A Road Network-Level Traffic Prediction Method Integrating Transfer Learning and Macro-Base Maps

Abstract

This invention relates to the field of intelligent transportation technology, specifically to a road network-level traffic prediction method integrating transfer learning and a macroscopic basic graph. The method includes: dividing the regional road network based on a semantic segmentation algorithm; determining the target domain and source domain based on the macroscopic basic graph; constructing a data modeling goodness of fit based on Pearson correlation coefficient and maximum mean difference, determining the target domain detector and source domain detector corresponding to the maximum data modeling goodness of fit, and generating a target domain training sample set and its corresponding ground truth set, and a source domain training sample set and its corresponding ground truth set using a sliding window method; training a traffic prediction model using the source domain training sample set and its corresponding ground truth set to obtain the source domain traffic prediction model; and adaptively training the source domain traffic prediction model using the filtered target domain training sample set and its corresponding ground truth set to obtain the target domain traffic prediction model. This invention improves the accuracy of traffic data prediction in areas with insufficient data, thereby improving traffic management efficiency and alleviating traffic congestion.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a road network-level traffic prediction method that integrates transfer learning and macroscopic basic graphs. Background Technology

[0002] With the increase in motor vehicles, various urban traffic problems are becoming increasingly serious, with congestion, traffic accidents, and exhaust emissions becoming common problems in daily life. Existing research has demonstrated in practice in different cities that Intelligent Transportation Systems (ITS) are an effective method for solving urban traffic problems. In order to better utilize Intelligent Transportation Systems (ITS) to solve urban traffic problems, it is necessary to predict urban traffic flow.

[0003] Currently, there is a large body of research on traffic flow prediction in existing technologies, much of which is based on data-driven methods, such as machine learning and deep learning, and these methods have shown good predictive performance. However, existing data-driven methods require extensive training of the traffic prediction model, resulting in high computational costs. Furthermore, the training process requires a large amount of traffic data, which typically needs to be sufficiently representative to cover all traffic modes in the region. However, few regions actually meet these data requirements, leading to low practicality and applicability of the prediction models. In addition, most regions suffer from data gaps and missing data, resulting in low accuracy of traffic flow prediction models when applied to most areas. Consequently, these methods cannot provide effective traffic flow prediction data for intelligent transportation systems and ultimately fail to effectively solve urban traffic problems. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problems of the high training cost of the prediction model due to the large amount of training required in the existing technology; the prediction model requires that the regional traffic data be sufficiently representative to represent all traffic modes in the region, but in reality most regions do not meet this requirement, resulting in poor practicality and applicability of the prediction model; and the prediction effect of the prediction model is relatively poor when regional traffic data is missing or lacking, which makes it impossible to provide effective traffic flow prediction data for intelligent transportation systems.

[0005] To address the aforementioned technical problems, this invention provides a road network-level traffic prediction method that integrates transfer learning and macroscopic basic graphs, comprising:

[0006] Based on historical traffic data of the regional road network and the preset number of regional divisions, the regional road network is divided using a semantic segmentation algorithm to obtain the various sub-regions of the regional road network.

[0007] Based on the macro-basic map of each sub-region, obtain the critical occupancy rate of each sub-region; select one sub-region as the target domain, and take other sub-regions whose critical occupancy rate is within the critical interval of the target domain as each critically related sub-region of the target domain. Within the critical interval of the target domain, calculate the area of ​​the curve enclosed by the macro-basic map of the target domain and the macro-basic maps of each critically related sub-region, and take the critically related sub-region corresponding to the smallest curve area as the source domain; if there is no critical occupancy rate of other sub-regions within the critical interval of the target domain, then re-divide the regional road network.

[0008] Taking each detector in the target domain as the center, obtain the flow rate values ​​of each modeled road in the target domain at each time moment, and establish the flow rate time matrix of each detector in the target domain; taking each detector in the source domain as the center, obtain the flow rate values ​​of each modeled road in the source domain at each time moment, and establish the flow rate time matrix of each detector in the source domain.

[0009] Based on the Pearson correlation coefficient and unbiased estimator between the flow-time matrices of each detector in the target domain and the flow-time matrices of each detector in the source domain, the goodness of data modeling between the flow-time matrices of each detector in the target domain and the flow-time matrices of each detector in the source domain is obtained.

[0010] Centered on the target domain detector and source domain detector corresponding to the highest obtained data modeling goodness, respectively, the flow rate and occupancy values ​​of each modeled road in the target domain and source domain at each time step are obtained. Then, using the sliding window method, the sliding window size is set to... The sliding step size is 1. For a modeling road, the sliding step size is 1. The traffic value or occupancy value at time n is used as a training sample, and the traffic value or occupancy value at time n is used as a training sample. The flow value or occupancy rate at any given time is used as the true value corresponding to the training sample, thereby generating the target domain training sample set and its corresponding true value set, and the source domain training sample set and its corresponding true value set.

[0011] The traffic prediction model is trained using the source domain training sample set and its corresponding real value set to obtain the source domain traffic prediction model.

[0012] By using the filtered target domain training sample set and its corresponding true value set, the source domain traffic prediction model is adaptively trained to obtain the target domain traffic prediction model.

[0013] Preferably, the traffic prediction model is a multi-convolutional neural network coupled traffic prediction model;

[0014] The multi-convolutional neural network coupled traffic prediction model includes three convolutional neural networks and two neural networks;

[0015] The source domain training set includes a traffic training sample set and a occupancy training sample set.

[0016] Each traffic training sample in the source domain training sample set is input into the first convolutional neural network, and the traffic prediction value corresponding to each traffic training sample is output.

[0017] The occupancy training samples in the source domain training sample set are input into the second convolutional neural network, and the occupancy prediction value corresponding to each occupancy training sample is output.

[0018] After the source domain training samples are normalized, each traffic training sample and each occupancy training sample are integrated to obtain each joint traffic-occupancy sample. The resulting joint traffic-occupancy sample is then input into the third convolutional neural network, which outputs the joint traffic-occupancy prediction value corresponding to each joint traffic-occupancy sample.

[0019] After standardizing the individual predicted values ​​of each flow, which are divided from the joint flow-occupancy predicted values ​​corresponding to each joint flow-occupancy sample, they are integrated with the flow predicted values ​​corresponding to each flow training sample, and then input into the first neural network to output the final predicted values ​​of the flow corresponding to each flow training sample.

[0020] After standardizing the individual predicted values ​​of each occupancy rate from the joint traffic-occupancy rate predicted values ​​corresponding to each joint traffic-occupancy rate sample, the predicted values ​​of each occupancy rate training sample are integrated with the predicted values ​​of each occupancy rate training sample, and then input into the second neural network to output the final predicted values ​​of the occupancy rate corresponding to each occupancy rate training sample.

[0021] Preferably, the step of training the traffic prediction model using the source domain training sample set and its corresponding real value set to obtain the source domain traffic prediction model includes:

[0022] Based on the first convolutional neural network, the spatiotemporal features of the flow values ​​at each time step within each flow training sample in the source domain training sample set are extracted to obtain the flow prediction value corresponding to each flow training sample.

[0023] Based on the second convolutional neural network, the spatiotemporal features of the occupancy values ​​at each time step within each occupancy training sample in the source domain training sample set are extracted to obtain the occupancy prediction value corresponding to each occupancy training sample.

[0024] The source domain training samples are integrated and normalized to obtain the traffic training samples and the occupancy training samples to obtain the joint traffic-occupancy samples. The joint traffic-occupancy samples are then input into the third convolutional neural network to output the joint traffic-occupancy prediction value corresponding to each joint traffic-occupancy sample.

[0025] After standardizing the individual predicted values ​​of each flow, which are divided from the joint flow-occupancy predicted values ​​corresponding to each joint flow-occupancy sample, they are integrated with the flow predicted values ​​corresponding to each flow training sample, and then input into the first neural network to output the final predicted values ​​of the flow corresponding to each flow training sample.

[0026] After standardizing the individual predicted values ​​of each occupancy rate from the joint traffic-occupancy rate predicted values ​​corresponding to each joint traffic-occupancy rate sample, they are integrated with the occupancy rate predicted values ​​corresponding to each occupancy rate training sample, and then input into the second neural network to output the final predicted values ​​of the occupancy rate corresponding to each occupancy rate training sample.

[0027] Based on the final predicted values ​​of traffic flow corresponding to each traffic flow training sample, the final predicted values ​​of occupancy corresponding to each occupancy training sample, and the true value set corresponding to the source domain training sample set, the loss function is calculated. By minimizing the loss function, the parameters of the multi-convolutional neural network coupled traffic prediction model are obtained, and the training of the multi-convolutional neural network coupled traffic prediction model is completed. The trained multi-convolutional neural network coupled traffic prediction model is used as the source domain traffic prediction model.

[0028] Preferably, the macroscopic basic graph of the target domain and its first... The area under the curve enclosed by the macroscopic basic plots of the critically correlated subregions is:

[0029]

[0030] in, A macroscopic basic diagram representing the target domain; The first term of the target domain One critically correlated sub-region; This indicates the critical occupancy rate of the target domain; Indicates the critical interval range of the target domain; This represents the interval threshold.

[0031] Preferably, calculating the Pearson correlation coefficient between the flow-time matrix of each detector in the target domain and the flow-time matrix of each detector in the source domain includes:

[0032] For any detector X in the target domain and any detector Y in the source domain, calculate the Pearson correlation coefficient between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is:

[0033]

[0034] ,

[0035]

[0036] in, This represents the flow-time matrix of the target domain detector X; This represents the flow-time matrix of the source domain detector Y; This represents the covariance between the flow-time matrix of the target domain detector X and the flow-time matrix of the source domain detector Y; This represents the flow-time matrix of the target domain detector X in a sample size of... The standard deviation in the sample; Indicates that the target domain detector X is in the first... The flow-time matrix at any given moment; This represents the mean of the flow-time matrix of the target domain detector X; This represents the flow-time matrix of the source domain detector Y with a sample size of The standard deviation in the sample; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The flow-time matrix at any given moment; This represents the mean of the flow time matrix of the source domain detector Y.

[0037] Preferably, calculating the unbiased estimator between the flow-time matrix of each detector in the target domain and the flow-time matrix of each detector in the source domain includes:

[0038] For any detector X in the target domain and any detector Y in the source domain, calculate the maximum mean difference between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is:

[0039]

[0040] in, Indicates that the target domain detector X is in the first... The flow-time matrix at any given moment; This represents the number of samples in the flow-time matrix of the target domain detector X; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The flow-time matrix at any given moment; This represents the number of samples in the flow-time matrix of the source domain detector Y;

[0041] Through mapping function Projecting the flow time matrices of target domain detector X and source domain detector Y into the kernel Hilbert space yields the kernel function between target domain detector X and source domain detector Y, whose expression is:

[0042]

[0043] in, The target domain detector X represents the first... The mapping function corresponding to the flow-time matrix at any given time; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The mapping function corresponding to the flow-time matrix at any given time;

[0044] Based on the kernel function expression, an unbiased estimator is calculated between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is:

[0045]

[0046] By performing kernel operations using the Gaussian kernel function, the kernel function value between the target domain detector X and the source domain detector Y is obtained, thus yielding the unbiased estimator value between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain; where the Gaussian kernel function expression is:

[0047]

[0048] in, This represents the width parameter of the Gaussian kernel function.

[0049] Preferably, the method of obtaining the data modeling goodness of the flow-time matrix between the detectors in the target domain and the detectors in the source domain based on the Pearson correlation coefficient and the unbiased estimator between the flow-time matrices of each detector in the target domain and the detectors in the source domain includes:

[0050] For any detector X in the target domain and any detector Y in the source domain, the Pearson correlation coefficient is used to determine the relationship between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. and unbiased estimators between them The goodness of data modeling between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain is obtained. Its expression is:

[0051]

[0052] in, Indicates the first weighting coefficient; Second weighting coefficient.

[0053] Preferably, taking the target domain detector corresponding to the highest data modeling goodness as the center, the flow rate and occupancy value of each modeling road within the target domain at each time step are obtained, and the sliding window method is used, with the sliding window size set to... The sliding step size is 1. For a modeling road, the sliding step size is 1. The traffic value or occupancy value at time n is used as a training sample, and the traffic value or occupancy value at time n is used as a training sample. The flow rate or occupancy rate at any given time is used as the true value corresponding to that training sample, thereby generating the target domain training sample set and its corresponding true value set, including:

[0054] Centered on the target domain detector corresponding to the highest data modeling excellence, the flow rate and occupancy rate of each modeling road in the target domain at each time are obtained, resulting in the target domain flow rate time matrix and the target domain occupancy rate time matrix.

[0055] Based on the target domain traffic time matrix, and according to the sliding window size, prediction window size, and sliding step size, a target domain traffic training sample set and its corresponding ground truth set are generated, with the following expressions:

[0056]

[0057] ,

[0058]

[0059]

[0060] in, , These represent the target domain traffic training sample set and the corresponding real value set, respectively. Indicates the first The target domain traffic training sample matrix contains the first target domain traffic training sample matrix. Each sliding step corresponds to a modeling path within the target domain. The flow rate at any given moment; Indicates the first The true value vector corresponding to the nth target domain traffic training sample matrix contains the nth target domain traffic training sample matrix. The first sliding step corresponds to the first modeling road in the target domain. The actual value of the flow at any given moment; Indicates the number of modeled roads within the target domain; Indicates the total number of moments; Indicates the size of the sliding window; Indicates the prediction window size. ; Indicates the number of slides;

[0061] Based on the target domain occupancy time matrix, and according to the sliding window size, prediction window size, and sliding step size, a training sample set and its corresponding true value set for the target domain occupancy are generated, with the following expressions:

[0062]

[0063] ,

[0064] ,

[0065]

[0066] in, , These represent the training sample set for target domain occupancy and the corresponding set of true values, respectively. Indicates the first The training sample matrix for the occupancy of the target domain contains the th target domain occupancy. Each sliding step corresponds to a modeling path within the target domain. The occupancy value at any given moment; Indicates the first The true value vector corresponding to the occupancy training sample matrix contains the th The first sliding step corresponds to the first modeling road in the target domain. The true value of time occupancy;

[0067] Based on the target domain traffic training sample set and its corresponding real value set, and the target domain occupancy training sample set and its corresponding real value set, construct the target domain training sample set and its corresponding real value set.

[0068] Preferably, taking the source domain detector corresponding to the obtained maximum data modeling goodness as the center, the flow rate and occupancy value of each modeling road in the source domain at each time moment are obtained, and the sliding window method is used, with the sliding window size set to... The sliding step size is 1. For a modeling road, the sliding step size is 1. The traffic value or occupancy value at time n is used as a training sample, and the traffic value or occupancy value at time n is used as a training sample. The flow rate or occupancy rate at any given time is used as the true value corresponding to that training sample, thereby generating the source domain training sample set and its corresponding true value set, including:

[0069] Centered on the source domain detector corresponding to the highest data modeling excellence, the flow rate and occupancy rate of each modeling road in the source domain at each time are obtained, resulting in the source domain flow time matrix and the source domain occupancy time matrix.

[0070] Based on the source domain traffic time matrix, and according to the sliding window size, prediction window size, and sliding step size, a source domain traffic training sample set and its corresponding ground truth set are generated, with the following expressions:

[0071]

[0072] ,

[0073]

[0074]

[0075] in, , These represent the source domain traffic training sample set and the corresponding real value set, respectively. Indicates the first The source domain traffic training sample matrix contains the first... Each sliding step corresponds to a modeling path within the source domain. The flow rate at any given moment; Indicates the first The true value vector corresponding to the nth source domain traffic training sample matrix contains the nth source domain traffic training sample matrix. The first sliding step corresponds to the first modeling path in the source domain. The actual value of the flow at any given moment; Indicates the number of modeled roads within the source domain; Indicates the total number of moments; Indicates the size of the sliding window; Indicates the prediction window size. ; Indicates the number of slides;

[0076] Based on the source domain occupancy time matrix, and according to the sliding window size, prediction window size, and sliding step size, a training sample set of source domain occupancy and its corresponding true value set are generated, with the following expressions:

[0077]

[0078] ,

[0079] ,

[0080]

[0081] in, , These represent the training sample set and the true value set corresponding to the source domain occupancy, respectively. Indicates the first The source domain occupancy training sample matrix contains the first source domain occupancy training sample matrix. Each sliding step corresponds to a modeling path in each source domain within the source domain. The occupancy value at any given moment; Indicates the first The true value vector corresponding to the source domain occupancy training sample matrix contains the first source domain occupancy training sample matrix. The first sliding step corresponds to the first modeling path in the source domain. The true value of time occupancy;

[0082] Based on the source domain traffic training sample set and its corresponding real value set, and the source domain occupancy training sample set and its corresponding real value set, a source domain training sample set and its corresponding real value set are constructed.

[0083] Preferably, filtering the target domain training sample set and its corresponding real value set includes:

[0084] By using the maximum weighted flow ratio and critical occupancy ratio between the source and target domains, the flow data and occupancy data in the target domain training sample set and its corresponding real value set are filtered to obtain the filtered target domain training sample set and its corresponding real value set.

[0085] The ratio of the maximum weighted traffic value between the source and target domains is used as the traffic filtering coefficient. Its expression is:

[0086]

[0087] in, This represents the maximum weighted flow value of the target domain; This represents the maximum weighted flow value of the source domain;

[0088] The critical occupancy ratio between the source and target domains is used as the occupancy filtering coefficient. Its expression is:

[0089]

[0090] in, This indicates the critical occupancy rate of the target domain; This represents the critical occupancy rate of the source domain.

[0091] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0092] (1) The road network-level traffic prediction method integrating transfer learning and macro-basic map described in this invention divides the current area into sub-regions and draws macro-basic maps of each sub-region; determines the target domain, and finds the sub-region with the most similar traffic pattern to the target domain as the source domain based on the macro-basic map of each sub-region; the source domain and the target domain found based on the macro-basic map have extremely high similarity in the changes of traffic flow data and occupancy data, laying the foundation for subsequent transfer learning; the traffic prediction model is trained using the traffic flow data and occupancy data of the source domain to obtain the source domain traffic prediction model, and then, based on the principle of transfer learning, the parts of the source domain traffic prediction model that are far from the input layer are frozen. After obtaining the parameters of multiple computational layers, the source domain traffic prediction model is adaptively trained using the traffic flow and occupancy data of the target domain to obtain the target domain traffic prediction model. This prediction process, which integrates transfer learning, can reduce the dependence on target domain data and the cost of training, while also improving the prediction effect on the target domain. That is, when the amount of target domain data is small or lacks data, the method designed in this invention can obtain high-precision prediction values ​​of target domain traffic flow and occupancy. In addition, the method designed in this invention enriches the practicality and applicability of the prediction model, can predict the traffic conditions of any area in advance, and can specifically solve the traffic congestion problems that may occur in various areas.

[0093] (2) The road network-level traffic prediction method that integrates transfer learning and macro-basic graph described in this invention finds the two detectors with the most similar traffic patterns in the target domain and the source domain through Pearson correlation coefficient and maximum mean difference during the data construction process. Based on the target domain detector and the source domain detector, data modeling is performed on the target domain and the source domain respectively to obtain the traffic flow time matrix sorted by the actual distance and time of the target domain detector and the traffic flow time matrix sorted by the actual distance and time of the source domain detector. These two traffic flow time matrices fully consider the temporal and spatial correlation between various links in the region and provide a dataset for subsequent training of the prediction model. This ensures the rationality of transfer learning between the target domain and the source domain.

[0094] (3) The road network-level traffic prediction method that integrates transfer learning and macro-basic graph described in this invention also designs a traffic prediction model coupled with multiple convolutional neural networks. This prediction model contains three convolutional neural networks: the first convolutional neural network is used to predict the flow rate; the second convolutional neural network is used to predict the occupancy rate; and the third convolutional neural network is used to predict the joint flow rate-occupancy rate. Compared with using only a single convolutional neural network to predict flow rate and occupancy rate, the prediction model designed in this invention can improve the prediction accuracy of flow rate and occupancy rate in the source domain and target domain. At the same time, after integrating transfer learning, it can further improve the prediction accuracy of flow rate and occupancy rate in the target domain, and can better predict the traffic conditions in any area, providing effective traffic flow prediction data for intelligent transportation systems. Attached Figure Description

[0095] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0096] Figure 1 This is a schematic diagram of a road network-level traffic prediction method that integrates transfer learning and macroscopic basic graphs, provided by the present invention.

[0097] Figure 2 It is a schematic diagram of the area of ​​the curve enclosed by the macroscopic basic map of each sub-region and the macroscopic basic map of the target domain;

[0098] Figure 3 This is a structural diagram of a traffic prediction model coupled with multiple convolutional neural networks;

[0099] Figure 4 This is a diagram of the structure of the third convolutional neural network;

[0100] Figure 5 This is a detailed framework diagram of the third convolutional neural network;

[0101] Figure 6 This is a schematic diagram of the model transfer process;

[0102] Figure 7 This is a schematic diagram of the study area; among which, Figure 7 (a) in the diagram represents a satellite schematic of the study area; Figure 7 (b) in the diagram represents the distribution of detectors in the study area;

[0103] Figure 8 This is a schematic diagram of the partitioning results of the study area and the MFD of each sub-region;

[0104] Figure 9 This is a comparison graph of the MFD and weighted flow of the source and destination domains; where, Figure 9(a) in the figure represents the MFD comparison diagram of the source domain and the target domain; Figure 9 (b) in the graph represents a weighted flow comparison diagram between the source and destination domains;

[0105] Figure 10 This is a comparison chart of MAE and MSE of traffic prediction values ​​and occupancy prediction values ​​based on source domain prediction datasets from different prediction models; among them, Figure 10 (a) in the figure represents a comparison of the MAE and MSE of the traffic forecasts from different forecasting models; Figure 10 (b) in the figure represents a comparison of the MAE and MSE of the market share predictions from different prediction models;

[0106] Figure 11 This is a trend comparison chart of predicted traffic flow and predicted occupancy rate with actual weighted traffic flow and occupancy rate from a multi-convolutional neural network coupled traffic prediction model.

[0107] Figure 12 This is a comparison chart of the prediction performance of the target domain model; among them, Figure 12 (a) in the figure represents a comparison of the MAE and MSE indices of various prediction models when predicting traffic flow on each road in the target domain; Figure 12 (b) in the figure represents a comparison of the NRMSE and SMAPE indices of various prediction models when predicting traffic flow on each road in the target domain; Figure 12 (c) in the figure represents a comparison of the MAE and MSE indices of various prediction models when predicting the occupancy of each road in the target domain; Figure 12 (d) in the figure represents a comparison of the NRMSE and SMAPE indices of various prediction models when predicting the occupancy of each road in the target domain.

[0108] Figure 13 This is a comparison chart of MFD-MATL predicted values ​​and measured values;

[0109] Figure 14 It is a traffic classification map of the target domain;

[0110] Figure 15 It is the MFD-MATL category prediction result in the target domain;

[0111] Figure 16 This is a schematic diagram of the macroeconomic fundamentals;

[0112] Figure 17 This is an overview diagram of transfer learning. Detailed Implementation

[0113] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0114] Example 1

[0115] Reference Figure 1 As shown, Figure 1 The flowchart of a road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs provided by this invention includes:

[0116] S1: Regional road network partitioning and selection of target and source domains, including:

[0117] S11: Based on the historical traffic data of the regional road network and the preset number of regional divisions, the regional road network is divided according to the semantic segmentation (N-Cut) algorithm to obtain the various sub-regions of the regional road network;

[0118] S12: Based on the macro-basic map of each sub-region, obtain the critical occupancy rate and maximum weighted flow value of each sub-region in the current region;

[0119] S13: Select any sub-region as the target domain, and take other sub-regions whose critical occupancy rate is within the critical interval range of the target domain as each critically related sub-region of the target domain. Within the critical interval range of the target domain, calculate the area of ​​the curve enclosed by the macro basic map of the target domain and the macro basic maps of each critically related sub-region, and take the critically related sub-region corresponding to the smallest curve area as the source domain; if there is no critical occupancy rate of other sub-regions within the critical interval range of the target domain, then re-divide the regional road network.

[0120] Among them, the macroscopic basic graph of the target domain and its first The area under the curve enclosed by the macroscopic basic plots of the critically correlated subregions is:

[0121]

[0122] in, A macroscopic basic diagram representing the target domain; The first term of the target domain One critically correlated sub-region; This indicates the critical occupancy rate of the target domain; Indicates the critical interval range of the target domain; Indicates the interval threshold;

[0123] by Figure 2 Taking the MFD of each sub-region as an example, with sub-region 2 as the target domain, the area of ​​the curve enclosed by sub-region 2 and sub-region 4 is... The expression is:

[0124]

[0125] in, This represents the macroscopic basic diagram of sub-region 4; This represents the macroscopic basic diagram of sub-region 2;

[0126] By calculating and filtering the critical occupancy difference and area difference of each sub-region, the sub-region with the smallest difference from the target domain MFD is obtained as the source domain for study, so as to ensure that the source domain and the target domain have similar traffic conditions.

[0127] S2: Data modeling, including:

[0128] When constructing model data, the temporal and spatial correlation between various links in the region should be fully considered. To this end, this invention takes the two detectors with the most similar traffic patterns in the target domain and the source domain as the center. Based on these two detectors, a traffic flow time matrix sorted by the actual distance and time of the detectors is established in their respective regions. Therefore, this invention finds the two detectors with the most similar traffic patterns in the target domain and the source domain through Pearson correlation coefficient and maximum mean difference.

[0129] S21: Taking each detector in the target domain as the center, obtain the flow rate values ​​of each modeled road in the target domain at each time, and establish the flow rate time matrix of each detector in the target domain; taking each detector in the source domain as the center, obtain the flow rate values ​​of each modeled road in the source domain at each time, and establish the flow rate time matrix of each detector in the source domain.

[0130] S22: Calculate the Pearson correlation coefficient between the flow-time matrix of each detector in the target domain and the flow-time matrix of each detector in the source domain, including:

[0131] For any detector X in the target domain and any detector Y in the source domain, calculate the Pearson correlation coefficient between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is:

[0132]

[0133] ,

[0134]

[0135] in, This represents the flow-time matrix of the target domain detector X; This represents the flow-time matrix of the source domain detector Y; This represents the covariance between the flow-time matrix of the target domain detector X and the flow-time matrix of the source domain detector Y; This represents the flow-time matrix of the target domain detector X in a sample size of... The standard deviation in the sample; Indicates that the target domain detector X is in the first... The flow time matrix at each moment; This represents the mean of the flow-time matrix of the target domain detector X; This represents the flow-time matrix of the source domain detector Y with a sample size of The standard deviation in the sample; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The flow time matrix at each moment; This represents the mean of the flow-time matrix of the source domain detector Y;

[0136] S23: Calculate the unbiased estimators between the flow-time matrices of each detector in the target domain and the flow-time matrices of each detector in the source domain, including:

[0137] For any detector X in the target domain and any detector Y in the source domain, calculate the maximum mean difference between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is:

[0138]

[0139] in, Indicates that the target domain detector X is in the first... The flow time matrix at each moment; This represents the number of samples in the flow-time matrix of the target domain detector X; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The flow time matrix at each moment; This represents the number of samples in the flow-time matrix of the source domain detector Y;

[0140] Through mapping function Project the flow time matrices of target domain detector X and source domain detector Y onto the kernel Hilbert space. In the process, the kernel function between the target domain detector X and the source domain detector Y is obtained, and its expression is:

[0141]

[0142] in, The target domain detector X represents the first... The mapping function corresponding to the flow time matrix at each time point; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The mapping function corresponding to the flow-time matrix at any given time; This corresponds to the inner product of the two domain features projected into the Hilbert space; by projecting the data into the kernel Hilbert space, the numerical difference between the two regions can be better measured.

[0143] By applying kernel tricks, the kernel function can be directly calculated. Based on the kernel function expression, the unbiased estimator between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain is calculated. Its expression is:

[0144]

[0145] By performing kernel operations using the Gaussian kernel function, the kernel function value between the target domain detector X and the source domain detector Y is obtained, thus yielding the unbiased estimator value between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain; where the Gaussian kernel function expression is:

[0146]

[0147] in, The parameter representing the width of the Gaussian kernel function;

[0148] Based on S23, in order to prevent excessive differences between regional traffic data determined solely by the Pearson correlation coefficient, this invention selects the Maximum Mean Difference (MMD) to measure the difference between the two regional data after modeling.

[0149] S24: Based on the Pearson correlation coefficient and unbiased estimator between the flow-time matrices of each detector in the target domain and the flow-time matrices of each detector in the source domain, the goodness of data modeling between the flow-time matrices of each detector in the target domain and the flow-time matrices of each detector in the source domain is obtained, including:

[0150] For any detector X in the target domain and any detector Y in the source domain, the Pearson correlation coefficient is used to determine the relationship between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. and unbiased estimators between them The goodness of data modeling between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain is obtained. Its expression is:

[0151]

[0152] in, Indicates the first weighting coefficient; Second weighting coefficient;

[0153] Taking into account both the regression prediction results of a single road and the differences between the modeling data of the two regions, the Pearson correlation coefficient was analyzed. and the corresponding data calculated after modeling Perform weighted calculations to obtain the goodness of data modeling. ;

[0154] Based on S21-S23, when performing similarity discrimination between two region detectors, the similarity between the two road detectors is calculated using kernel spatial distance and Pearson correlation coefficient;

[0155] S25: Using the target domain detector and source domain detector corresponding to the obtained maximum data modeling goodness as the center, obtain the flow rate and occupancy value of each modeling road in the target domain and source domain at each time step, and use the sliding window method, setting the sliding window size to... The sliding step size is 1. For a modeling road, the sliding step size is 1. The traffic value or occupancy value at time n is used as a training sample, and the traffic value or occupancy value at time n is used as a training sample. The traffic value or occupancy rate at any given time is used as the true value corresponding to the training sample, thereby generating the target domain training sample set and its corresponding true value set, and the source domain training sample set and its corresponding true value set, including:

[0156] (1) Generate the target domain training sample set and its corresponding real value set, including:

[0157] Centered on the target domain detector corresponding to the highest data modeling excellence, the flow rate and occupancy rate of each modeling road within the target domain at each time point are obtained, resulting in the target domain flow rate time matrix and the target domain occupancy rate time matrix. Each row in the target domain flow rate time matrix represents a time series of flow rates for a modeling road at each time point; each row in the target domain occupancy rate time matrix represents a time series of occupancy rates for a modeling road at each time point.

[0158] Based on the target domain traffic time matrix, and according to the sliding window size, prediction window size, and sliding step size, a target domain traffic training sample set and its corresponding ground truth set are generated, with the following expressions:

[0159]

[0160] ,

[0161] ,

[0162]

[0163] in, , These represent the target domain traffic training sample set and the corresponding real value set, respectively. Indicates the first The target domain traffic training sample matrix contains the first target domain traffic training sample matrix. Each sliding step corresponds to a modeling path within the target domain. The traffic flow value at time n, where each row represents the traffic flow time series of a modeled road within the target domain. Time to the The flow rate at any given moment; Indicates the first The true value vector corresponding to the nth target domain traffic training sample matrix contains the nth target domain traffic training sample matrix. The first sliding step corresponds to the first modeling road in the target domain. The actual value of the traffic flow at time t, where each row represents the traffic flow time series of a modeled road within the target domain. The flow rate at any given moment; Indicates time; Indicates the size of the sliding window; This represents the prediction window size, i.e., the number of true values ​​contained in each row of each true value vector. This invention sets... ; Indicates the number of slides;

[0164] Based on the target domain occupancy time matrix, and according to the sliding window size, prediction window size, and sliding step size, a training sample set and its corresponding true value set for the target domain occupancy are generated, with the following expressions:

[0165]

[0166] ,

[0167] ,

[0168]

[0169] in, , These represent the training sample set for target domain occupancy and the corresponding set of true values, respectively. Indicates the first The training sample matrix for the occupancy of the target domain contains the th target domain occupancy. Each sliding step corresponds to a modeling path within the target domain. The occupancy value at time n, where each row represents the occupancy value of a modeled road in the target domain at time n. Time to the Occupancy rate at any given moment; Indicates the first The true value vector corresponding to the training sample matrix of the target domain occupancy rate contains the first... The first sliding step corresponds to the first modeling road in the target domain. The true value of occupancy at time point, where each row represents the occupancy time series of a modeled road within the target domain. Occupancy rate at any given moment;

[0170] Based on the target domain traffic training sample set and its corresponding real value set, and the target domain occupancy training sample set and its corresponding real value set, construct the target domain training sample set and its corresponding real value set.

[0171] (2) Generate the source domain training sample set and its corresponding real value set, including:

[0172] Centered on the source domain detector corresponding to the highest data modeling excellence, the flow rate and occupancy rate of each modeling road within the source domain at each time point are obtained, resulting in the source domain flow rate time matrix and the source domain occupancy rate time matrix. Each row in the source domain flow rate time matrix represents a time series of flow rates for each modeling road at each time point; similarly, each row represents the flow rate time series of each modeling road within the source domain. Similarly, each row in the source domain occupancy rate time matrix represents a time series of occupancy rates for each modeling road at each time point; similarly, each row represents the occupancy rate time series of each modeling road within the source domain.

[0173] Based on the source domain traffic time matrix, and according to the sliding window size, prediction window size, and sliding step size, a source domain traffic training sample set and its corresponding ground truth set are generated, with the following expressions:

[0174]

[0175] ,

[0176] ,

[0177]

[0178] in, , These represent the source domain traffic training sample set and the corresponding real value set, respectively. Indicates the first The source domain traffic training sample matrix contains the first... Each sliding step corresponds to a modeling path within the source domain. The traffic flow value at time n, where each row represents the traffic flow time series of a modeled road within the source domain. Time to the The flow rate at any given moment; Indicates the first The true value vector corresponding to the nth source domain traffic training sample matrix contains the nth source domain traffic training sample matrix. The first sliding step corresponds to the first modeling path in the source domain. The actual value of the traffic flow at time step, where each row represents the traffic flow time series of a modeled road within the source domain at the specified time step. The flow rate at any given moment; Indicates the number of modeled roads within the source domain; Indicates the total number of moments; Indicates the size of the sliding window; Indicates the prediction window size. ; Indicates the number of slides;

[0179] Based on the source domain occupancy time matrix, and according to the preset input window size, prediction window size, and sliding step size, the source domain occupancy input set and source domain occupancy output set are generated, with the following expressions:

[0180]

[0181] ,

[0182] ,

[0183]

[0184] in, , These represent the training sample set and the true value set corresponding to the source domain occupancy, respectively. Indicates the first The source domain occupancy training sample matrix contains the first source domain occupancy training sample matrix. Each sliding step corresponds to a modeling path in each source domain within the source domain. The occupancy value at time n, where each row represents the occupancy value of a modeled road in the source domain at time n. Time to the Occupancy rate at any given moment; Indicates the first The true value vector corresponding to the source domain occupancy training sample matrix contains the first source domain occupancy training sample matrix. The first sliding step corresponds to the first modeling path in the source domain. The true value of occupancy at time point, where each row represents the occupancy time series of a modeled road within the source domain. Occupancy rate at any given moment;

[0185] Based on the source domain traffic training sample set and its corresponding real value set, and the source domain occupancy training sample set and its corresponding real value set, a source domain training sample set and its corresponding real value set are constructed.

[0186] S3: Establish a multi-convolutional neural network coupled traffic prediction model. The structure diagram of the multi-convolutional neural network coupled traffic prediction model is shown below. Figure 3 As shown; specifically including:

[0187] The multi-convolutional neural network coupled traffic prediction model includes three convolutional neural networks and two neural networks;

[0188] The source domain training set includes a traffic training sample set and a occupancy training sample set.

[0189] Each traffic training sample in the source domain training sample set is input into the first convolutional neural network, and the traffic prediction value corresponding to each traffic training sample is output.

[0190] The occupancy training samples in the source domain training sample set are input into the second convolutional neural network, and the occupancy prediction value corresponding to each occupancy training sample is output.

[0191] After the source domain training samples are normalized, each traffic training sample and each occupancy training sample are integrated to obtain each joint traffic-occupancy sample. The resulting joint traffic-occupancy sample is then input into the third convolutional neural network, which outputs the joint traffic-occupancy prediction value corresponding to each joint traffic-occupancy sample.

[0192] After standardizing the individual predicted values ​​of each flow, which are divided from the joint flow-occupancy predicted values ​​corresponding to each joint flow-occupancy sample, they are integrated with the flow predicted values ​​corresponding to each flow training sample, and then input into the first neural network to output the final predicted values ​​of the flow corresponding to each flow training sample.

[0193] After standardizing the individual predicted values ​​of each occupancy rate from the joint traffic-occupancy rate predicted values ​​corresponding to each joint traffic-occupancy rate sample, they are integrated with the occupancy rate predicted values ​​corresponding to each occupancy rate training sample, and then input into the second neural network to output the final predicted values ​​of the occupancy rate corresponding to each occupancy rate training sample.

[0194] In summary, the Multi-Convolutional Neural Network Coupled Traffic Prediction Model (M-CNN-FPM) contains three convolutional neural networks (CNNs) for different feature extraction. The first CNN extracts the spatiotemporal features of traffic flow, and the second CNN extracts the spatiotemporal features of occupancy. Both the first and second CNNs only consider the trend and features of one variable, without considering the direct real-time relationship between traffic flow and occupancy. Therefore, a third CNN is introduced to extract the relationship between the two. Unlike the first two CNNs, the input to the third CNN is the integrated value of traffic flow and occupancy, making its spatial dimension twice that of the first two CNNs.

[0195] The structure diagram of the third convolutional neural network is as follows: Figure 4 As shown, the traffic and occupancy data are normalized separately and then integrated. The resulting joint traffic-occupancy sample is then input into the third convolutional neural network to obtain the joint traffic-occupancy prediction value corresponding to the joint traffic-occupancy sample. Finally, the traffic data and occupancy data in the obtained joint traffic-occupancy prediction value are standardized to restore the data to its original magnitude.

[0196] After performing CNN predictions on traffic and occupancy separately, a set of predicted values ​​for traffic and occupancy can be obtained. Furthermore, a third convolutional neural network can produce another set of traffic and occupancy predictions. To integrate these two sets of predicted data, two additional neural networks are added after each of the three CNNs to combine the predicted values ​​for traffic and occupancy. By combining the individual and shared features of traffic and occupancy extracted by the first three CNNs, along with their connections, the accuracy of the predictions is further improved. The model framework of the third convolutional neural network is as follows: Figure 5 As shown.

[0197] S4: Using the source domain training sample set and its corresponding real value set, train the traffic prediction model to obtain the source domain traffic prediction model, including:

[0198] Based on the first convolutional neural network, the spatiotemporal features of the flow values ​​at each time step within each flow training sample in the source domain training sample set are extracted to obtain the flow prediction value corresponding to each flow training sample.

[0199] Based on the second convolutional neural network, the spatiotemporal features of the occupancy values ​​at each time step within each occupancy training sample in the source domain training sample set are extracted to obtain the occupancy prediction value corresponding to each occupancy training sample.

[0200] The source domain training samples are integrated and normalized to obtain the traffic training samples and the occupancy training samples to obtain the joint traffic-occupancy samples. The joint traffic-occupancy samples are then input into the third convolutional neural network to output the joint traffic-occupancy prediction value corresponding to each joint traffic-occupancy sample.

[0201] After standardizing the individual predicted values ​​of each flow, which are divided from the joint flow-occupancy predicted values ​​corresponding to each joint flow-occupancy sample, they are integrated with the flow predicted values ​​corresponding to each flow training sample, and then input into the first neural network to output the final predicted values ​​of the flow corresponding to each flow training sample.

[0202] After standardizing the individual predicted values ​​of each occupancy rate from the joint traffic-occupancy rate predicted values ​​corresponding to each joint traffic-occupancy rate sample, they are integrated with the occupancy rate predicted values ​​corresponding to each occupancy rate training sample, and then input into the second neural network to output the final predicted values ​​of the occupancy rate corresponding to each occupancy rate training sample.

[0203] Based on the final predicted values ​​of traffic flow corresponding to each traffic flow training sample, the final predicted values ​​of occupancy corresponding to each occupancy training sample, and the true value set corresponding to the source domain training sample set, the loss function is calculated. By minimizing the loss function, the parameters of the multi-convolutional neural network coupled traffic prediction model are obtained, and the training of the multi-convolutional neural network coupled traffic prediction model is completed. The trained multi-convolutional neural network coupled traffic prediction model is used as the source domain traffic prediction model.

[0204] S5: Using the filtered target domain training sample set and its corresponding true value set, adaptively train the source domain traffic prediction model to obtain the target domain traffic prediction model.

[0205] Because of the variability between networks in different traffic regions, after filtering out similar traffic regions, it is necessary to filter and process the data in the source or target domain to adapt to the data changes between different research areas. In this invention, the target domain data is filtered and processed by the critical occupancy rate between the target domain and the source domain and the proportional relationship between the maximum weighted traffic, so as to fit the prediction model trained from the source domain dataset.

[0206] Filtering the target domain training sample set and its corresponding ground truth set includes:

[0207] By using the maximum weighted flow ratio and critical occupancy ratio between the source and target domains, the flow data and occupancy data in the target domain training sample set and its corresponding real value set are filtered to obtain the filtered target domain training sample set and its corresponding real value set.

[0208] The ratio of the maximum weighted flow value between the source and target domains is used as the flow filtering coefficient. Its expression is:

[0209]

[0210] in, This represents the maximum weighted flow value of the target domain; This represents the maximum weighted flow value of the source domain;

[0211] The critical occupancy ratio between the source and target domains is used as the occupancy filtering coefficient. Its expression is:

[0212]

[0213] in, This indicates the critical occupancy rate of the target domain; This represents the critical occupancy rate of the source domain;

[0214] In summary, the source domain with the most similar traffic conditions to the target domain is selected through the macro fundamental map (MFD), and model transfer is carried out between these two subdomains. However, due to the differences in traffic flow between different regions, even the two subdomains that have been selected have differences of orders of magnitude. Therefore, this invention filters the target domain traffic volume data by the ratio of the maximum traffic flow values ​​between the source domain and the target domain, and performs the same operation on the occupancy data.

[0215] After filtering the traffic data in the target domain, the data between the two regions are made to be on the same order of magnitude. The M-CNN-FPM traffic prediction model is trained using the source domain training set data to obtain the source domain traffic prediction model. Then, several computational layer parameters of the source domain traffic prediction model are frozen, and a small portion of the target domain data (training set data) is input into the frozen source domain model for adaptive model training. The model trained in this way can capture the common features between the two traffic zones at the same time, and also improve the prediction performance of the traffic prediction model in the target domain.

[0216] Specifically, in the strategy of integrating transfer learning and macro-level fundamental graphs (MFD-MATL), three CNNs are treated as one module, and the subsequent two neural networks are also treated as one module. During model transfer, the CNN module closest to the input is frozen at the computational layer, while other layers and neural network modules are not frozen. The specific model transfer is as follows: Figure 8 As shown, the transfer method is used to transfer common features between two domains and adapt them to the target domain, thereby obtaining better prediction results in the target domain.

[0217] Example 2

[0218] To verify the performance of the proposed M-CNN-FPM prediction model, this invention used data obtained from UTD19 in central London, UK, with satellite views of the area as follows: Figure 7 As shown; the reason for choosing the traffic area in the city center for research is that the city center, as a compact area, has more similar road adjacencies, which is suitable for the transfer of the proposed model; after adjusting the remaining detectors and filtering out detectors with data anomalies, a selection of detectors were chosen for research, such as Figure 7 As shown, Figure 7 The red dots in the image represent the selected detectors.

[0219] The research data is generated based on a sliding window, with a preset sliding window size of 12, a prediction window size of 1, and a sliding step size of 1 (using data from the previous hour to predict traffic flow data for the next 5 minutes). To validate the model, the data generated by the sliding window is divided into two groups: a training set and a test set, with different proportions for the source and target domains. In the source domain, the first 90% of the traffic data is used for training, and the last 10% is used for prediction and evaluation. In the target domain, the first 10% of the traffic data from the source domain is used for adaptive training of the model, and the last 90% of the traffic data is used for prediction and evaluation of the model's predictive effectiveness. The traffic data includes flow rate data and occupancy data.

[0220] This invention, based on the N-cut clustering algorithm, considers a series of factors such as model feature extraction, internal homogeneity, model generalization, and computational cost, dividing the study area into five homogeneous sub-regions; the specific partitions and MFDs within each sub-region are as follows: Figure 8 As shown in Table 1, the MFD and parameters of each sub-region are recorded. By analyzing the critical occupancy rate of each sub-region, the two sub-regions with the highest similarity are Region3 and Region4.

[0221] Table 1 shows the partitioning results and the MFD and parameters of each subregion.

[0222]

[0223] In the MFD parameters shown in Table 1, Num_detector is the number of detectors in the region, MFD_a and MFD_b are the coefficients of the fitted quadratic function, and the expression for the maximum weighted flux of any sub-region is:

[0224]

[0225] in, This indicates the critical occupancy rate of the current sub-region; This represents the maximum weighted flow value corresponding to the current sub-region; This indicates the goodness of fit of the MFD curve fitted to the current sub-region to the whole. The larger the value, the more accurate the fit, and the better the MFD can represent the regional traffic pattern.

[0226] Region 3 was selected as the source region (source domain) and Region 4 as the target region (target domain); the scatter plot of the average occupancy weighted flow of the two regions and MFD is shown below. Figure 9 As shown in (a), it is clear that the two regions and their respective MFD shapes are more similar, but they are not on the same order of magnitude. Figure 9 (b) It shows the dynamic relationship between the weighted traffic data of the 643-step sequence of the target domain training set and the corresponding source domain data. It also clearly shows the similar traffic scenarios and numerical differences between the two. Figure 9 The similarity in traffic conditions between the two regions also proves that the transfer of the prediction model based on the minimum value in this invention is reasonable.

[0227] The M-CNN-FPM traffic prediction model represents a multi-convolutional neural network coupled traffic prediction model trained in the source domain. Two individual CNNs from the source domain (one predicting traffic flow and the other predicting occupancy) are selected for prediction and fused with transfer learning, named the Source-CNN-MATL traffic prediction model. That is, the Source-CNN-MATL traffic prediction model consists of two convolutional neural networks fused with transfer learning. Meanwhile, to further explore the advantages and disadvantages of direct model transfer based on feature space, the source domain data is filtered before training. The FSATL traffic prediction model corresponds to the model trained on the filtered source domain dataset, which is also based on a single CNN. MAE and MSE are selected to evaluate the predictive effectiveness of the two models.

[0228] Figure 10 The results of three traffic prediction models predicting traffic flow and occupancy on the source domain test dataset are shown. The prediction results of the Source-CNN-MATL traffic prediction model and the M-CNN-FPM traffic prediction model are better than those of the FSATL traffic prediction model, proving that filtering the source domain data does affect the prediction results of the source domain model. At the same time, it is shown that the proposed M-CNN-FPM traffic prediction model has better prediction performance.

[0229] Figure 11 The figure shows a comparison of the trends of the proposed M-CNN-FPM traffic prediction model with the actual weighted traffic flow and occupancy rate; Figure 11 The size and color of each point represent the difference between the predicted occupancy rate and the actual value in this invention; the results demonstrate that the traffic prediction model proposed in this invention can provide good predictions in the source domain.

[0230] After obtaining the source domain traffic prediction model, it is also necessary to perform domain adaptation on the source domain traffic prediction model in the target domain so that the prediction model can extract traffic knowledge and features from less traffic data in the target domain.

[0231] To verify the advantages of the MFD-MATL algorithm proposed in this invention in transfer learning, other modeling methods were added and compared. In addition to FSATL, Direct Transfer Learning (DTL) was also included, which directly applies the traffic prediction model trained in the source domain to the target domain data prediction after adaptation. A single CNN network trained with target domain data (Target-CNN) was also included for comparison. Among the various prediction methods studied, the CNN layer and parameter structure were identical. To further evaluate the model's superiority, in addition to the previously mentioned MAE and MSE, evaluation metrics such as NRMSE and SMAPE were introduced.

[0232] Figure 12 This paper demonstrates a comparison of the MAE, MSE, NRMSE, and SMAPE metrics of various prediction models when predicting traffic flow on each road in the target domain, even with insufficient training data. Through a comparison of the evaluation metrics under various models, the proposed MFD-MATL algorithm has predictive advantages and can provide more accurate traffic flow occupancy predictions for each link in the target domain.

[0233] Figure 13 The comparison between the actual values ​​and the MFD-MATL predicted values ​​in the target domain test set shows that the difference between the two is very small, indicating that accurate predictions can be made for most traffic conditions, and that the predicted data points can reflect most traffic conditions in the target domain.

[0234] To further compare the accuracy of the proposed model in predicting traffic conditions, the traffic conditions in the target domain were divided into three categories based on the K-means algorithm: free flow, steady flow, and unstable flow. The partitioning results and the prediction results of the proposed model are shown below. Figure 14 and Figure 15 As shown, the results indicate that the model proposed in this invention can provide relatively accurate traffic condition predictions even when target domain data is insufficient.

[0235] Table 2 compares the prediction performance of various models in the target domain with insufficient data; the first row shows traffic flow prediction, and the second row shows occupancy prediction. Source-CNN-MATL consists of two independent CNNs in the source domain, predicting traffic flow and occupancy respectively. Its prediction framework in the target domain is the same as MFD-MATL, i.e., it only filters the target domain data. For further comparison, this invention also uses Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) for traffic flow prediction in the target domain. Accuracy represents the prediction result after classifying the traffic conditions of the target domain cells. Comparison of various metrics shows that the MFD-MATL proposed in this paper has good prediction performance in cells with insufficient data.

[0236] Table 2 Comparison of Evaluation Indicators

[0237]

[0238] Based on Embodiment 1 and Embodiment 2, the road network-level traffic prediction method that integrates transfer learning and macroscopic basic graphs provided by this invention is generally divided into two parts.

[0239] In the first part, this invention divides the network into several regions based on the N-cut algorithm, draws the MFD scatter plot of each region, and then regresses to obtain the MFD curve and its key parameters for each region. In order to filter the traffic regions most suitable for transfer learning, this invention proposes a two-step traffic region similarity discrimination method based on the critical MFD occupancy rate and the area between curves, and finds the sub-region most similar to the macroscopic basic graph of the target domain as the source domain. The traffic prediction model M-CNN-FPM (Multi-CNNs-Fusion Prediction Model) coupled with the multi-convolutional neural network proposed in this invention is trained with traffic data from the source domain, and some layers of the obtained prediction model are frozen to prepare for subsequent transfer learning. The frozen layers of the obtained prediction model are several consecutive computational layers close to the input.

[0240] In the second part, the traffic flow occupancy data of the target domain is filtered based on the ratio of the critical occupancy to the maximum weighted flow value obtained from the MFD of the two regions; a small amount of data from the target domain is input to adapt the source domain model to the target domain, so that the model can obtain better prediction results in the regression task of the target domain; finally, the adjusted prediction model is used to predict the test data of the target domain and a rationality analysis is performed.

[0241] This invention utilizes macroscopic fundamental graphs, transfer learning, and deep learning, the specific principles of which are as follows:

[0242] (1) The Macroscopic Fundamental Diagram (MFD) was first proposed by Godfrey in 1969. It reflects the unimodal relationship between network weighted flow and average density (occupancy). Since the MFD is independent of the traffic network and does not change with the traffic flow of the road network, it can measure the traffic status of the road network and is an intrinsic property of the network. According to the MFD, the unimodal relationship between network weighted flow and average density (occupancy) can be described as a quadratic or cubic function relationship.

[0243] Taking a quadratic function relationship as an example, the basic curve of MFD is as follows: Figure 1 As shown; in Figure 1 middle, This represents the average occupancy rate of the road network; while This represents the weighted flow value of the road network; This represents the critical occupancy rate of the road network, at which point the corresponding weighted traffic flow value is the maximum weighted traffic flow value. The corresponding area's road network achieves maximum traffic efficiency; when the average road network occupancy is less than When the average occupancy rate of the road network is greater, the weighted traffic flow is greater; when the average occupancy rate of the road network is greater than... At that time, the higher the average occupancy of the road network, the lower its weighted traffic flow; by controlling the number of vehicles in the road network, the average occupancy can be controlled within a certain range. In the vicinity, the road network's traffic efficiency can be maintained at a high level, thereby alleviating regional congestion. This includes the weighted flow rate of the road network. The expressions for average market share are as follows:

[0244]

[0245]

[0246] in, , , They represent roads in the road network. The weighted traffic volume, occupancy rate, and road length.

[0247] (2) Transfer learning is a powerful machine learning method. Its core idea is to transfer knowledge or models learned on one task (source task) to another related task (target task) to improve the performance of the new task. This method makes full use of existing data and model resources, reduces the target task's dependence on a large amount of new data, speeds up model training, and improves the model's generalization ability; the overview diagram of transfer learning is shown below. Figure 2 As shown;

[0248] Transfer learning is mainly divided into three categories:

[0249] Model-based transfer learning: directly using the pre-trained model of the source task as the starting point for the target task, and fine-tuning the model parameters to adapt to the new task; when there is little traffic data in the target domain, using the pre-trained model can significantly improve model performance.

[0250] Feature-based transfer: extracting feature representations from the source task and then training a model for the target task on these features; specifically, transferring features between the source and target domains through feature transformation to reduce the gap between them; or transforming the data features of the source and target domains into a unified feature space and then using traditional machine learning methods for classification and recognition.

[0251] Relationship-based transfer: learn the relationships between data from the source task and apply these relationships to the target task; in this invention, by learning traffic data in the source domain and their interrelationships, and then applying them to the data-scarce target domain, the dependence on the target domain data is low.

[0252] (3) Deep learning is a machine learning algorithm based on artificial neural networks. Its core lies in building multi-layered neural network models to achieve automatic learning and feature extraction of complex data. This algorithm imitates the interaction between neurons in the human brain and uses a large amount of data for training to identify patterns, classify and predict.

[0253] In deep learning, a Convolutional Neural Network (CNN) is a type of feedforward neural network widely used in computer vision and image recognition tasks. The design of CNNs is inspired by research into biological visual cognitive mechanisms, particularly the layout of local connections between neurons in the visual cortex. It automatically extracts features by performing a series of convolutional operations on the input image, and these features can be used as the basis for various tasks such as image classification, object detection, and face recognition.

[0254] CNNs are mainly composed of the following layers:

[0255] Convolutional Layer: This is the core component of a CNN. In a convolutional layer, one or more convolutional kernels (also called filters) slide across the input image, generating an output feature map through dot product operations.

[0256] Activation Layer: Activation layers typically follow convolutional layers and are used to increase the non-linearity of the network.

[0257] Pooling Layer: Pooling layers (also known as downsampling layers) are used to reduce the spatial size of data, thereby reducing the number of parameters and computational cost in the network, and also helping to reduce overfitting.

[0258] Fully Connected Layer: After several convolutional and pooling layers, one or more fully connected layers are typically followed. Each neuron in a fully connected layer is connected to all neurons in the previous layer. Its function is to combine the features extracted by the convolutional and pooling layers for the final classification or regression task.

[0259] Output Layer: The output layer is usually a special case of a fully connected layer, used to output the final result.

[0260] The advantage of CNNs lies in their ability to automatically extract useful features from raw images without requiring the manual design of complex feature extractors. This makes CNNs promising for applications in fields such as image recognition, video analysis, and autonomous driving.

[0261] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs, characterized in that, include: Based on historical traffic data of the regional road network and the preset number of regional divisions, the regional road network is divided using a semantic segmentation algorithm to obtain the various sub-regions of the regional road network. Based on the macro-basic map of each sub-region, obtain the critical occupancy rate of each sub-region; select one sub-region as the target domain, and take other sub-regions whose critical occupancy rate is within the critical interval of the target domain as each critically related sub-region of the target domain. Within the critical interval of the target domain, calculate the area of ​​the curve enclosed by the macro-basic map of the target domain and the macro-basic maps of each critically related sub-region, and take the critically related sub-region corresponding to the smallest curve area as the source domain; if there is no critical occupancy rate of other sub-regions within the critical interval of the target domain, then re-divide the regional road network. Taking each detector in the target domain as the center, obtain the flow rate values ​​of each modeled road in the target domain at each time moment, and establish the flow rate time matrix of each detector in the target domain; taking each detector in the source domain as the center, obtain the flow rate values ​​of each modeled road in the source domain at each time moment, and establish the flow rate time matrix of each detector in the source domain. Based on the Pearson correlation coefficient and unbiased estimator between the flow-time matrices of each detector in the target domain and the flow-time matrices of each detector in the source domain, the goodness of data modeling between the flow-time matrices of each detector in the target domain and the flow-time matrices of each detector in the source domain is obtained. Centered on the target domain detector and source domain detector corresponding to the highest obtained data modeling goodness, respectively, the flow rate and occupancy values ​​of each modeled road in the target domain and source domain at each time step are obtained. Then, using the sliding window method, the sliding window size is set to... The sliding step size is 1. For a modeling road, the sliding step size is 1. The traffic value or occupancy value at time n is used as a training sample, and the traffic value or occupancy value at time n is used as a training sample. The flow value or occupancy rate at any given time is used as the true value corresponding to the training sample, thereby generating the target domain training sample set and its corresponding true value set, and the source domain training sample set and its corresponding true value set. The traffic prediction model is trained using the source domain training sample set and its corresponding real value set to obtain the source domain traffic prediction model. Specifically, it includes: Based on the first convolutional neural network, the spatiotemporal features of the flow values ​​at each time step within each flow training sample in the source domain training sample set are extracted to obtain the flow prediction value corresponding to each flow training sample. Based on the second convolutional neural network, the spatiotemporal features of the occupancy values ​​at each time step within each occupancy training sample in the source domain training sample set are extracted to obtain the occupancy prediction value corresponding to each occupancy training sample. The source domain training samples are integrated and normalized to obtain the traffic training samples and the occupancy training samples to obtain the joint traffic-occupancy samples. The joint traffic-occupancy samples are then input into the third convolutional neural network to output the joint traffic-occupancy prediction value corresponding to each joint traffic-occupancy sample. After standardizing the individual predicted values ​​of each flow, which are divided from the joint flow-occupancy predicted values ​​corresponding to each joint flow-occupancy sample, they are integrated with the flow predicted values ​​corresponding to each flow training sample, and then input into the first neural network to output the final predicted values ​​of the flow corresponding to each flow training sample. After standardizing the individual predicted values ​​of each occupancy rate from the joint traffic-occupancy rate predicted values ​​corresponding to each joint traffic-occupancy rate sample, they are integrated with the occupancy rate predicted values ​​corresponding to each occupancy rate training sample, and then input into the second neural network to output the final predicted values ​​of the occupancy rate corresponding to each occupancy rate training sample. Based on the final predicted values ​​of traffic flow corresponding to each traffic flow training sample, the final predicted values ​​of occupancy corresponding to each occupancy training sample, and the true value set corresponding to the source domain training sample set, the loss function is calculated; by minimizing the loss function, the parameters of the multi-convolutional neural network coupled traffic prediction model are obtained, and the training of the multi-convolutional neural network coupled traffic prediction model is completed; the trained multi-convolutional neural network coupled traffic prediction model is used as the source domain traffic prediction model. By using the filtered target domain training sample set and its corresponding true value set, the source domain traffic prediction model is adaptively trained to obtain the target domain traffic prediction model.

2. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs as described in claim 1, characterized in that, The traffic prediction model is a multi-convolutional neural network coupled traffic prediction model; The multi-convolutional neural network coupled traffic prediction model includes three convolutional neural networks and two neural networks; The source domain training sample set includes a traffic training sample set and a occupancy training sample set. Each traffic training sample in the source domain training sample set is input into the first convolutional neural network, and the traffic prediction value corresponding to each traffic training sample is output. The occupancy training samples in the source domain training sample set are input into the second convolutional neural network, and the occupancy prediction value corresponding to each occupancy training sample is output. After the source domain training samples are normalized, each traffic training sample and each occupancy training sample are integrated to obtain each joint traffic-occupancy sample. The resulting joint traffic-occupancy sample is then input into the third convolutional neural network, which outputs the joint traffic-occupancy prediction value corresponding to each joint traffic-occupancy sample. After standardizing the individual predicted values ​​of each flow, which are divided from the joint flow-occupancy predicted values ​​corresponding to each joint flow-occupancy sample, they are integrated with the flow predicted values ​​corresponding to each flow training sample, and then input into the first neural network to output the final predicted values ​​of the flow corresponding to each flow training sample. After standardizing the individual predicted values ​​of each occupancy rate from the joint traffic-occupancy rate predicted values ​​corresponding to each joint traffic-occupancy rate sample, the predicted values ​​of each occupancy rate training sample are integrated with the predicted values ​​of each occupancy rate training sample, and then input into the second neural network to output the final predicted values ​​of the occupancy rate corresponding to each occupancy rate training sample.

3. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs as described in claim 1, characterized in that, The macroscopic basic graph of the target domain and its first The area under the curve enclosed by the macroscopic basic plots of the critically correlated subregions is: ， in, A macroscopic basic diagram representing the target domain; The first term of the target domain One critically correlated sub-region; This indicates the critical occupancy rate of the target domain; Indicates the critical interval range of the target domain; This represents the interval threshold.

4. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs as described in claim 1, characterized in that, The calculation of the Pearson correlation coefficient between the flow-time matrix of each detector in the target domain and the flow-time matrix of each detector in the source domain includes: For any detector X in the target domain and any detector Y in the source domain, calculate the Pearson correlation coefficient between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is: ， ， ， in, This represents the flow-time matrix of the target domain detector X; This represents the flow-time matrix of the source domain detector Y; This represents the covariance between the flow-time matrix of the target domain detector X and the flow-time matrix of the source domain detector Y; This represents the flow-time matrix of the target domain detector X in a sample size of... The standard deviation in the sample; Indicates that the target domain detector X is in the first... The flow-time matrix at any given moment; This represents the mean of the flow-time matrix of the target domain detector X; This represents the flow-time matrix of the source domain detector Y with a sample size of The standard deviation in the sample; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The flow-time matrix at any given moment; This represents the mean of the flow time matrix of the source domain detector Y.

5. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs according to claim 1, characterized in that, The calculation of the unbiased estimator between the flow-time matrix of each detector in the target domain and the flow-time matrix of each detector in the source domain includes: For any detector X in the target domain and any detector Y in the source domain, calculate the maximum mean difference between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is: ， in, Indicates that the target domain detector X is in the first... The flow-time matrix at any given moment; This represents the number of samples in the flow-time matrix of the target domain detector X; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The flow-time matrix at any given moment; This represents the number of samples in the flow-time matrix of the source domain detector Y; Through mapping function Project the flow time matrices of target domain detector X and source domain detector Y onto the kernel Hilbert space. In the process, the kernel function between the target domain detector X and the source domain detector Y is obtained, and its expression is: ， in, The target domain detector X represents the first... The mapping function corresponding to the flow-time matrix at any given time; The flow-time matrix of the source domain detector Y represents the flow-time matrix in the th... The mapping function corresponding to the flow-time matrix at any given time; Based on the kernel function expression, an unbiased estimator is calculated between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. Its expression is: ， By performing kernel operations using the Gaussian kernel function, the kernel function value between the target domain detector X and the source domain detector Y is obtained, thus yielding the unbiased estimator value between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain; where the Gaussian kernel function expression is: ， in, This represents the width parameter of the Gaussian kernel function.

6. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs according to claim 1, characterized in that, The data modeling goodness of the flow-time matrices of each detector in the target domain and each detector in the source domain, based on the Pearson correlation coefficient and unbiased estimator, includes: For any detector X in the target domain and any detector Y in the source domain, the Pearson correlation coefficient is used to determine the relationship between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain. and the unbiased estimators between them The goodness of data modeling between the flow-time matrix of detector X in the target domain and the flow-time matrix of detector Y in the source domain is obtained. Its expression is: ， in, Indicates the first weighting coefficient; Second weighting coefficient.

7. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs as described in claim 1, characterized in that, Centering on the target domain detector corresponding to the highest data modeling goodness, we obtain the traffic flow and occupancy values ​​of each modeled road within the target domain at each time step, and then use the sliding window method, setting the sliding window size to... The sliding step size is 1. For a modeling road, the sliding step size is 1. The traffic value or occupancy value at time n is used as a training sample, and the traffic value or occupancy value at time n is used as a training sample. The flow rate or occupancy rate at any given time is used as the true value corresponding to that training sample, thereby generating the target domain training sample set and its corresponding true value set, including: Centered on the target domain detector corresponding to the highest data modeling excellence, the flow rate and occupancy rate of each modeling road in the target domain at each time are obtained, resulting in the target domain flow rate time matrix and the target domain occupancy rate time matrix. Based on the target domain traffic time matrix, and according to the sliding window size, prediction window size, and sliding step size, a target domain traffic training sample set and its corresponding ground truth set are generated, with the following expressions: ， ， ， ， in, , These represent the target domain traffic training sample set and the corresponding real value set, respectively. Indicates the first The target domain traffic training sample matrix contains the first target domain traffic training sample matrix. Each sliding step corresponds to a modeling path within the target domain. The flow rate at any given moment; Indicates the first The true value vector corresponding to the nth target domain traffic training sample matrix contains the nth target domain traffic training sample matrix. The first sliding step corresponds to the first modeling road in the target domain. The actual value of the flow at any given moment; Indicates the number of modeled roads within the target domain; Indicates the total number of moments; Indicates the size of the sliding window; Indicates the prediction window size. ; Indicates the number of slides; Based on the target domain occupancy time matrix, and according to the sliding window size, prediction window size, and sliding step size, a training sample set and its corresponding true value set for the target domain occupancy are generated, with the following expressions: ， ， ， ， in, , These represent the training sample set for target domain occupancy and the corresponding set of true values, respectively. Indicates the first The training sample matrix for the occupancy of the target domain contains the th target domain occupancy. Each sliding step corresponds to a modeling path within the target domain. The occupancy value at any given moment; Indicates the first The true value vector corresponding to the occupancy training sample matrix contains the th The first sliding step corresponds to the first modeling road in the target domain. The true value of time occupancy; Based on the target domain traffic training sample set and its corresponding real value set, and the target domain occupancy training sample set and its corresponding real value set, construct the target domain training sample set and its corresponding real value set.

8. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs according to claim 1, characterized in that, Centering on the source domain detector corresponding to the highest obtained data modeling goodness, the flow rate and occupancy values ​​of each modeled road within the source domain at each time step are obtained. Then, using the sliding window method, the sliding window size is set to... The sliding step size is 1. For a modeling road, the sliding step size is 1. The traffic value or occupancy value at time n is used as a training sample, and the traffic value or occupancy value at time n is used as a training sample. The flow rate or occupancy rate at any given time is used as the true value corresponding to that training sample, thereby generating the source domain training sample set and its corresponding true value set, including: Centered on the source domain detector corresponding to the highest data modeling excellence, the flow rate and occupancy rate of each modeling road in the source domain at each time are obtained, resulting in the source domain flow time matrix and the source domain occupancy time matrix. Based on the source domain traffic time matrix, and according to the sliding window size, prediction window size, and sliding step size, a source domain traffic training sample set and its corresponding ground truth set are generated, with the following expressions: ， ， ， ， in, , These represent the source domain traffic training sample set and the corresponding real value set, respectively. Indicates the first The source domain traffic training sample matrix contains the first... Each sliding step corresponds to a modeling path within the source domain. The flow rate at any given moment; Indicates the first The true value vector corresponding to the nth source domain traffic training sample matrix contains the nth source domain traffic training sample matrix. The first sliding step corresponds to the first modeling path in the source domain. The actual value of the flow at any given moment; This indicates the number of modeled roads within the source domain; Indicates the total number of moments; Indicates the size of the sliding window; Indicates the prediction window size. ; Indicates the number of slides; Based on the source domain occupancy time matrix, and according to the sliding window size, prediction window size, and sliding step size, a training sample set of source domain occupancy and its corresponding true value set are generated, with the following expressions: ， ， ， ， in, , These represent the training sample set and the true value set corresponding to the source domain occupancy, respectively. Indicates the first The source domain occupancy training sample matrix contains the first source domain occupancy training sample matrix. Each sliding step corresponds to a modeling path in each source domain within the source domain. The occupancy value at any given moment; Indicates the first The true value vector corresponding to the source domain occupancy training sample matrix contains the first source domain occupancy training sample matrix. The first sliding step corresponds to the first modeling path in the source domain. The true value of time occupancy; Based on the source domain traffic training sample set and its corresponding real value set, and the source domain occupancy training sample set and its corresponding real value set, a source domain training sample set and its corresponding real value set are constructed.

9. The road network-level traffic prediction method integrating transfer learning and macroscopic basic graphs according to claim 1, characterized in that, Filtering the training sample set of the target domain and its corresponding real value set includes: By using the maximum weighted flow ratio and critical occupancy ratio between the source and target domains, the flow data and occupancy data in the target domain training sample set and its corresponding real value set are filtered to obtain the filtered target domain training sample set and its corresponding real value set. The ratio of the maximum weighted traffic value between the source and target domains is used as the traffic filtering coefficient. Its expression is: ， in, This represents the maximum weighted flow value of the target domain; This represents the maximum weighted flow value of the source domain; The critical occupancy ratio between the source and target domains is used as the occupancy filtering coefficient. Its expression is: ， in, This indicates the critical occupancy rate of the target domain; This represents the critical occupancy rate of the source domain.

Images