained according to the starting and ending toll stations; (2) The vehicle path dataset detected by the toll gantries of ETC sections at the district and county level, and the vehicle travel path information dataset is obtained based on the location of the gantries; (3) Vehicle ETC identification ID code to obtain vehicle category information.

3. A highway traffic assignment method based on neural network according to claim 2, It is characterized in that In step c1, the travel path information data set is divided into paths by probability through a probabilistic operation to obtain the probability of a vehicle selecting a travel path under the same starting and ending conditions.

4. A highway traffic assignment method based on neural network according to claim 1, It is characterized in that In step c2, The traffic community population data includes population size, population growth rate and land area; The economic development data of the traffic community include GDP, primary industry, secondary industry and tertiary industry; The transportation zone travel impedance data includes the average travel time and travel cost between transportation zones.

5. A highway traffic assignment method based on neural network according to claim 4, It is characterized in that The population data, economic development data and travel impedance data of traffic districts are standardized through the Min-Max normalization method.

6. A highway traffic assignment method based on neural network according to claim 1, It is characterized in that Step c3 includes: c31, input and output construction Model input: The population data, economic development data, and travel impedance data of the traffic district processed in step c2, thereby obtaining a list input; Model output: For the district-level traffic distribution OD matrix obtained in step c1, two probability distribution matrices are obtained by calculating the probability in the row direction and the column direction, and they are used as the output results of the model; c32. Model building The model obtains the fusion vector of input features through the graph attention layer; For the probability prediction stage: the probability value of the target is obtained through multiple 1×1 one-dimensional convolution layers, graph attention convolution layers and two-dimensional softmax layers in sequence; c33, model training and validation The model is trained using the training set and verified using the validation set. The cross entropy loss function is used for back propagation calculation of the model loss, and the model convergence is finally achieved through continuous iteration.

7. A highway traffic assignment method based on neural network according to claim 1, It is characterized in that In step c4, the test set is input into the trained model to achieve the prediction of the traffic distribution OD matrix of the target year.

8. A highway traffic assignment method based on neural network according to claim 1, It is characterized in that The method also includes training the highway path selection model in step c5. The input data in the training process is the vehicle travel path probability selection data, including the number of different paths selected by the traveling vehicles between all toll station OD point pairs.

9. A highway traffic assignment method based on neural network according to claim 1, It is characterized in that In step c5, the vehicle travel path probability selection data is input to train the XGBoost algorithm to obtain a highway path selection model based on XGBoost, and the predicted traffic distribution OD matrix is ​​distributed to the road network to realize the distribution of traffic demand.

10. A highway traffic assignment system based on neural network, It is characterized in that include: The data acquisition module is used to obtain the vehicle toll data set and path information data set detected by the district and county ETC gantry; divide the vehicle into three types according to the attributes of the whole sample, cars and trucks, and process different types of district and county traffic distribution OD matrices and OD distribution probability value matrices; A data combination module is used to obtain the population data of the district-level traffic community, the economic development data of the traffic community, and the travel impedance data of the traffic community, and combine them with the OD distribution probability value matrix into a data set, and divide it into training, verification and test sets; The deep learning gravity model building module is used to take the historical traffic distribution OD matrix, traffic community population data, economic development data, and travel impedance data as model inputs, and use the obtained probability values ​​as the model output label set for supervised training and iteration until the model effect is optimal; The deep learning gravity model prediction module is used to test the model in the test set, obtain the latest district-level traffic distribution OD matrix status, as well as traffic community population data, economic development data, travel impedance data, and predict the future traffic community traffic distribution OD matrix; The traffic demand allocation module is used to build a highway path selection model based on the XGBoost algorithm, allocate the predicted traffic distribution OD matrix to the road network, and realize the allocation of traffic demand.

Images