Simulation parameter calibration method and electronic device

Abstract

The application provides a simulation parameter calibration method and electronic equipment. In the method, for a time period of an OD estimation problem, the time period is divided into a plurality of time periods connected at the head and tail, each sub time period includes a plurality of continuous time periods, and there is an overlap of time periods between adjacent two sub time periods. For a road network of the OD estimation problem, the road network is divided into at least two independent sub road networks. The sub time periods and the sub road networks are combined correspondingly to obtain a plurality of OD estimation sub problems. An OD estimation sub matrix of each OD estimation sub problem is calculated. The OD estimation sub matrices are integrated to obtain an OD estimation matrix of the OD estimation problem. The OD estimation matrix of the OD estimation problem is used as OD configuration information in a traffic simulation model. The application can calibrate traffic simulation parameters more accurately.

Description

Technical Field

[0001] This application relates to the field of traffic simulation technology, and in particular to a simulation parameter calibration method and electronic device. Background Technology

[0002] Traffic simulation utilizes computer simulation technology to replicate and extrapolate complex traffic systems, making it an important tool for design and planning departments and traffic management departments to understand and manage complex traffic systems. With the continuous improvement of computer computing power, the individual behaviors and traffic phenomena depicted by traffic simulation tools are becoming increasingly complex and detailed.

[0003] The sophisticated simulation tools inevitably bring a massive amount of simulation parameters. Before conducting traffic simulations, how to accurately calibrate the simulation parameters to make the simulation results more accurate is a problem that needs to be solved. Summary of the Invention

[0004] This application provides a simulation parameter calibration method and electronic device, which can perform more accurate calibration of traffic simulation parameters.

[0005] Firstly, embodiments of this application provide a simulation parameter calibration method, comprising: for the time period of the OD estimation problem, dividing the time period into several consecutive time periods, each sub-time period including several consecutive time periods, with overlap between adjacent sub-time periods; for the road network of the OD estimation problem, dividing the road network into at least two independent sub-road networks; combining the sub-time periods and sub-road networks to obtain multiple OD estimation sub-problems; calculating the OD estimation sub-matrix for each OD estimation sub-problem; integrating the OD estimation sub-matrices to obtain the OD estimation matrix of the OD estimation problem; and using the OD estimation matrix of the OD estimation problem as OD configuration information in the traffic simulation model. In this method, the OD estimation problem is divided into OD estimation sub-problems, and the OD estimation sub-matrices are calculated separately and then integrated to obtain the OD estimation matrix, which is then used as the OD configuration information in the traffic simulation model. In the above method, the OD estimation problem is divided into estimation sub-problems, the OD estimation sub-matrices of the OD estimation sub-problems are calculated separately, and then integrated into the OD estimation matrix, thereby enabling accurate calibration of the OD estimation matrix in the traffic simulation parameters, and reducing the computational load of calculating the OD estimation matrix.

[0006] In one possible implementation, the OD estimation submatrix for each OD estimation subproblem is computed, including:

[0007] For an OD estimation subproblem, an objective function is constructed. Within the feasible region, n sampling points are randomly selected from the OD estimation matrix of the OD estimation subproblem. These n sampling points are then input into the simulation system to obtain the objective function values ​​corresponding to the n sampling points. Based on the n sampling points and their corresponding objective function values, the mean and variance are calculated using Gaussian process regression, serving as a surrogate model for the functional relationship between the objective function and the OD estimation matrix of the OD estimation subproblem. Based on this surrogate model, a data acquisition function is constructed, and the extreme points of this function are solved to serve as new sampling points. These new sampling points are then input into the simulation system. The simulation system obtains the objective function value corresponding to the new sampling point. Based on all sampling points, including the n sampling points and the new sampling point, and the objective function value corresponding to the sampling point, it is determined whether the iteration stopping condition is met. If yes, the OD estimation submatrix of the OD estimation subproblem is determined based on the optimal sampling point among all sampling points. If no, the n sampling points are replaced with all sampling points, and the step of calculating the mean and variance based on the n sampling points and the objective function value corresponding to the n sampling points through Gaussian process regression is returned. Then, n is replaced with n+1, and so on, until the iteration stopping condition is met.

[0008] One possible implementation also includes: calculating the OD prediction matrix using an OD prediction model based on the OD estimation matrix and the OD data of the target road network;

[0009] Configure the OD prediction matrix as the OD prediction information in the traffic simulation model.

[0010] In one possible implementation, the OD prediction model is pre-trained. The training method includes: obtaining historical road network traffic statistics and OD data statistics; constructing an initial model of the OD prediction model; converting the statistical data into tensors that match the initial model; and using the tensors to train and validate the initial model to obtain the OD prediction model.

[0011] In one possible implementation, the method further includes: calculating the micro-parameters of the traffic simulation model and using the calculated micro-parameters as the configuration information of the traffic simulation model; the micro-parameters include: parameters of the kinematic model, and / or, parameters of the path selection model; the kinematic model includes: a car-following model, and / or, a lane-changing model.

[0012] In one possible implementation, the micro-parameters of the traffic simulation model are calculated, and the calculated micro-parameters are used as the configuration information of the traffic simulation model. This includes: extracting car-following behavior data of vehicles from vehicle trajectory data; solving the optimization problem of a preset car-following model based on the car-following behavior data to obtain car-following model parameters that satisfy the optimization problem, wherein the optimization problem includes the car-following model parameters; and using the car-following model parameters as the configuration information of the car-following model in the traffic simulation model.

[0013] In one possible implementation, the optimization objective of the car-following model optimization problem includes:

[0014]

[0015] P * =argmin p WMAPE

[0016] Among them, v i '(t) is the estimated speed of the i-th vehicle at time t, v i (t) is the actual speed of the i-th vehicle at time t, g i '(t) is the estimated distance between the i-th vehicle and the preceding vehicle at time t, g i (t) is the actual distance between the front and rear vehicles of the i-th vehicle at time t.

[0017] In one possible implementation, the optimization problem of a preset car-following model is solved based on car-following behavior data to obtain car-following model parameters that satisfy the optimization problem. This includes: using a heuristic algorithm based on car-following behavior data to solve the optimization problem of a preset car-following model to obtain car-following model parameters that satisfy the optimization problem.

[0018] In one possible implementation, the process of calculating the micro-parameters of the traffic simulation model and using the calculated micro-parameters as configuration information of the traffic simulation model further includes: extracting lane-changing behavior data of vehicles from vehicle trajectory data; solving the optimization problem of a preset lane-changing model based on the lane-changing behavior data to obtain lane-changing parameters that satisfy the optimization problem of the lane-changing model, wherein the optimization problem of the lane-changing model includes the lane-changing parameters of the lane-changing model; and using the lane-changing model parameters as configuration information of the lane-changing model in the traffic simulation model.

[0019] In one possible implementation, the optimization objective of the lane-changing model optimization problem includes:

[0020]

[0021] Where, N lc’ N is the estimated number of lane changes for the vehicle. lc This refers to the actual number of lane changes made by the vehicle.

[0022] In one possible implementation, the optimization problem of a preset lane-changing model is solved based on lane-changing behavior data to obtain lane-changing parameters that satisfy the optimization problem of the lane-changing model. This includes: using a method of traversing from small to large within the feasible range of lane-changing times to obtain lane-changing parameters that satisfy the optimization problem of the lane-changing model.

[0023] In one possible implementation, the process of calculating the micro-parameters of the traffic simulation model and using the calculated micro-parameters as configuration information for the traffic simulation model further includes: extracting vehicle travel paths from the target road network to obtain the feasible path set for each pair of ODs in the target road network; acquiring segment travel time and path flow data of the target road network; calculating the probability of each vehicle choosing each path in the feasible path set based on the segment travel time and path flow data using a preset path selection model; constructing a likelihood function; solving for the parameters in the path selection model with the objective of maximizing the likelihood function; and using the solved parameters as configuration parameters for the path selection model in the traffic simulation model.

[0024] In one possible implementation, the process of calculating the micro-parameters of the traffic simulation model and using these calculated micro-parameters as configuration information for the traffic simulation model includes: acquiring the traffic flow information, segment length, free-flow speed, and capacity information of the target road network; for the parameters given in the preset path selection model, performing random user equilibrium allocation based on the path selection model according to the traffic flow information, segment length, free-flow speed, and capacity information, and recording the road network traffic flow when equilibrium is reached; calculating the relative error between the road network traffic flow at equilibrium and the actual road network traffic flow; solving for the parameters of the path selection model with the objective of minimizing the relative error; and using the solved parameters as configuration parameters of the path selection model in the traffic simulation model.

[0025] In one possible implementation, the method further includes: calculating macroscopic parameters of the traffic simulation model and using the calculated macroscopic parameters as configuration information of the traffic simulation model; the macroscopic parameters include: driver speed distribution parameters, and / or, free-flow vehicle speed.

[0026] In one possible implementation, the macroscopic parameters of the traffic simulation model are calculated, and the calculated macroscopic parameters are used as the configuration information of the traffic simulation model. This includes: calculating the free-flow velocity of the target road network using a supervised learning model based on the free-flow velocity observation data of the target road network; the supervised learning model is pre-trained; and the free-flow velocity is used as the road network speed limit in the traffic simulation model.

[0027] In one possible implementation, the supervised learning model is pre-trained, and the training method includes: acquiring training samples; the training samples include observation data of free-flow velocity; using the training samples to train the pre-set model to obtain the supervised learning model, which is used to calculate the free-flow velocity of the target road network.

[0028] One possible implementation further includes: acquiring test samples, which include observational data of free-flow velocity, and the test samples do not overlap with the training samples; using the test samples to test the prediction accuracy of the supervised learning model for free-flow velocity; and adjusting the model parameters of the supervised learning model based on the test results until the supervised learning model reaches a preset prediction accuracy.

[0029] In one possible implementation, the calculation of macroscopic parameters of the traffic simulation model and the use of the calculated macroscopic parameters as configuration information of the traffic simulation model also includes: fitting the speed distribution of the driver group based on the driving behavior data of the target road network to obtain the speed distribution parameters of the driver group; and using the obtained speed distribution parameters as the speed distribution parameters of the driver group in the traffic simulation model.

[0030] In one possible implementation, calculating the macroscopic parameters of the traffic simulation model further includes: establishing an objective function; obtaining feasible solutions for n sets of driver speed distributions as n sampling points, inputting the n sampling points into the simulation system to obtain the objective function values ​​corresponding to the n sampling points; establishing a surrogate model between the objective function and the speed distribution through Gaussian process regression based on the n sampling points and the objective function values ​​corresponding to the n sampling points; constructing a sampling function based on the surrogate model, solving for the extreme points of the sampling function, and using them as new sampling points; inputting the new sampling points into the simulation system to obtain the objective function values ​​corresponding to the new sampling points; determining whether the objective function values ​​have reached the convergence condition; if so, determining the speed distribution parameters of the driver group based on the optimal sampling point among all sampling points; if not, returning to the step of obtaining feasible solutions for n sets of driver speed distributions, and replacing n with n+1, and so on, until the iteration stopping condition is met.

[0031] One possible implementation includes: acquiring predicted and real-time weather information for the actual simulation period; determining the weather-affected area in the target road network based on the predicted and real-time weather information; extracting road network-related data and the current free-flow speed within the weather-affected area; inputting the road network-related data and free-flow speed into a supervised learning model of free-flow vehicle speed to obtain a correction value for the free-flow speed under the predicted weather information; and correcting the free-flow speed in the traffic simulation model based on the correction value.

[0032] One possible implementation also includes: performing a simulation based on a traffic simulation model to obtain traffic simulation results; determining whether the traffic simulation results have reached the preset simulation accuracy; if so, ending the simulation parameter calibration; otherwise, returning to the step of calculating the traffic demand parameters of the traffic simulation model until the traffic simulation results reach the preset simulation accuracy.

[0033] In a second aspect, embodiments of this application provide an electronic device, including: a processor; a memory; and a computer program, wherein the computer program is stored in the memory, and when the computer program is executed by the electronic device, it causes the electronic device to perform the method of any one of the first aspects.

[0034] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the method of any one of the first aspects. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of one embodiment of the simulation parameter calibration method of this application;

[0037] Figure 2A This is a flowchart of another embodiment of the simulation parameter calibration method of this application;

[0038] Figure 2B This is a flowchart of yet another embodiment of the simulation parameter calibration method of this application;

[0039] Figure 3 This is a flowchart illustrating the implementation of step 201 in the simulation parameter calibration method of this application.

[0040] Figure 4A This is a flowchart illustrating the differential evolution algorithm of this application;

[0041] Figure 4B This application provides a schematic diagram of lane change error variation.

[0042] Figure 5 This is a flowchart illustrating the implementation of step 205 in the simulation parameter calibration method of this application.

[0043] Figure 6A The road network diagram provided for this application;

[0044] Figure 6B A time segmentation diagram provided for this application;

[0045] Figure 6C A schematic diagram of the regional division of the sub-road network provided in this application;

[0046] Figure 6D A schematic diagram of the OD estimation submatrix integration method provided in this application;

[0047] Figure 7A Another road network diagram provided for this application;

[0048] Figure 7B This is a flowchart of one implementation of step 202 in the simulation parameter calibration method of this application;

[0049] Figure 7C The likelihood function variation graph provided for this application;

[0050] Figure 8A This is a schematic diagram of another implementation process for step 202 in the simulation parameter calibration method of this application;

[0051] Figure 8B This is a diagram illustrating the comparison between the road network traffic and the actual traffic in this application;

[0052] Figure 9 This is a schematic diagram of one implementation process of step 203 in the simulation parameter calibration method of this application;

[0053] Figure 10A This is a schematic diagram of one implementation process of step 204 in the simulation parameter calibration method of this application;

[0054] Figure 10B A schematic diagram of the final calibrated velocity distribution provided for this application;

[0055] Figure 11 This is a schematic diagram of one implementation process of step 207 in the simulation parameter calibration method of this application;

[0056] Figure 12 This is a schematic diagram of one implementation process of step 208 in the simulation parameter calibration method of this application. Detailed Implementation

[0057] The terminology used in the implementation section of this application is for the purpose of explaining specific embodiments of this application only, and is not intended to limit this application.

[0058] The proliferation of sophisticated simulation tools inevitably leads to a massive number of simulation parameters. To ensure that these tools can accurately and precisely depict complex traffic systems and individual behaviors, these parameters need to be set according to the specific simulation environment and actual observation data. This is particularly true when dealing with real-time traffic simulations, where rapid and accurate calibration of a large number of simulation parameters is crucial. Against this backdrop, numerous simulation parameter calibration techniques have been proposed. However, current engineering practice shows that existing techniques still struggle to accurately calibrate the numerous and complex macroscopic and microscopic parameters within simulation tools in real time, and they also cannot accurately depict the impact of real-world environmental factors such as road conditions and weather on the simulation.

[0059] In one example of a simulation parameter calibration method, real-time traffic data is acquired and multi-source data fusion processing is performed to determine real-time traffic signal data and real-time traffic state data. The real-time traffic signal data and real-time traffic state data are then input into a traffic simulation model to determine the simulation results. When the simulation results do not meet the preset accuracy requirements, the traffic simulation model is checked and updated. This includes adjusting key parameters based on the deviation between the default values ​​of the traffic simulation model and the real-time calculated values ​​of the parameters, or the deviation between the default values ​​of the parameters and the average value of historical accumulated data. The values ​​of OD and path flow distribution and prediction model parameters in the traffic simulation model are also adjusted. Finally, the simulation is re-performed based on the checked and updated traffic simulation model.

[0060] In another simulation parameter calibration method, parameter calibration is performed in the simulation software. A simulated road network is built using paramics based on actual road conditions. Measured travel time and flow rate, along with their respective ranges, are input as the basis for calibration. The simulation software automatically calls the program, calculates the initial evaluation simulation count, and performs the simulation to determine if parameter calibration is needed. If calibration is required, the program performs sensitivity analysis on the simulation model parameters, selects parameters with significant impact for calibration, uses a genetic simulated annealing algorithm to search for the optimal parameter combination, automatically calls the Modeller module to run the simulation, collects statistical data after simulation, and compares it with measured data. If the error does not meet the requirements, the parameters are automatically adjusted, and the simulation is repeated until the error meets the requirements, at which point the optimized parameter values ​​are output.

[0061] However, the above-mentioned technologies mainly calibrate macroscopic parameters and cannot take into account the influence of environmental factors such as weather and road conditions and characterize them in the simulation, resulting in relatively low accuracy of simulation parameter calibration.

[0062] Therefore, this application proposes a simulation parameter calibration method that can more accurately calibrate traffic simulation parameters.

[0063] Figure 1 A flowchart of one embodiment of the simulation parameter calibration method of this application is shown below. Figure 1 As shown, the method may include:

[0064] Step 101: Calculate the micro parameters of the traffic simulation model based on traffic observation data, and use the calculated micro parameters as the configuration information of the traffic simulation model; proceed to step 103.

[0065] Traffic observation data may include, but is not limited to: vehicle trajectory data, and / or, road network traffic flow, etc.

[0066] The traffic simulation model may include, but is not limited to, a kinematic model and / or a path selection model.

[0067] Optionally, kinematic model parameters can be calculated based on vehicle trajectory data in this step.

[0068] The vehicle trajectory data in this step is obtained by sampling the motion process of one or more vehicles in a spatiotemporal environment according to a preset sampling time interval. This sampled data constitutes the trajectory data in chronological order of sampling time. Each sampled data point may include sampling time, vehicle ID, latitude and longitude, speed, and other information. The sampling time records the sampling time of this sample, the vehicle ID identifies the vehicle, the latitude and longitude records the vehicle's position, and the speed records the vehicle's current speed.

[0069] The vehicle trajectory data in this step can be the trajectory data of vehicles in the first road network in the first time period; the first time period can be the target time period, a part of the target time period, or other time periods outside the target time period; the first road network can be the target road network, a part of the target road network, or other road network areas outside the target road network, which is not limited in this application embodiment.

[0070] The target time period and target road network mentioned above refer to the time period and road network for which the simulation system is expected to perform traffic simulation.

[0071] Kinematic model parameters may include: car-following model parameters and / or lane-changing model parameters.

[0072] Calculating kinematic model parameters based on vehicle trajectory data can specifically include: calculating the model parameters of the car-following model and / or lane-changing model based on the vehicle trajectory data.

[0073] Alternatively, the following model may include, but is not limited to, the IDM model or the Krass model.

[0074] Optionally, in this step, the path selection model parameters can be calculated based on traffic observation data such as traffic flow and vehicle trajectories of the third road network during the third time period. Traffic observation data may include, but is not limited to, vehicle path information and vehicle travel time within the road network.

[0075] Optionally, the travel time of a vehicle in the road network can be calculated by inversely estimating the vehicle's speed in the road network. Therefore, the aforementioned travel time of a vehicle in the road network can be replaced by the vehicle's speed.

[0076] The third time period and the third road network may or may not be related to the first time period and the first road network. Preferably, the third time period includes the first time period and the third road network includes the first road network, which can improve the simulation results of subsequent playback simulations.

[0077] Optionally, the third time period and the third road network in this step can be the same as the second time period and the second road network in step 103. In this case, the simulation results of subsequent playback simulation can be improved.

[0078] Step 102: Calculate the macroscopic parameters of the traffic simulation model based on historical traffic observation data, and use the calculated macroscopic parameters as the configuration information of the traffic simulation model.

[0079] Alternatively, macroscopic parameters can also be calculated based on weather data.

[0080] Macroscopic parameters may include, but are not limited to: free-flow vehicle speed, and / or, speed distribution parameters of the driver group. Among them, free-flow vehicle speed can be used to configure the road network speed limit in traffic simulation.

[0081] Optionally, this step may include:

[0082] The free-flow vehicle speed of the target road network is calculated using a supervised learning model, and this free-flow vehicle speed is set as the vehicle speed limit of the target road network.

[0083] Optionally, feature learning can be performed on the free-flow vehicle speed to obtain a supervised learning model of the free-flow vehicle speed. In this step, feature learning can be performed on the free-flow vehicle speed based on historical data such as vehicle speed, weather information, and / or road condition information of the fourth road network in the fourth time period to obtain a supervised learning model of the free-flow vehicle speed.

[0084] The input to the supervised learning model is the vehicle speed, weather information, and road condition information of the target road network, and the output is the free-flow vehicle speed of the target road network. Optionally, when performing traffic simulation on the target road network, the vehicle speed, weather information, and / or road condition information of the target road network can be input into the supervised learning model to obtain the free-flow vehicle speed of the target road network, and the free-flow vehicle speed can be set as the vehicle speed limit of the target road network.

[0085] In this step, the fourth road network may or may not include the first and third road networks mentioned above, but it is preferable to include them. Generally, the range of the fourth road network is larger than that of the first and third road networks mentioned above.

[0086] Optionally, this step may include:

[0087] The speed distribution parameters of the driver group are calculated using the driving behavior data of the drivers, and the calculated speed distribution parameters are used as the speed distribution parameters of the driver group in the traffic simulation model.

[0088] The driver behavior data used in this step can be driver behavior data from the target road network or driver behavior data from other road networks. In other words, the data source used for calibration in this step is not limited to driver behavior data observed from drivers in the target road network.

[0089] The road network to which the driving behavior data used in this step belongs may or may not be related to the first to fourth road networks mentioned above, and this application embodiment does not limit this.

[0090] Step 103: Calculate the traffic demand parameters of the traffic simulation model and use the calculated traffic demand parameters as the configuration information of the traffic simulation model.

[0091] Traffic demand parameters may include, but are not limited to: OD estimation matrix, and / or OD prediction matrix.

[0092] Optionally, offline OD back-calculation can be performed in this step to obtain the OD estimation matrix.

[0093] The OD estimation matrix records the flow rates between ODs.

[0094] In this step, offline OD back-calculation can be performed based on the historical traffic data of the second road network in the second time period to obtain the OD estimation matrix.

[0095] The second time period and the second road network can be related to or unrelated to the first time period and the first road network in step 101. Preferably, the second time period includes the first time period and the second road network includes the first road network, which can improve the simulation results of subsequent playback simulations.

[0096] Optionally, in this step, the kinematic model parameters calibrated in step 101 can be input into the simulation tool, and traffic demand parameters can be calculated using a Bayesian optimization algorithm that considers time correlation, based on the observation data.

[0097] Optionally, when performing online simulation, this step may include: calculating the OD prediction matrix. The OD prediction matrix records the predicted flow rates between ODs.

[0098] Optionally, calculating the OD prediction matrix may specifically include:

[0099] The OD prediction model is preset, and the OD estimation matrix and the OD data of the target road network are input into the OD prediction model to obtain the OD prediction matrix.

[0100] The method for establishing the OD prediction model is described in subsequent embodiments and will not be repeated here.

[0101] Step 104: Perform a simulation based on the traffic simulation model to obtain traffic simulation results.

[0102] The implementation of this step can be referenced from relevant simulation techniques, which will not be elaborated here.

[0103] Step 105: Determine whether the traffic simulation results have reached the preset simulation accuracy. If yes, the simulation parameter calibration ends; otherwise, return to step 103 and recalculate the traffic demand parameters.

[0104] In this step, to determine whether the traffic simulation results have reached the preset simulation accuracy, you can refer to relevant simulation technologies, which will not be elaborated here.

[0105] It should be noted that when recalculating traffic demand parameters in step 105, some or all of the traffic demand parameters can be calculated to optimize the traffic demand parameters and thus optimize the traffic simulation model. This application embodiment does not limit this. For example, offline OD back-calculation can be performed again, or, when there is a lack of measured driver group speed distribution statistics, the driver group speed distribution parameters can be recalculated, or, when there is a lack of road segment free-flow data statistics, the free-flow vehicle speed can also be recalculated, thereby adjusting the road speed limits in the target road network.

[0106] See Figure 2A The flowchart of another embodiment of the simulation parameter calibration method of this application is shown, as follows: Figure 2A As shown, it includes:

[0107] Step 201: Calculate the kinematic model parameters and use the calculated kinematic parameters to configure the traffic simulation model.

[0108] Step 202: Calculate the path selection model parameters and configure the traffic simulation model using the calculated path selection model parameters.

[0109] Step 203: Calculate the free-flow vehicle speed of the target road network using a preset supervised learning model, and set the free-flow vehicle speed as the vehicle speed limit of the target road network in the traffic simulation model.

[0110] Step 204: Calculate the speed distribution parameters of the driver group and use the calculated speed distribution parameters to configure the traffic simulation model.

[0111] Step 205: Perform offline OD back-calculation to obtain the OD estimation matrix, and use the OD estimation matrix to configure the traffic simulation model.

[0112] Step 206: Generate a replay simulation based on the configured traffic simulation model.

[0113] After step 205 is completed, a simulation environment with calibrated parameters is obtained, and the simulation system can generate a playback simulation of the target road network.

[0114] Figure 2BThis is a flowchart of another embodiment of the simulation parameter calibration method of this application, as shown below. Figure 2B As shown, the method may include: replacing step 206 with the following steps 207 to 209, specifically,

[0115] Step 207: Generate an OD prediction matrix based on the historical road network traffic and OD estimation matrix of the target road network, and use the OD prediction matrix to configure the traffic simulation model.

[0116] After step 205 is completed, a simulation environment with calibrated parameters is obtained, and the simulation system enters the online operation stage. However, since the OD is time-varying during the simulation process, it is necessary to predict the OD in real time online in order to improve the prediction and inference accuracy of the simulation system.

[0117] Step 208: Calculate the free-flow vehicle speed of the target road network based on the weather forecast information for the simulation period, and adjust the road network speed limit in the traffic simulation model according to the free-flow vehicle speed.

[0118] Step 209: Based on the traffic simulation model calibrated in the previous steps, generate the predicted simulation results of the target road network.

[0119] Figure 2B The method shown can achieve predictive simulation of the target road network.

[0120] The following sections will discuss the above. Figures 1-2B The specific implementation of the steps is illustrated by example.

[0121] pass Figure 3 The implementation of calculating the kinematic model parameters in step 201 is illustrated by an example. This example uses a kinematic model that includes both a car-following model and a lane-changing model. Figure 3 As shown, calculating the kinematic model parameters can include:

[0122] Step 301: Obtain trajectory data and extract traffic micro-behavioral features from the trajectory data.

[0123] Trajectory data can be high-precision trajectory data obtained through methods such as high-point cameras, drone aerial photography, and radar-visual fusion.

[0124] Traffic micro-behavioral characteristics may include, but are not limited to: each vehicle's following behavior, lane-changing behavior, and the distance between the vehicle and the vehicle in front of it.

[0125] For example, in this embodiment, the German HighD dataset (2018), containing the driving trajectories of 1047 vehicles, can be selected as the data source. The HighD dataset (Highway Drone Dataset) is a new dataset of natural vehicle trajectories recorded on German highways. By using a camera-equipped drone, the position and movement of each vehicle are measured from a bird's-eye view. This approach allows for unobstructed capture of vehicle longitudinal and lateral dimensions at high resolution. Vehicle height information, which cannot be directly obtained from a top-down view, can be predicted according to vehicle type, overcoming typical limitations of established traffic data collection methods, such as occlusion, from an aerial perspective.

[0126] The dataset contains trajectory data from 110,000 vehicles across 6 locations, spanning 11.5 hours, with a total measured mileage of 45,000 km. It also includes 5,600 complete lane change records. The dataset is evaluated based on its quantity, variety, and included context. Using state-of-the-art computer vision algorithms, the localization error is typically less than ten centimeters.

[0127] Table 1 shows the descriptions of the main fields of the HighD data.

[0128]

[0129] Table 1

[0130] The extracted traffic micro-behavioral features may include: vehicle ID (i.e., the vehicle ID of this vehicle in the following text), vehicle speed, distance between vehicles (distance between this vehicle and the vehicle in front), speed of the vehicle in front, vehicle ID of the vehicle in front, vehicle ID of the vehicle in front to the left, vehicle ID of the vehicle in front to the right, lane number, vehicle type, etc.

[0131] For example, the data is categorized according to vehicle ID, and the relevant data for each vehicle is stored in a dictionary. For each vehicle, the process with the longest following time is extracted, and vehicles with a following time of less than 4 seconds are removed. From this, 72 'car' and 72 'truck' vehicles are selected for the experiment, with approximately 40,000 data points. Since the driving behaviors of 'car' and 'truck' are very different, different parameter search ranges are set for the two types of vehicles.

[0132] Step 302: Select the car-following model to be calibrated and establish an optimization problem for the parameters of the car-following model.

[0133] In one possible implementation, the car-following model can be an Intelligent Driver Model (IDM). It should be noted that the IDM model is merely an example; in practical applications, other models, such as the Krass model, can also be used for car-following. This application does not limit the specific implementation. The formula for the IDM model is shown below.

[0134]

[0135]

[0136] The first equation is the main expression of the IDM model, and the second equation is the driver's desired distance s in the IDM model. i * The expression for (t). Where A is the maximum acceleration, δ is the acceleration exponent (δ>0), T is the safety time interval, b is the comfortable deceleration, s0 is the safe distance from rest, and v f For the desired speed, all six parameters are to be calibrated. Among them, v i (t) represents the speed of the i-th vehicle at time t, and a i (t) represents the acceleration of the i-th vehicle at time t.

[0137] Considering that acceleration can easily generate large cumulative errors, and that there is no upper limit to vehicle position, this application embodiment selects to consider only speed and the distance between vehicles in front and behind, and constructs a comprehensive evaluation index, the Weighted Mean Absolute Percentage Error (WMAPE).

[0138]

[0139] Among them, v i '(t) is the estimated speed of the i-th vehicle at time t, v i (t) is the actual speed of the i-th vehicle at time t, g i '(t) is the estimated distance between the i-th vehicle and the preceding vehicle at time t, g i (t) represents the actual distance between the i-th vehicle and the vehicle in front at time t. The actual speed corresponds to the vehicle speed in the above traffic micro-behavioral characteristics, and the actual distance between the i-th vehicle and the vehicle in front corresponds to the distance between the i-th vehicle and the vehicle in front.

[0140] Therefore, the car-following model parameter calibration problem can be transformed into a WMAPE minimization problem:

[0141] P*=argmin p WMAPE

[0142] Step 303: Based on the micro-traffic behavior characteristics, use a heuristic algorithm to calculate the above optimization problem and obtain the optimal solution of the car-following model parameters, which will be used as the parameter configuration information of the car-following model in the simulation model.

[0143] Optionally, the above-mentioned heuristic algorithms may include, but are not limited to, differential evolution algorithms, genetic algorithms, etc.

[0144] This application uses the Differential Evolutionary Algorithm (DEA) as an example. DEA is a heuristic random search algorithm based on population differences. It represents the solution to a problem as a survival-of-the-fittest process of "chromosomes." Through generations of evolution of the "chromosome" population, including operations such as replication, crossover, and mutation, it eventually converges to the individual "most adapted to the environment," thus obtaining the optimal or satisfactory solution to the problem. This algorithm was proposed by R. Storm and K. Price for solving Chebyshev polynomials and is a meta-heuristic algorithm for optimization problems. Essentially, it is a greedy genetic algorithm based on real-number encoding with a superiority-preserving principle.

[0145] The characteristics of the differential evolution algorithm are:

[0146] Genetic operations are mainly performed using the differential mutation operator, which has a simple structure and is easy to use.

[0147] It boasts superior performance, exhibiting good reliability, robustness, and high efficiency.

[0148] The differential mutation operator can be a constant or an adaptive operator with variable time length and search direction, thus exhibiting strong adaptability.

[0149] It possesses inherent parallelism and can perform collaborative search; under the same requirements, the differential evolution algorithm has a faster convergence speed.

[0150] Differential evolution is similar to genetic algorithm, including mutation, crossover, and elimination mechanisms. However, the difference between differential evolution and genetic algorithm is that the mutation part is the difference between the variables of two selected solution members, which is then added to the variables of the current solution member after scaling. Therefore, differential evolution does not need to use probability distribution to generate the next generation of solution members.

[0151] like Figure 4A As shown, the differential evolution algorithm can include processes such as population initialization, mutation, crossover, and selection. Detailed explanations are as follows:

[0152] The optimization problem that needs to be solved for minimizing WMAPE as described above can be identified as follows:

[0153] min f(x1,x2,...x D )

[0154]

[0155] Where x1, x2, ..., x D These are the parameters to be adjusted: L is the lower bound, U is the upper bound, and x is the lower bound. j This refers to the range of values ​​for the j-th parameter, where j can take values ​​from 1, 2, ..., D.

[0156] Continuing with the previous example, since the car-following model requires calibration of 6 parameters, the value of D is 6. The range of values ​​for each parameter is not limited in this embodiment and can be set independently in practical applications.

[0157] 1) Initialize the population

[0158] The classic differential evolution algorithm uses real number encoding, which makes it more suitable for solving real number optimization problems.

[0159]

[0160] Where X i (0) represents the i-th individual, NP represents the population size, and D represents the problem dimension.

[0161] Randomly generate solution vectors between the upper and lower bounds:

[0162]

[0163] 2) Variation

[0164] Individual mutation can be achieved through differential mutation strategies. A common differential mutation strategy is to randomly select two different individuals in the population, scale their vector difference, and then synthesize the vector with the individual to be mutated.

[0165] V i (g+1)=X r1 (g)+F(X r2 (g)-X r3 (g))

[0166] Where r1, r2, and r3 are three random numbers in the interval [1, NP]. F is the scaling factor, also known as the mutation rate, and g represents the g-th generation.

[0167] Generally, mutation operators are often taken as constants, which are difficult to determine accurately. A high mutation rate leads to a low global optimum, while a low mutation rate reduces population diversity and can easily cause premature convergence. Therefore, the adaptive mutation operator in this embodiment can be:

[0168]

[0169] Thus, the mutation operator is initially 2F0, which maintains diversity and prevents premature convergence in the early stages. As the process progresses, the mutation operator decreases until it eventually becomes F0, preventing the optimal solution from being compromised.

[0170] The most important element in the differential evolution algorithm is the differential mutation strategy, from which the algorithm derives its name. Price et al. also proposed nine mutation strategies, as shown below:

[0171] V i (g+1)=X best (g)+F(X r2 (g)-X r3 (g))

[0172] V i (g+1)=X r1 (g)+F(X r2 (g)-X r3 (g))+F(X r4 (g)-X r5 (g))

[0173] V i (g+1)=X r1 (g)+F(X r2 (g)-X r3 (g))+F(X best (g)-X r1 (g))

[0174] Where X best (g) represents the best individual in the current group.

[0175] 3) Cross

[0176] The purpose of crossover is to randomly select individuals, as differential evolution is also a stochastic algorithm. Differential evolution uses discrete crossover factors, including binomial crossover and exponential crossover. The crossover operator generates a mutation vector V through the mutation operator. i With parent individual vector X i Discrete hybridization is performed to obtain the trial vector U. i The method for crossover operations is:

[0177]

[0178] Here, CR is called the crossover probability, with an interval of [0,1], and new individuals are generated based on the probability.

[0179] 4) Select

[0180] After generating a child population through mutation and crossover operators, the differential evolution algorithm employs a greedy selection strategy, choosing the better individual as the new individual. A one-to-one competitive selection operator compares the child individual with its corresponding parent individual, and the better one is retained in the next generation. For the minimum optimization problem, its selection operator can be described as:

[0181]

[0182] Since the differential evolution algorithm uses a one-to-one competitive selection, it is a steady-state evolution algorithm that preserves elite individuals.

[0183] Once a new population is formed, the differential evolution algorithm continues to evolve the population through mutation, crossover, and selection operators until a termination condition is reached and the program exits. Based on the above differential evolution algorithm, the algorithm flow for solving the optimization problem of the car-following model in this application embodiment is as follows:

[0184]

[0185] In other words, this step may include:

[0186] An initial population is generated, and the value of each individual in the initial population is within a preset range.

[0187] WMAPE was calculated using the initial population and traffic micro-behavioral characteristics.

[0188] If WMAPE meets the convergence accuracy, the initial population will be used as a parameter in the race-following model.

[0189] If WMAPE does not meet the convergence accuracy, the initial population is mutated and crossovered to generate a new population.

[0190] WMAPE was calculated using new populations and traffic microbehavioral characteristics.

[0191] If WMAPE meets the convergence accuracy, the new population will be used as the parameters of the carousel model.

[0192] If WMAPE does not meet the convergence accuracy, the new population is mutated and crossovered to generate a new population; this process is repeated until the number of iterations reaches the preset iteration threshold or the preset convergence accuracy is achieved, thereby obtaining the parameters of the carousel model.

[0193] The calculation of WMAPE using the initial population and traffic micro-behavioral characteristics can include:

[0194] Substitute the initial population and traffic micro-behavioral characteristics into the above car-following model to calculate the estimated speed of the i-th vehicle at time t and the estimated distance between the front and rear vehicles.

[0195] WMAPE is calculated based on the actual speed, estimated speed, actual distance between vehicles in front and behind, and estimated distance between vehicles in front and behind at time t.

[0196] For example, in the above steps, the population size NP can be 15, the iteration threshold can be 100, and the convergence accuracy can be ≤6.5%.

[0197] The above steps describe the parameter calculation method for the car-following model. The following section explains the implementation of parameter calibration for the lane-changing model.

[0198] Step 304: Select the lane-changing model to be calibrated, and establish an optimization problem for the lane-changing model based on its parameters.

[0199] The sensitivity of each parameter of the LC2013 model was analyzed using the Sobal global sensitivity analysis method, and the key parameter (lcSpeedGain) affecting lane changing was identified, thereby reducing the complexity of the optimization problem.

[0200] Lane change frequency was selected as the evaluation metric.

[0201]

[0202] Where, N lc’ N is the estimated number of lane changes for the vehicle. lc This refers to the actual number of lane changes made by the vehicle.

[0203] The optimization problem is to find the optimal lane-changing parameters that minimize the difference between the actual number of lane changes and the estimated number of lane changes.

[0204] Step 305: Calculate the optimization problem of the lane-changing model based on the micro-traffic behavior characteristics to obtain the optimal solution of the lane-changing model parameters, which serves as the parameter configuration information for the lane-changing model.

[0205] When there is only one key parameter, a traversal approach can be used, gradually increasing the parameter from small to large within its feasible value range until the optimization objective is reached. If there are multiple key parameters, other heuristic algorithms such as differential evolution can also be used to solve the optimization problem. The setting of the feasible value range of the parameter can be set independently in practical applications, and this application embodiment does not limit it.

[0206] For example, 54 left lane change data points can be extracted from the HighD dataset. The lcSpeedGain parameter is incremented from 0.001, increasing by 0.01 in each iteration. During the entire car-keeping process, if the speedGainProbability parameter reaches the lane change threshold (ProbabilityThreshold), the lane change is considered successful. Different lcSpeedGain values ​​correspond to different numbers of lane changes. The final test results are as follows: Figure 4B The graph shows the lane-changing error variation, used to record the calibration error. The iteration number in the graph represents the number of iterations, corresponding to the lcSpeedGain parameter, and Num_error represents the error value, corresponding to the aforementioned evaluation metric ε. As the lcSpeedGain parameter increases, the rate of decrease in the relative error of the number of lane changes slows down. In actual SUMO simulations, this parameter should not be set too high. Therefore, in one possible implementation, the convergence accuracy can be set to ε = 10%. The lcSpeedGain parameter represents the vehicle's intention to change lanes to gain a faster speed; generally, a larger value indicates a stronger intention. The speedGainProbability parameter measures the likelihood that the vehicle will change lanes to increase speed.

[0207] There is no restriction on the execution order between steps 302-303 and steps 304-305.

[0208] The parameters for the car-following model and the lane-changing model can be obtained through the above calculations, and these parameters can be stored in the simulation configuration.

[0209] pass Figure 5 The implementation of offline OD back-reasoning in step 205 of this application is illustrated by example. Figure 5 As shown, step 205 may include:

[0210] Step 501: Extract basic information such as traffic flow and traffic micro-behavioral features from the trajectory data, and input the traffic micro-behavioral features into the simulation tool to initialize the simulation tool.

[0211] For example: Figure 6A Taking the road network shown as an example, a section of a two-way expressway network about 4km long is selected. The road network includes one toll station and one entrance / exit. The test road network expressway has a total of 3 traffic zones (i.e., the cross sections at both ends of the road network and the entrance / exit zones), with a total of 6 OD pairs.

[0212] Assuming that OD is adjusted every 15 minutes (min) to correct for traffic demand over a 2-hour period (h), the dynamic OD estimation problem has 6*8=48 dimensions. The study road network is equipped with 5 road detectors, which collect traffic flow data every 5 minutes.

[0213] In this step, basic information such as traffic flow is extracted from the observation data, and the micro-traffic behavior characteristics are input into the simulation tool to initialize the tool. An example of the detector's actual observation data is shown in Table 2 below:

[0214]

[0215]

[0216] Table 2

[0217] Step 502: Divide the OD estimation problem into multiple sub-problems according to time and / or spatial characteristics.

[0218] This step may include: dividing the time period into 'a' sub-time periods and the road network into 'b' sub-road networks, resulting in 'ab' OD estimation sub-problems; where two adjacent sub-time periods overlap in time. 'a' and 'b' are natural numbers, and at least one of 'a' and 'b' is greater than 1.

[0219] Optionally, a time segmentation strategy such as Figure 6B As shown, the time period can be divided into 3a-1 time periods, with each 3 consecutive time periods as a sub-time period, and there is an overlap of 1 time period between two adjacent sub-time periods. That is, the last time period of the previous sub-time period overlaps with the first time period of the next sub-time period. Thus, each sub-problem estimates the OD of three time periods, and there is an overlap of 1 time period between every two sub-problems.

[0220] Optionally, in one spatial segmentation method, the second road network can be divided into several sub-road networks, and the location where the road network is divided is set as a virtual traffic zone. The virtual traffic zone belongs to two sub-road networks at the same time.

[0221] It should be noted that the location where the road network is split (i.e., the location where virtual traffic zones are set) should generally avoid ordinary traffic zones to avoid ambiguity; the location where the road network is split (i.e., the location where virtual traffic zones are set) can preferably be a location with a flow detector, so that the flow detector can be used as a constraint condition when solving the OD estimation matrix of the two sub-road networks at the same time, ensuring that the arrival flow and departure flow of the virtual traffic zone are close to consistent.

[0222] It should be noted that the arrival and departure flows calculated at the virtual traffic cell for two adjacent sub-road networks may not be exactly the same in each sub-time period (e.g., 15 minutes), but the arrival flow accumulated in the virtual traffic cell over the entire simulation duration (e.g., 2 hours or more) should generally be approximately equal to the departure flow.

[0223] For example, such as Figure 6C The simulated road network shown originally had 7 traffic zones, labeled 1 to 7, including three road network sections and four highway entrance / exit toll stations. For OD pairs, the spatial complexity increases significantly; therefore, the road network can be spatially decomposed. Figure 6C In one example of a splitting method, the road network is divided into two sub-road networks. A virtual traffic zone is added at the cut point, namely traffic zone 8 shown by the dashed line. The sub-road network shown by the rectangle below has four traffic zones. For the OD pair, the sub-network shown by the rectangle above has 5 traffic zones. For OD.

[0224] The sub-time periods obtained by time segmentation are combined with the sub-road networks obtained by spatial segmentation to obtain the OD estimation sub-problem composed of the sub-time periods and the sub-road networks.

[0225] Step 503: Apply the Bayesian optimization algorithm to solve each OD estimation subproblem to obtain the OD estimation submatrix for each OD estimation subproblem.

[0226] To distinguish it from the OD estimation matrix of the OD estimation problem, in this embodiment, the OD matrix obtained by solving each OD estimation subproblem is called the OD estimation submatrix. The elements in the OD estimation submatrix are the flows of OD pairs.

[0227] The methods for solving an OD subproblem using Bayesian optimization algorithms in this step can include:

[0228] Step S1: Construct the objective function and constraints.

[0229] Constraints can include: boundary constraints, time-dependent constraints, total constraints, etc.

[0230] The objective function can be the mean absolute percentage error (MAPE) between the flow rate output by the detector in the simulation and the flow rate measured by the actual detector, where K is the number of detectors and T is the number of time periods detected by the detectors. and These represent the output flow rates of the actual and simulated detectors at the k-th detector during the t-th detection period, respectively.

[0231]

[0232] Set boundary constraints:

[0233]

[0234] Setting time-related constraints: Time-related constraints refer to constraining the deviation of traffic demand between adjacent time periods of the same set of origin and destination (OD).

[0235]

[0236] Step S2: Randomly select n feasible solutions of OD estimation submatrices within the feasible region Ω as n sampling points Q = {Q1, ..., Q...} n Input the data into the simulation system to obtain the objective function values ​​y = {f1, ..., fn} corresponding to n sampling points. n}

[0237] Where Q represents the set of sampling points, Q1, ..., Q2. n Let Q represent n sampling points, where Q1 is the first feasible solution to the OD estimation submatrix, i.e., the first sampling point. n This is the nth feasible solution to the OD estimation submatrix, which is also the nth sampling point.

[0238] Where y represents the set of objective function values, f1, ..., f n Let f represent the objective function values ​​corresponding to the n sampling points, and f1 be the objective function value corresponding to the first feasible solution of the OD estimation submatrix, that is, the objective function value corresponding to the first sampling point. n This is the objective function value corresponding to the nth feasible solution of the OD estimation submatrix, which is also the objective function value corresponding to the nth sampling point.

[0239] Step S3: Based on n sampling points Q = {Q1, ..., Q...} n The objective function values ​​y = {f1, ..., fn} corresponding to n sampling points. n The mean and variance are calculated using Gaussian process regression, which serves as a surrogate model for the functional relationship between the objective function and the OD estimator matrix of the OD estimator subproblem.

[0240] The method for calculating the mean and variance using Gaussian process regression in this step will not be elaborated further in this application's embodiment.

[0241] Step S4: Construct a sampling function based on the above surrogate model, and solve for the extreme points of the sampling function, which are used as new sampling points Q. n+1 .

[0242] Step S5: Transfer the new sampling point Q n+1 Input the data into the simulation system to obtain the new sampling point Q. n+1 The objective function value f n+1 .

[0243] Step S6: Update Q = Q∪{Q} n+1},y=y∪{f n+1} Determine whether the iteration stopping condition has been met. If so, determine the OD estimation submatrix of the OD estimation subproblem based on the optimal sampling point among all sampling points. Otherwise, let n = n + 1 and repeat steps S2 to S6 until the iteration stopping condition is met.

[0244] The iteration stopping conditions may include: reaching the required number of iterations, or reaching the required convergence accuracy.

[0245] For example, in steps S2 to S6 above, the Bayesian optimization algorithm considering time correlation is used to solve each subproblem, with an upper limit of 80 sampling points; the initial number of sampling points is set to 10; the time segmentation strategy corrects the OD estimation subproblems within 3 time periods in the next step; and the number of overlapping estimation time periods for adjacent OD estimation subproblems is set to 1. The comparison results with the classic Bayesian optimization algorithm are shown in Table 3 below:

[0246] Solving the model CPU Time (h) MAPE Classical Algorithm 144.6h 33.2％ Improved Algorithm 10.53h 17.4％

[0247] Table 3

[0248] It should be noted that the specific implementation of steps S1 to S6 above can refer to the relevant technologies of Bayesian optimization algorithm, and will not be elaborated here.

[0249] Step 504: Integrate the OD estimation sub-matrices of each OD estimation subproblem to obtain the final OD estimation matrix.

[0250] This step can be considered as integrating multiple OD estimation sub-matrices into a single OD estimation matrix. Specific methods may include:

[0251] The OD estimation sub-matrices are merged pairwise until they are finally merged into a single OD estimation matrix. During the pairwise merging, for the two OD estimation sub-matrices to be merged, the virtual traffic cells that are only included in one OD estimation sub-matrice are treated as ordinary traffic cells and their matrix elements are merged. The matrix elements corresponding to the virtual traffic cells included in both OD estimation sub-matrices are split into matrix elements corresponding to the OD pairs formed by the ordinary traffic cells (including ordinary traffic cells and virtual traffic cells included in one of the two sub-road networks) between the two sub-road networks. The two sub-road networks are the sub-road networks corresponding to the OD estimation sub-matrices.

[0252] The following explains the pairwise merging process of the OD estimation submatrices:

[0253] Assuming that both sub-network 1 and sub-network 2 include virtual traffic zones 1, the OD estimation submatrix corresponding to sub-network 1 is OD estimation submatrix 1, and the OD estimation submatrix corresponding to sub-network 2 is OD estimation submatrix 2. The matrix obtained by merging OD estimation submatrix 1 and OD estimation submatrix 2 is called the merged matrix.

[0254] For matrix element 1 corresponding to an OD pair 1 in OD estimation submatrix 1:

[0255] If neither the origin (O) nor the destination (D) of OD pair 1 is a virtual traffic cell 1 (including ordinary traffic cells and virtual traffic cells that belong only to sub-network 1 and not to sub-network 2), then the matrix element 1 corresponding to OD pair 1 (i.e. the flow of OD pair 1) is used as the matrix element corresponding to OD pair 1 in the merge matrix.

[0256] If the starting point of OD pair 1 is not virtual traffic zone 1, but the ending point is virtual traffic zone 1, calculate the diversion ratio based on the matrix elements corresponding to all OD pairs with virtual traffic zone 1 as the starting point in the OD estimation submatrix 2. Based on the diversion ratio and the matrix element 1 corresponding to OD pair 1, calculate the matrix elements corresponding to OD pairs in the merge matrix that have the same starting point as OD pair 1, but whose ending point is a zone other than virtual traffic zone 1 in sub-network 2 (including ordinary traffic zones and virtual traffic zones that belong only to sub-network 1 and not to sub-network 2).

[0257] If the starting point of OD pair 1 is virtual traffic zone 1 and the ending point is not virtual traffic zone 1, then it is used to calculate the diversion ratio of the corresponding matrix element in OD estimation submatrix 2. Based on the element corresponding to the OD pair with the ending point of virtual traffic zone 1 in another OD estimation submatrix, the matrix element corresponding to the OD pair formed by the zones other than virtual traffic zone 1 between the two sub-road networks is calculated. For details, please refer to the case where the starting point of OD pair 1 in OD estimation submatrix 1 is not virtual traffic zone 1 but the ending point is virtual traffic zone 1, which will not be elaborated here.

[0258] Specific examples are as follows:

[0259] See Figure 6D As shown, matrices A and B are OD estimation submatrix 1 and OD estimation submatrix 2, respectively, and matrix C is the merge matrix. The virtual traffic cell included in both matrices A and B is virtual traffic cell 8. At this time:

[0260] For the matrix elements corresponding to OD pairs in matrices A and B whose starting and ending points are not virtual traffic zone 8 (e.g., matrix elements in matrix A whose row is zone 1 and column is zone 1), they can be directly moved from matrices A and B to matrix C and used as the matrix elements corresponding to the OD pair in matrix C.

[0261] For matrix elements in matrices A and B that are related to virtual traffic zone 8 (e.g., matrix elements in matrix A with rows representing zone 1, 2, or 3 and columns representing zone 8), traffic can be diverted proportionally to other traffic zones outside of virtual traffic zone 8. Diversion methods include:

[0262] The matrix elements corresponding to the OD pairs in matrix A, where the starting point is cell 1 to cell 3 and the ending point is cell 8 (i.e., 80, 60, and 40 in matrix A), are proportionally calculated to correspond to the matrix elements corresponding to the OD pairs in the merged matrix where the starting point is cell 1 to cell 3 and the ending point is cell 4 to cell 7 in another sub-network. This proportion can be referenced to the matrix elements corresponding to the OD pairs in matrix B, where the starting point is cell 4 to cell 7 and the ending point is cell 8 (i.e., 30, 80, 50, and 40 in matrix B).

[0263] Then, the matrix element in matrix A with the starting point of cell 1 and the ending point of cell 8 (i.e., 80 in matrix A) is proportionally divided into matrix elements in the merging matrix with the starting point of cell 1 and the ending points of cells 4 to 7, as follows:

[0264] In matrix C, the matrix element with starting point 1 and ending point 4 is 80*30 / 200. Here, 200 is the sum of the matrix elements in matrix B with starting points 4 to 7 and ending point 8 (that is, the sum of 30, 80, 50, and 40 in matrix B), and 30 is the matrix element in matrix B with starting point 4 and ending point 8 (that is, 30 in matrix B).

[0265] Similarly, the matrix elements in matrix C with the starting point being cell 1 and the ending point being cell 5 are: 80*80 / 200; the matrix elements in matrix C with the starting point being cell 1 and the ending point being cell 6 are: 80*50 / 200; the matrix elements in matrix C with the starting point being cell 1 and the ending point being cell 7 are: 80*40 / 200.

[0266] The calculation results are written into the corresponding positions in matrix C.

[0267] The matrix element in matrix A with a starting point of cell 2 and an ending point of cell 8 (i.e., 80 in matrix A) is proportionally divided into matrix elements in the merging matrix with a starting point of cell 2 and ending points of cells 4-7, as follows:

[0268] In matrix C, the matrix element with starting point 2 and ending point 4 is 60*30 / 200; the matrix element with starting point 2 and ending point 5 is 60*80 / 200; the matrix element with starting point 2 and ending point 6 is 60*50 / 200; and the matrix element with starting point 2 and ending point 7 is 60*40 / 200.

[0269] The calculation results are written into the corresponding positions in matrix C.

[0270] The matrix element in matrix A with a starting point of cell 3 and an ending point of cell 8 (i.e., 80 in matrix A) is proportionally divided into matrix elements in the merging matrix with a starting point of cell 3 and ending points of cells 4-7, as follows:

[0271] In matrix C, the matrix element with starting point 3 and ending point 4 is 40*30 / 200; the matrix element with starting point 3 and ending point 5 is 40*80 / 200; the matrix element with starting point 3 and ending point 6 is 40*50 / 200; and the matrix element with starting point 3 and ending point 7 is 40*40 / 200.

[0272] The calculation results are written into the corresponding positions in matrix C.

[0273] The above process calibrates the OD estimation matrix in traffic simulation parameters, making the traffic simulation parameters more accurate. Furthermore, by dividing the OD estimation problem into sub-problems, the computational cost of calculating the OD estimation matrix is ​​simplified.

[0274] Furthermore, when dividing the sub-time periods, there is overlap between the sub-time periods. When dividing the sub-road network, virtual traffic zones are set at the locations where the road network is split, which can prevent jumps in the estimated values ​​and make the final calculated OD estimation matrix more accurate.

[0275] The following is an exemplary description of the implementation of calibrating the path selection model parameters in step 202 of this application.

[0276] Generally, on highways, due to the limited number of available paths, the path selection model has little impact on the simulation. This step can be skipped for traffic simulations of highways where there are no two or more available paths. However, for urban road simulations, it is necessary to calibrate the path selection model based on actual road network data.

[0277] The Logit route selection model reflects travelers' sensitivity to travel costs, embodying a group characteristic. When simulating urban roads, it can be assumed that the same city exhibits consistent group characteristics, and the Logit route selection model can be calibrated using actual data from a specific area within that city. Similarly, when simulating highways, it can be assumed that the same province exhibits consistent group characteristics, and the model can be calibrated using actual data from a specific area within that province. Based on these assumptions, the data requirements can be reduced. Optionally, this calibration process can be completed offline, thereby improving the overall reliability of the simulation.

[0278] In the traffic assignment problem, each path is considered as a set of options, and the explanatory variables are the travel costs (time, expense, etc.) of each path. The logit model expression in SUMO is:

[0279]

[0280] Among them, P s K represents the probability of choosing path s. od Let Cost(s) represent the set of alternative paths between OD pairs, and Cost(s) represent the travel cost of path s. s Let CF represent the common factor between path s and the other paths, and Cost(l) represent the travel cost of path l. l This represents the common factor between path l and the other paths.

[0281]

[0282] L i Let L be the length of path i. j L is the length of path j. ij Let be the length of the road network where path i and path j overlap, and β and γ be the parameters.

[0283] In one embodiment, the problem can be simplified and the data requirements reduced by setting β = 0. In this case, the model is converted into a standard Logit model, and the parameter to be calibrated is θ.

[0284] For example, suppose for example Figure 7A The SiouxFalls network shown is used for path selection model calibration and contains 24 nodes and 76 road networks.

[0285] Step 202 can be implemented in two scenarios: one where the source data for parameter calibration is complete, and the other where the source data for parameter calibration is limited. These will be explained separately below:

[0286] Scenario 1: The source data for parameter calibration is complete and specific, including the road network traffic and vehicle travel time for the third time period and the third road network.

[0287] Alternatively, since the road network travel time of a vehicle can be inferred from the vehicle's speed in the road network, in other possible implementations, the road network travel time of a vehicle can be replaced by the vehicle's travel speed.

[0288] At this time, as Figure 7B As shown, step 202 can be achieved through the following steps:

[0289] Step 701: Extract the vehicle's driving path from the traffic observation data to obtain the feasible path set for each pair of ODs.

[0290] Step 702: Obtain the segment travel time and path flow data of the target road network, and calculate the probability of each vehicle choosing each path in the feasible path set based on the segment travel time and path flow data using a preset path selection model (e.g., logit model).

[0291] in, This represents the flow of the k-th path between OD and rs.

[0292] Step 703: Construct the likelihood function and solve for the parameter θ with the goal of maximizing the likelihood function (i.e., perform maximum likelihood estimation);

[0293]

[0294]

[0295] like Figure 7C The figure shows the change of the likelihood function. Convergence of calibration indicates the convergence of calibration, Likelihood indicates the likelihood function value, and Iteration number indicates the number of iterations. Based on this figure, it can be seen that the convergence time of the algorithm is less than 1 minute.

[0296] Scenario 2: The source data for parameter calibration is limited, specifically the road network traffic in the third time period and the third road network.

[0297] At this time, as Figure 8A As shown, step 202 can be achieved through the following steps:

[0298] Step 801: Extract information such as traffic flow, road network length, free-flow speed (or speed limit), and capacity from traffic observation data.

[0299] Step 802: For a given parameter θ, perform random user equilibrium (SUE) allocation based on a path selection model (e.g., logit model) and record the network traffic when equilibrium is reached.

[0300] Step 803: Calculate the relative error between the equilibrium network traffic and the actual network traffic, using the following formula:

[0301]

[0302] Where RP represents the relative error, x SUE x represents the network flow at equilibrium. real This represents the actual traffic flow on the road network.

[0303] The actual road network traffic can be calculated based on the road network trajectory data used in step 103.

[0304] Step 804: Solve for the parameters of the path selection model with the objective of minimizing the relative error RP.

[0305]

[0306] The following is a comparison of the results in the actual implementation: It took approximately 30 minutes. Logit-based SUE allocation was performed based on the calibrated parameter values, and the network traffic was compared with the actual traffic. The results are as follows... Figure 8B As shown.

[0307] The JS divergence value is 6.288 × 10⁻⁶. -5 The relative entropy value is 2.505 × 10⁻⁶. -4 It can be seen that the difference between the actual traffic and the allocated traffic is minimal.

[0308] The following is an exemplary description of the implementation of the supervised learning model for training the free-flow velocity in step 203.

[0309] Free-flow speed, or simply free speed, refers to the speed a driver naturally chooses based on road characteristics under conditions free from interference from other vehicles, obvious speed enforcement, and other external environmental factors. In other words, it is the speed at which the driver feels comfortable under current road conditions. In the simulation, it corresponds to the legal speed limit. Since each road in the simulation can have a legal speed limit set, which includes a speed limit and a speed factor, the speed factor, used to reflect differences in the distribution of speeds across the group, can remain unchanged. In this embodiment, the average vehicle speed on the target road network is mainly reflected by adjusting the legal speed limit, taking into account factors such as road width and weather conditions.

[0310] Free-flow velocity is difficult to measure in reality due to stringent observation conditions. It is usually obtained by observing the average speed of passenger cars under low to medium flow conditions (not exceeding 1300 standard vehicles / hour / lane), and the observation method adopts continuous observation or systematic sampling with a sample size of no less than 100 cars.

[0311] Taking the simulation software SUMO as an example, for macroscopic parameters other than traffic demand, they are classified according to the adjustable and non-adjustable macroscopic parameters in the simulation, and their mapping relationship is shown in Table 4 below:

[0312]

[0313] Table 4

[0314] In this embodiment, the characterization of free-flow velocity is divided into two cases: one is where the free-flow velocity of the target road network can be observed, i.e., the free-flow velocity of the target road network can be obtained; the other is where the free-flow velocity of the target road network cannot be observed. The specific implementation of step 203 in each of the two cases is illustrated below.

[0315] Scenario 1: When the free-flow velocity of the target road network (e.g., a highway network) is observable, a supervised learning model can be constructed to estimate the free-flow velocity of the target road network. In this case, the external conditions that can be considered are relatively flexible and extensive. For example... Figure 9 As shown, step 203 may include the following steps:

[0316] Step 901: Obtain observation data of free-flow velocity of the target road network.

[0317] Optionally, the free-flow velocity of the target road network can be observed in advance to obtain the free-flow velocity of the target road network. Furthermore, the external characteristic data corresponding to the free-flow velocity can be recorded, including but not limited to: weather, lane width, number of lanes, road curvature, slope and other external factor data.

[0318] Step 902: Divide the observation data into training set and test set.

[0319] Specifically, the observation data can be divided into two parts, serving as the training set and the test set. It should be noted that the observation data included in the training set and the test set do not overlap.

[0320] Step 903: Train the supervised learning model using the training set.

[0321] Optionally, the initial model for the supervised learning model can be a neural network model, a decision tree model, or the like.

[0322] Step 904: Test the trained supervised learning model using the test set, and adjust the supervised learning model based on the test results until the test is passed.

[0323] It should be noted that if the test passes, the supervised learning model does not need to be adjusted based on the test results in this step.

[0324] Specifically, metrics such as MAE, RMSE, and MAPE can be used to evaluate the prediction accuracy of free-flow velocity in supervised learning models.

[0325]

[0326]

[0327]

[0328] In actual simulations, the speed limit configuration of each road network in the simulation is modified based on various external feature data, thereby indirectly reflecting the impact of external information such as road network width and number of lanes on the simulation. The impact of weather factors on road network speed limits is based on future weather forecasts and can be updated at a frequency of 1 hour. The specific method is explained in step 207 and will not be repeated here.

[0329] Scenario 2: When the free-flow velocity of the target road network (e.g., highway network) is missing or unavailable, it can be estimated by consulting the "U.S. Highway Capacity Manual":

[0330] FFS = BFFS - f LW -f LC -f N -f ID

[0331] in:

[0332] FFS - Estimated free-flow velocity (km / h);

[0333] BFFS - Basic free-flow speed, 110 km / h (urban areas) or 120 km / h (rural areas);

[0334] f LW - Lane width correction values ​​(km / h) can be obtained from a table;

[0335] f LC - The right-side lateral clearance correction value (km / h) can be obtained from a table;

[0336] f N - The lane number correction value (km / h) can be obtained from a table;

[0337] fID - The density correction value for grade separation (km / h) can be obtained from a table.

[0338] The implementation of step 204 will be illustrated below.

[0339] like Figure 10A As shown, step 204 can be achieved through the following steps:

[0340] Step 1001: Determine if there is driving behavior data of drivers in the target road network. If there is, proceed to step 1002; otherwise, proceed to step 1003.

[0341] Step 1002: Analyze the drivers' driving behavior data, fit the speed distribution of the driver group, and this branch of the process ends.

[0342] Step 1003: Establish the objective function, obtain the feasible solutions of n sets of driver speed distribution as n sampling points, input the n sampling points into the simulation system respectively, and obtain the simulation values ​​corresponding to the n sampling points. These simulation values ​​are the simulation values ​​of cross-sectional flow or cross-sectional speed of the target road network.

[0343] Among them, the mean absolute percentage error (MAPE) between the simulated value and the actual value of cross-sectional flow or cross-sectional velocity can be used as the objective function.

[0344] Step 1004: Calculate the objective function value based on the mean absolute percentage error (MAPE) between the simulated and actual values ​​of the cross-sectional flow rate or velocity.

[0345] Where K is the number of detectors, and T is the number of time periods detected by the detectors. and Let represent the output flow rate (velocity) of the actual and simulated detectors at the k-th detector during the t-th detection period, respectively:

[0346]

[0347] For example, given the actual value data of the detector, an example of detector cross-sectional data is shown below:

[0348]

[0349] Table 5

[0350] Step 1005: Construct a Gaussian process model and establish a surrogate model between the objective function value and the velocity distribution parameters.

[0351] Step 1006: Construct a sampling function based on the surrogate model, and solve for the extreme points of the sampling function as new sampling points.

[0352] Step 1007: Input the new sampling points into the simulation to obtain the simulation values ​​corresponding to the new sampling points, and then calculate the objective function values ​​of the new sampling points.

[0353] The optimal speed distribution was obtained. In the example, the final calibrated speed distribution parameters were: expected speed distribution of the large vehicle = 0.914, variance of the large vehicle speed distribution = 0.19, expected speed distribution of the small vehicle = 0.749, variance of the small vehicle speed distribution = 0.09, and the time taken was 2620s. The MAPE value decreased from 36.5% of the initial parameters to 10.8%.

[0354] Step 1008: Determine whether the iteration stopping condition is met based on all sampling points, including n sampling points and the new sampling point, and the objective function value corresponding to the sampling points. If yes, determine the velocity distribution parameter based on the optimal sampling point among all sampling points. Otherwise, let n = n + 1 and repeat steps 1004 to 1008.

[0355] like Figure 10B As shown, in one embodiment, the final calibrated speed distribution parameters are: expected speed distribution of the main vehicle = 0.914, variance of speed distribution of the main vehicle = 0.19, expected speed distribution of the trolley = 0.749, variance of speed distribution of the trolley = 0.09, time taken 2620s, and the MAPE value decreased from 36.5% of the initial parameters to 10.8%.

[0356] The implementation of step 207 is illustrated below. Figure 11 As shown, step 207 can be achieved through the following steps:

[0357] Step 1101: Obtain historical road network traffic statistics and OD data.

[0358] Historical road network traffic and historical OD data can be collected, and the above data can be aggregated and statistically analyzed to obtain the above historical road network traffic statistics and OD statistics.

[0359] Step 1102: Construct an OD prediction model.

[0360] There are adjacency relationships between traffic lanes, and their adjacency matrix is ​​suitable for capturing the spatial features generated by OD using convolutional neural networks. At the same time, recurrent neural networks can capture temporal features. Combining the two in modeling can improve the overall prediction accuracy and robustness of the model.

[0361] Step 1103: Convert the above statistical data into tensors that match the OD prediction model.

[0362] The main purpose of this step is to convert the statistical data into tensors according to a preset format, so that the tensors can be matched with the OD prediction model to execute subsequent steps. The implementation of this step can be found in relevant technologies and will not be elaborated here.

[0363] Step 1104: Use the above tensors to train and validate the OD prediction model.

[0364] The methods for training and validating the model in this step can be found in relevant technologies, and will not be elaborated here.

[0365] Step 1105: Model parameter tuning and optimization.

[0366] Step 1105 is an optional step. If the OD prediction model obtained in step 1104 does not meet the preset accuracy requirements, then step 1105 is executed to adjust and optimize the parameters of the OD prediction model obtained in step 1104 so that the OD prediction model meets the preset accuracy requirements. If the OD prediction model obtained in step 1104 meets the preset accuracy requirements, then step 1105 can be skipped.

[0367] The specific implementation of this step can be found in relevant technical implementations, which will not be elaborated here.

[0368] Step 1106: Save and generalize the optimal model to obtain the OD prediction model.

[0369] The specific implementation of this step can be found in relevant technical implementations, which will not be elaborated here.

[0370] Step 1107: During real-time simulation, input the historical and current data of the target road network into the OD prediction model to obtain the OD prediction matrix.

[0371] The OD prediction matrix can then be configured as OD prediction information in the traffic simulation model.

[0372] The implementation of step 208 is illustrated below. Figure 12 As shown, step 208 can be achieved through the following steps:

[0373] Step 1201: Obtain the predicted weather information and real-time weather information for the actual simulation period.

[0374] The actual simulation period refers to the time period during which the simulation system is triggered to perform traffic simulation.

[0375] Step 1202: Determine the weather-affected areas in the target road network based on the predicted weather information and the real-time weather information.

[0376] Step 1203: Extract road network-related data and current free-flow velocity within the weather-affected area.

[0377] Step 1204: Input the road network related data and the free-flow speed into the supervised learning model of free-flow vehicle speed to obtain the corrected value of the free-flow speed under the predicted weather information.

[0378] Step 1205: Correct the free-flow velocity in the traffic simulation model according to the correction value.

[0379] Steps 1201 to 1205 above can be implemented with reference to relevant technologies, and will not be elaborated here.

[0380] The parameter calibration method provided in this application addresses the shortcomings of existing traffic simulation parameter calibration methods by providing a comprehensive and systematic analysis of the simulation parameters. It integrates multi-source data and the data requirements for different parameter calibrations, combining the inherent properties of the parameters with the data requirements of different parameter calibrations to construct a macro-micro integrated traffic simulation parameter calibration method. Furthermore, it considers the impact of external environmental parameters on the simulation by mapping environmental factors to simulation parameters. Simultaneously, it introduces a Bayesian optimization-based simulation optimization method during the parameter calibration process to improve the method's computational efficiency and ability to handle high-dimensional parameters. Ultimately, it systematically integrates the calibration of various parameters to form a logically complete, macro-micro integrated, and highly interpretable traffic simulation parameter calibration method.

[0381] This application's embodiments calibrate the microscopic kinematic model parameters and macroscopic traffic demand and speed distribution parameters in microscopic traffic simulation. First, the microscopic parameters are calibrated to achieve a reasonable characterization of individual behavior. Then, the macroscopic parameters are iteratively calibrated to calibrate traffic demand, macroscopic group parameters, and environmental parameters. Furthermore, offline and online calibrations are separated, ensuring accuracy while meeting real-time requirements. Thus, by constructing a rational framework, both microscopic and macroscopic simulation parameters can be calibrated simultaneously. This solves the problems of insufficient systematicity and comprehensiveness in traffic simulation parameter calibration, and the difficulty in simultaneously calibrating high-dimensional and diverse parameters, improving the accuracy and rationality of simulation parameter calibration and increasing computational efficiency.

[0382] Furthermore, an improved method for estimating dynamic traffic demand parameters is proposed, which uses an improved Bayesian optimization method to perform global optimization for high-dimensional dynamic traffic demand. This solves the problems of low efficiency and difficulty in global optimization for high-dimensional dynamic traffic demand parameter calibration, and improves the efficiency and accuracy of high-dimensional dynamic traffic demand calibration.

[0383] Furthermore, the estimation methods for environmental parameters and macroscopic group characteristics are improved, a mapping relationship between environmental parameters and simulation is constructed, and feature learning and Bayesian optimization methods are used to estimate the parameters. This solves the problem of difficulty in characterizing the influence of environmental parameters in simulation and improves the accuracy of traffic simulation parameters.

[0384] This application also provides an electronic device, including: a processor; a memory; and a computer program, wherein the computer program is stored in the memory, and when the electronic device executes the computer program, it causes the electronic device to perform this application. Figures 1-12 The method provided in the illustrated embodiment.

[0385] This application also provides a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to execute this application. Figures 1-12 The method provided in the illustrated embodiment.

[0386] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0387] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0388] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0389] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0390] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A simulation parameter calibration method, characterized in that, include: For the time period of the OD estimation problem, the time period is divided into several consecutive time periods, each sub-time period includes several consecutive time periods, and there is an overlap between two adjacent sub-time periods; For the road network in the OD estimation problem, the road network is divided into at least two independent sub-road networks; By combining the sub-time periods with the corresponding sub-road networks, multiple OD estimation sub-problems are obtained; Calculate the OD estimation submatrix for each of the OD estimation subproblems; By integrating the OD estimation submatrices, the OD estimation matrix of the OD estimation problem is obtained; The OD estimation matrix of the aforementioned OD estimation problem is used as the OD configuration information in the traffic simulation model; The calculation of the OD estimation sub-matrix for each of the OD estimation sub-problems includes: For an OD estimation subproblem, construct the objective function; Within the feasible region, n sampling points of the OD estimation submatrix of the OD estimation subproblem are randomly selected, and the n sampling points are input into the simulation system respectively to obtain the objective function values ​​corresponding to the n sampling points; Based on the n sampling points and the objective function values ​​corresponding to the n sampling points, the mean and variance are calculated through Gaussian process regression, which serves as a proxy model for the functional relationship between the objective function and the OD estimation submatrix of the OD estimation subproblem; Based on the aforementioned proxy model, a sampling function is constructed, and the extreme points of the sampling function are solved as new sampling points. The new sampling points are input into the simulation system to obtain the objective function values ​​corresponding to the new sampling points; Based on all sampling points, including the n sampling points and the new sampling point, and the objective function values ​​corresponding to the sampling points, determine whether the iteration stopping condition is met. If yes, determine the OD estimation submatrix of the OD estimation subproblem based on the optimal sampling point among all sampling points. If no, replace the n sampling points with all sampling points, return to the step of calculating the mean and variance through Gaussian process regression based on the n sampling points and the objective function values ​​corresponding to the n sampling points, and replace n with n+1, and so on, until the iteration stopping condition is met.

2. The method according to claim 1, characterized in that, Also includes: The OD prediction matrix is ​​calculated using the OD prediction model based on the OD estimation matrix and the OD data of the target road network. Configure the OD prediction matrix as OD prediction information in the traffic simulation model.

3. The method according to claim 2, characterized in that, The OD prediction model is pre-trained, and the training method includes: Obtain statistical data on historical road network traffic statistics and OD data; Construct the initial model for OD prediction; The statistical data is converted into tensors that match the initial model; The initial model is trained and validated using the tensor to obtain the OD prediction model.

4. The method according to any one of claims 1 to 3, characterized in that, Also includes: Calculate the micro-parameters of the traffic simulation model and use the calculated micro-parameters as the configuration information of the traffic simulation model; The micro-parameters include: parameters of the kinematic model, and / or parameters of the path selection model; the kinematic model includes: a car-following model, and / or a lane-changing model.

5. The method according to claim 4, characterized in that, The calculated micro-parameters of the traffic simulation model are used as configuration information for the traffic simulation model, including: Extract vehicle following behavior data from vehicle trajectory data; The optimization problem of the preset car-following model is solved based on the car-following behavior data to obtain the car-following model parameters that satisfy the optimization problem. The optimization problem includes the car-following model parameters of the car-following model. The car-following model parameters are used as the configuration information for the car-following model in the traffic simulation model.

6. The method according to claim 5, characterized in that, The optimization objectives of the car-following model optimization problem include: ； in, It is the estimated speed of the i-th vehicle at time t. It is the actual speed of the i-th vehicle at time t. It is the estimated distance between the i-th vehicle and the vehicle in front at time t. It is the actual distance between the front and rear vehicles of the i-th vehicle at time t.

7. The method according to claim 5, characterized in that, The step of solving the optimization problem of the preset car-following model based on the car-following behavior data to obtain the car-following model parameters that satisfy the optimization problem includes: Based on the car-following behavior data, a heuristic algorithm is used to solve the optimization problem of the preset car-following model, and the car-following model parameters that satisfy the optimization problem are obtained.

8. The method according to claim 4, characterized in that, The calculation of the micro-parameters of the traffic simulation model, using the calculated micro-parameters as the configuration information of the traffic simulation model, also includes: Extract lane-changing behavior data from vehicle trajectory data; The optimization problem of the preset lane-changing model is solved based on the lane-changing behavior data to obtain the lane-changing parameters that satisfy the optimization problem of the lane-changing model. The optimization problem of the lane-changing model includes the lane-changing parameters of the lane-changing model. The lane-changing model parameters are used as the configuration information for the lane-changing model in the traffic simulation model.

9. The method according to claim 8, characterized in that, The optimization objectives of the lane-changing model optimization problem include: ； in, This is the estimated number of lane changes for the vehicle. This refers to the actual number of lane changes made by the vehicle.

10. The method according to claim 8, characterized in that, The step of solving the optimization problem of the preset lane-changing model based on the lane-changing behavior data to obtain lane-changing parameters that satisfy the optimization problem of the lane-changing model includes: By traversing the feasible range of lane change counts from smallest to largest, the lane change parameters that satisfy the optimization problem of the lane change model are obtained.

11. The method according to claim 4, characterized in that, The calculation of the micro-parameters of the traffic simulation model, using the calculated micro-parameters as the configuration information of the traffic simulation model, also includes: Based on the target road network, the driving paths of vehicles are extracted to obtain the feasible path set for each pair of ODs in the target road network; Obtain the segment travel time and path flow data of the target road network, and use a preset path selection model to calculate the probability of each vehicle choosing each path in the feasible path set based on the segment travel time and path flow data; Construct a likelihood function and solve for the parameters in the path selection model with the goal of maximizing the likelihood function; The solved parameters are used as configuration parameters for the path selection model in the traffic simulation model.

12. The method according to claim 4, characterized in that, The calculation of the micro-parameters of the traffic simulation model, using the calculated micro-parameters as the configuration information of the traffic simulation model, also includes: Obtain traffic flow information, segment length, free-flow velocity, and capacity information of the target road network; For the parameters given in the preset path selection model, random user equilibrium allocation is performed based on the path selection model according to the traffic information, road segment length, free flow speed and capacity information, and the road network traffic when equilibrium is reached is recorded. Find the relative error between the equilibrium network traffic and the actual network traffic. The parameters of the path selection model are solved with the objective of minimizing the relative error; The solved parameters are used as configuration parameters for the path selection model in the traffic simulation model.

13. The method according to any one of claims 1 to 3, characterized in that, Also includes: Calculate the macroscopic parameters of the traffic simulation model and use the calculated macroscopic parameters as the configuration information of the traffic simulation model; The macroscopic parameters include: driver speed distribution parameters, and / or, free-flow vehicle speed.

14. The method according to claim 13, characterized in that, The macroscopic parameters of the traffic simulation model are calculated, and the calculated macroscopic parameters are used as the configuration information of the traffic simulation model, including: The free-flow velocity of the target road network is calculated using a supervised learning model based on free-flow velocity observation data of the target road network; the supervised learning model is pre-trained. The free-flow velocity is used as the speed limit for the road network in the traffic simulation model.

15. The method according to claim 14, characterized in that, The supervised learning model is pre-trained, and the training method includes: Obtain training samples; the training samples include observational data of free-flow velocity; The training samples are used to train the preset model to obtain a supervised learning model, which is used to calculate the free flow velocity of the target road network.

16. The method according to claim 15, characterized in that, Also includes: Obtain test samples, which include observational data of free-flow velocity, and the test samples do not overlap with the training samples; The accuracy of the supervised learning model in predicting free-flow velocity was tested using test samples. The model parameters of the supervised learning model are adjusted based on the test results until the supervised learning model reaches the preset prediction accuracy.

17. The method according to claim 13, characterized in that, The calculation of macroscopic parameters of the traffic simulation model, using the calculated macroscopic parameters as configuration information for the traffic simulation model, also includes: Based on the driving behavior data of the target road network, the speed distribution of the driver group is obtained by fitting, and the speed distribution parameters of the driver group are obtained. The obtained speed distribution parameters are used as the speed distribution parameters of the driver group in the traffic simulation model.

18. The method according to claim 13, characterized in that, The macroscopic parameters of the traffic simulation model also include: Establish the objective function; Obtain n feasible solutions for driver speed distributions as n sampling points, input the n sampling points into the simulation system respectively, and obtain the objective function values ​​corresponding to the n sampling points; Based on the n sampling points and the objective function values ​​corresponding to the n sampling points, a surrogate model between the objective function and the velocity distribution is established through Gaussian process regression; Based on the aforementioned proxy model, a sampling function is constructed, and the extreme points of the sampling function are solved as new sampling points. The new sampling points are input into the simulation system to obtain the objective function values ​​corresponding to the new sampling points; Determine whether the objective function value has reached the convergence condition. If yes, determine the speed distribution parameters of the driver group based on the optimal sampling point among all the sampling points. If no, return to the step of obtaining the feasible solution of n groups of driver speed distributions, and replace n with n+1, and so on, until the iteration stopping condition is met.

19. The method according to claim 1, characterized in that, Also includes: Obtain predicted and real-time weather information for the actual simulation period; The weather-affected areas in the target road network are determined based on the predicted weather information and the real-time weather information. Extract road network-related data and current free-flow velocity within the area affected by the weather; The road network-related data and the free-flow speed are input into the supervised learning model of free-flow vehicle speed to obtain the corrected value of the free-flow speed under the predicted weather information. The free-flow velocity in the traffic simulation model is corrected based on the correction value.

20. The method according to any one of claims 1, 2, 3, or 19, characterized in that, Also includes: Based on the traffic simulation model, a simulation was performed to obtain traffic simulation results; Determine whether the traffic simulation results have reached the preset simulation accuracy; if so, the simulation parameter calibration ends. Otherwise, the OD configuration information in the traffic simulation model is redefined until the traffic simulation results reach the preset simulation accuracy.

21. An electronic device, characterized in that, include: Processor; memory; And a computer program, wherein the computer program is stored in the memory, and when the computer program is executed by the electronic device, causes the electronic device to perform the method of any one of claims 1 to 20.

22. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1 to 20.

Images