Traffic jam distribution state prediction method and system based on networked vehicle GPS track data
By constructing a three-dimensional tensor matrix and a CNN-GRU model, the problem of insufficient modeling of nonlinear spatiotemporal relationships in connected vehicle GPS trajectory data is solved, enabling accurate prediction of traffic congestion under extreme weather conditions and improving computational efficiency and prediction accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies are insufficient in modeling nonlinear spatiotemporal relationships, have low computational efficiency, and are particularly poor in terms of accuracy and robustness in traffic flow prediction under extreme weather conditions.
A traffic congestion distribution prediction method based on connected vehicle GPS trajectory data is adopted. By constructing a three-dimensional tensor matrix, the model is optimized using truncated high-order singular value decomposition and gradient descent algorithms. Combined with the CNN-GRU prediction model, traffic speed tensor completion and congestion state prediction are performed.
It improves the accuracy and computational efficiency of traffic congestion prediction under different weather conditions, making it suitable for large-scale industrial applications.
Smart Images
Figure CN121789446A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent traffic control technology, specifically relating to a method and system for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles. Background Technology
[0002] With the advancement of the new urbanization strategy, my country's transportation system is facing a transformation from scale expansion to quality improvement. Current research mainly focuses on three key areas: traffic data completion, traffic flow prediction, and congestion status identification. Regarding data completion, existing methods can be divided into three categories: spatiotemporal interpolation-based methods, statistical learning-based methods, and deep learning-based methods. In the field of traffic flow prediction, although deep learning models such as LSTM and Transformer have demonstrated superior performance, they still lack in-depth modeling of the dynamic coupling mechanism between meteorological factors (such as rainfall intensity and visibility) and traffic flow parameters. While research on congestion state identification has developed various technical approaches based on cluster analysis, probabilistic reasoning, and deep learning, there is still significant room for improvement in the accuracy and robustness of identification under extreme weather conditions, particularly in handling abnormal congestion patterns caused by severe weather.
[0003] In summary, although these methods in the existing technology have made some progress, they still face challenges such as insufficient modeling of nonlinear spatiotemporal relationships and low computational efficiency. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles, so as to solve the problems of insufficient nonlinear spatiotemporal relationship modeling and low computational efficiency in existing technologies.
[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for predicting traffic congestion distribution based on connected vehicle GPS trajectory data includes the following steps: Step 1: Collect multiple sets of raw GPS trajectory data from connected vehicles to form a raw trajectory dataset; preprocess the raw trajectory dataset to obtain a preprocessed dataset. Step 2: Construct a three-dimensional tensor matrix from the preprocessed dataset, identify the location of the original missing traffic data in the three-dimensional tensor matrix and fill it in, and set the tensor used for filling in the missing data as the optimal initial tensor. Step 3: Use the truncated higher-order singular value decomposition method to extract the three-dimensional principal component features of the optimal initial tensor, and construct the core tensor based on the three-dimensional principal component features; Construct a tensor completion model and use the gradient descent algorithm to iteratively optimize the parameters to obtain a tensor completion optimized model. Step 4: Use the tensor completion optimization model to perform traffic speed tensor completion on each set of connected vehicle GPS raw trajectory data in the preprocessed dataset to obtain the training set; Step 5: Build a CNN-GRU prediction model. Train the model using the dataset obtained in Step 4 to obtain the trained CNN-GRU prediction model. Step 6: Use the trained CNN-GRU prediction model to predict the distribution of traffic congestion.
[0006] The present invention also has the following features: Furthermore, step 1 includes the following sub-steps: Step 11: Collect multiple sets of raw GPS trajectory data from connected vehicles to form a raw trajectory dataset; perform abnormal data cleaning on the raw trajectory dataset, specifically including: S101. Delete duplicate data; S102. Delete data where signal interference causes abnormal positioning points; S103. Delete data where the speed exceeds the threshold; S104. Delete data where the position speed is continuously zero; S105. Delete data with consecutive missing values. Step 12: For the original trajectory dataset after anomaly data cleanup in Step 11, perform geographic coordinate transformation on each set of connected vehicle GPS original trajectory data, converting the GCJ-02 coordinate system to the WGS84 coordinate system, as shown in the following formula:
[0007]
[0008]
[0009]
[0010]
[0011] in, , These represent the latitude and longitude coordinates of any set of raw GPS trajectory data for connected vehicles in the GCJ-02 coordinate system; , These represent the latitude and longitude coordinates of any set of raw GPS trajectory data for connected vehicles in the WGS-84 coordinate system; , These represent the random deviation values of latitude and longitude in different coordinate systems; , express , Transition value; , ; Step 13: For each set of raw GPS trajectory data of connected vehicles that underwent geographic coordinate transformation in Step 12, perform time format conversion to form a preprocessed dataset; the time format conversion is as follows:
[0012] in, Represents a timestamp; The time format is YYYY / M / D HH:MM:SS.
[0013] Furthermore, step 2 includes the following sub-steps: Step 21, set the road detection equipment to operate at fixed time intervals. t Traffic flow data was collected, and the number of data points collected per day was: The number of road sections is I ; The preprocessed dataset is constructed into a three-dimensional tensor matrix. M Specifically, it is expressed as:
[0014] Step 22, using non-negative weight tensors Identify the locations of missing original traffic data, where the non-negative weight tensor As shown in the following formula:
[0015] in, ; Step 23: Introduce spatiotemporal regularization constraints into the standard Tucker decomposition model, use a low-rank completion model to determine the tensor element used to fill in the original missing traffic data, use this tensor element to fill in the original missing traffic data, and designate this tensor element as the optimal initialization tensor, denoted as . ; Step 24: Use the truncated higher-order singular value decomposition method to extract the three-dimensional principal component features of the optimal initial tensor, specifically including temporal principal component features, spatial principal component features, and flow principal component features; Calculate the initial feature matrix of the three-dimensional principal component features , and .
[0016] Step 25: Construct the graph Laplacian optimization strategy, specifically by defining a graph Laplacian matrix. , ; in, It is an adjacency matrix. It is a diagonal matrix; The core tensor is calculated using a space regularization constraint model, as shown in the following equation:
[0017] in These are three-dimensional tensor matrices. M The factor matrix derived from decomposition; For the core tensor; * indicates element-wise product; It is a regularization term; For regularization parameters; Step 26: Introduce a time-dimension regularization term as a constraint, and perform iterative parameter optimization using the gradient descent algorithm to obtain the trained tensor completion optimization model.
[0018] Furthermore, in step 26, the time-dimension regularization term is specifically represented as follows: ;in, It is a time constraint matrix with a Toplitz structure, specifically as follows: .
[0019] Furthermore, the tensor completion optimization model in step 26 is as follows:
[0020] in, A factor matrix representing the daily time intervals; Regularization term It is a time interval constraint within one day; It is the regularization parameter.
[0021] Furthermore, the CNN-GRU prediction model described in step 5 includes a 1D-CNN layer, a GRU layer, and a fully connected layer; The 1D-CNN layer is used to extract spatial features between adjacent road segments in the original GPS trajectory data of connected vehicles, and then generate a feature matrix; The GRU layer is used for time-series modeling, capturing the dynamic evolution of traffic states through a gating mechanism, and outputting high-level features. The fully connected layer is used to output traffic congestion distribution predictions based on high-level features.
[0022] Furthermore, the hyperparameters are specifically selected during the training of the CNN-GRU prediction model as follows: Window size = 18, batch size = 18, learning rate = 0.001, training epochs = 300.
[0023] A traffic congestion distribution prediction module based on GPS trajectory data of connected vehicles, specifically including the following modules: Data processing module: used to receive raw trajectory datasets formed by multiple sets of GPS raw trajectory data of connected vehicles; to preprocess the raw trajectory datasets to obtain preprocessed datasets; The 3D tensor matrix construction and filling module is used to construct a 3D tensor matrix from the preprocessed dataset, identify the location of the original missing traffic data in the 3D tensor matrix and fill it in, and set the tensor used for filling as the optimal initial tensor; Tensor completion optimization model construction module: used to extract the three-dimensional principal component features of the optimal initial tensor using the truncated higher-order singular value decomposition method, and construct the core tensor based on the three-dimensional principal component features; Construct a tensor completion model and use the gradient descent algorithm to iteratively optimize the parameters to obtain a tensor completion optimized model. Traffic speed tensor completion module: Used to perform traffic speed tensor completion on each set of connected vehicle GPS raw trajectory data in the preprocessed dataset using the tensor completion optimization model to obtain the training set; CNN-GRU Prediction Model Training Module: Used to build a CNN-GRU prediction model. The model is trained using the dataset obtained from the Traffic Speed Tensor Completion Module to obtain the trained CNN-GRU prediction model. Prediction module: Used to predict the distribution of traffic congestion using the trained CNN-GRU prediction model.
[0024] Compared with the prior art, the present invention has the following technical effects: The traffic congestion distribution prediction method based on connected vehicle GPS trajectory data of the present invention exhibits significant temporal heterogeneity and pattern differentiation characteristics in the temporal distribution pattern of traffic congestion based on connected vehicle GPS trajectory data. Specifically, it manifests as differentiated response mechanisms on weekdays and weekends, as well as the lag effect of rainfall. Rainfall reconstructs the spatiotemporal distribution pattern of traffic congestion by reducing road traffic efficiency, changing driving behavior, and delaying the recovery process of traffic flow. By jointly modeling the spatiotemporal correlation between traffic speed and rainfall data, the method improves performance efficiency while achieving accurate prediction of road network congestion under different weather conditions, making it suitable for large-scale industrial use and promotion. Attached Figure Description
[0025] Figure 1(a) is a schematic diagram of the state before map matching in one embodiment of the present invention; Figure 1(b) is a schematic diagram of the map matching state in one embodiment of the present invention; Figure 2(a) is a comparison of speeds on sunny and rainy days during a weekday in one embodiment of the present invention; Figure 2(b) is a comparison chart of weekday traffic congestion perception in one embodiment of the present invention; Figure 3(a) is a comparison of speeds on sunny and rainy days during a rest day in one embodiment of the present invention; Figure 3(b) is a comparison of traffic congestion perception on sunny and rainy days during a rest day in one embodiment of the present invention; Figure 4(a) is a diagram of road network congestion distribution on a sunny weekday according to one embodiment of the present invention; Figure 4(b) is a diagram showing the road network congestion distribution on a weekday during rainy weather, according to one embodiment of the present invention. Figure 5(a) is a diagram showing the road network congestion distribution on a sunny rest day according to one embodiment of the present invention; Figure 5(b) is a diagram showing the road network congestion distribution on a rainy day during a rest day, according to one embodiment of the present invention. Figure 6(a) is a diagram of the traffic congestion perception and prediction results on a weekday rainy day in one embodiment of the present invention; Figure 6(b) is a diagram showing the traffic congestion perception and prediction results on a rainy day during a rest day in one embodiment of the present invention; Detailed Implementation The following are specific embodiments of the present invention. It should be noted that the present invention is not limited to the following specific embodiments. All equivalent modifications made based on the technical solutions of this application fall within the protection scope of the present invention.
[0026] A method for predicting traffic congestion distribution based on connected vehicle GPS trajectory data includes the following steps: Step 1: Collect multiple sets of raw GPS trajectory data of connected vehicles, preprocess each set of GPS trajectory data of connected vehicles, and obtain the preprocessed dataset. Step 2: Construct a three-dimensional tensor matrix from the preprocessed dataset, identify the location of the original missing traffic data in the three-dimensional tensor matrix, fill in the original missing traffic data, and set the tensor used for filling in the missing data as the optimal initial tensor. The truncated higher-order singular value decomposition method is used to extract the three-dimensional principal component features of the optimal initial tensor, and the core tensor is constructed based on the three-dimensional principal component features. A tensor completion model is constructed by introducing regularization and time regularization terms. The parameters are iteratively optimized using the gradient descent algorithm to train the tensor completion model and obtain the optimized tensor completion model. Step 4: Use the tensor completion optimization model to perform traffic speed tensor completion on each set of connected vehicle GPS trajectory data in the preprocessed dataset to obtain the training set; Step 5: Build a CNN-GRU prediction model. Train the model using the dataset obtained in Step 4 to obtain the trained CNN-GRU prediction model. Step 6: Use the trained CNN-GRU prediction model to predict the distribution of traffic congestion.
[0027] Furthermore, step 1 includes the following sub-steps: Step 11: Collect multiple sets of raw GPS trajectory data for connected vehicles, and perform abnormal data cleanup on each set of raw GPS trajectory data for connected vehicles. Specifically, this includes: S101. Delete duplicate data; S102. Delete data where signal interference causes abnormal positioning points; S103. Delete data where the speed exceeds the threshold; S104. Delete data where the position speed is continuously zero; S105. Delete data with consecutive missing values. Multiple sets of raw GPS trajectory data of connected vehicles after abnormal data cleanup were used as the cleaned dataset. Regarding geographic coordinate system transformation, this invention relates to two coordinate systems: the GCJ-02 coordinate system (Chinese encryption standard) used in taxi GPS data and the WGS84 coordinate system (international standard) used by OpenStreetMap. Due to the nonlinear offset between the two coordinate systems, this study utilizes a specialized transformation algorithm to convert GCJ-02 coordinates to WGS84 coordinates, ensuring that trajectory data and road network information are matched and analyzed under a unified reference system. This transformation process accurately handles coordinate offset characteristics, providing a reliable spatial foundation for subsequent research.
[0028] Step 12: Perform geographic coordinate transformation on each group of raw GPS trajectory data of connected vehicles in the cleaned dataset, converting the GCJ-02 coordinate system to the WGS84 coordinate system, as shown in the following formula:
[0029]
[0030]
[0031]
[0032]
[0033] in, , These represent the latitude and longitude coordinates of any set of raw GPS trajectory data for connected vehicles in the GCJ-02 coordinate system; , These represent the latitude and longitude coordinates of any set of raw GPS trajectory data for connected vehicles in the WGS-84 coordinate system; , These represent the random deviation values of latitude and longitude in different coordinate systems; , express , Transition value; , ; Regarding time format conversion, the time information in the GPS track data underwent standardization. Time anomalies in the original data (such as incorrect years or time fields containing letters) were rigorously removed. The original time format was "YYYY / M / D HH:MM:SS", which has been uniformly converted to the Unix timestamp format (cumulative seconds since January 1, 1970, 00:00:00 UTC) for easier calculation and analysis. This conversion ensures the standardization of time data and cross-system compatibility, providing a unified time benchmark for subsequent data processing and cluster analysis.
[0034] Step 13: Perform time format conversion on each group of raw GPS trajectory data of connected vehicles in the cleaned dataset to obtain preprocessed GPS trajectory data of connected vehicles, forming a preprocessed dataset; the time format conversion is as follows:
[0035] in, Represents a timestamp; The time format is YYYY / M / D HH:MM:SS; Complete the preprocessing of all connected vehicle GPS raw trajectory data to obtain the preprocessed dataset.
[0036] Furthermore, this embodiment constructs a spatiotemporally regularized Tucker decomposition model to achieve optimal initialization decomposition by fusing spatiotemporal feature constraints and truncating low-rank tensors for completion, and optimizes the core tensor rank selection based on dynamic thresholds. Traffic data is constructed as a three-dimensional tensor (time × space × velocity). Specifically, step 2 includes the following sub-steps: Step 21, assume the road detection equipment collects traffic flow data at fixed time intervals t, and the number of data points collected per day is... The number of road segments is I; In this embodiment, the road detection equipment is set to operate at fixed time intervals. Traffic flow data is collected every 5 minutes; the corresponding number of data points collected per day. for ; For traffic data in a third-order tensor, we have:
[0037] The preprocessed dataset is constructed into a 3D tensor matrix X. Specifically, the 3D tensor matrix X is:
[0038] The set of observation locations is The corresponding missing region is denoted as .
[0039] Traffic observation data constructed using tensors Estimate missing data values .
[0040] Step 22, using non-negative weight tensors Identify the location of the original missing traffic data using a non-negative weighted tensor. As shown in the following formula:
[0041] in, ; Step 23: Introduce spatiotemporal regularization constraints into the standard Tucker decomposition model, use a low-rank completion model to determine tensor elements, use this tensor to fill in the original missing traffic data, and designate this tensor as the optimal initialization tensor, denoted as . ; Step 24: Use the truncated higher-order singular value decomposition method to extract the three-dimensional principal component features of the optimal initial tensor, specifically including temporal principal component features, spatial principal component features, and flow principal component features; Calculate the initial feature matrix of the three-dimensional principal component features , and .
[0042] Step 25: Construct a graph Laplacian optimization strategy, specifically by defining a graph Laplacian matrix. , ; in, It is an adjacency matrix. It is a diagonal matrix; The space regularization constraint model is as follows:
[0043] in It is a factor matrix; Specifically, This matrix combines all dimensions (J, K) except for the spatial dimension (I) into a single column. Each row of this matrix represents the complete traffic pattern of a road segment across all days and all time points. This merges all dimensions (K, I) except for the time dimension (J). Each row represents the traffic flow pattern across all road segments and all days at a given time point. This merges all dimensions (I, J) except for the number of days (K). Each row represents the overall traffic pattern for all road segments and all time points within a day.
[0044] For the core tensor; * indicates element-wise product.
[0045] It is a space-related regularization term; For regularization parameters; Step 26: Introduce a time-dimension regularization term as a constraint to establish a tensor completion optimization model.
[0046] Furthermore, in step 26, the time dimension regularization term is specifically represented as follows: ;in, It is a time constraint matrix with a Toplitz structure, as shown below: .
[0047] Furthermore, the tensor completion optimization model in step 26 is as follows:
[0048] in, A factor matrix representing the daily time intervals; Regularization term It is a time interval constraint within one day; It is the regularization parameter.
[0049] in , , It is a factor matrix; This is the core tensor; * denotes element-wise multiplication. Matrix It is a space-dependent factor matrix. It is a space-related regularization term. This is the regularization parameter.
[0050] The tensor completion optimization model is represented as follows:
[0051] in, Factor matrix corresponding to daily time intervals; regularization term It is a time interval constraint within one day; It is the regularization parameter.
[0052] Specifically, the final optimization objective function is:
[0053] in, For regularization terms; This is the regularization parameter.
[0054] The gradient descent method is used to find the optimal solution of the objective function, through calculation. For core tensor and each factor matrix The partial derivatives are used to obtain the gradient of the objective function, assuming... , Then its expression is:
[0055] function For decision variables The partial derivatives are used to obtain the decision variables:
[0056]
[0057]
[0058]
[0059] in is the learning rate for the gradient descent algorithm.
[0060] Furthermore, the CNN-GRU prediction model in step 5 includes 1D-CNN layers, GRU layers, and fully connected layers; The 1D-CNN layer is used to extract spatial features between adjacent road segments, and then generate a feature matrix; The GRU layer performs time-series modeling, captures the dynamic evolution of traffic states through a gating mechanism, and outputs high-level features; The fully connected layer outputs a traffic congestion distribution prediction based on the characteristics of higher layers.
[0061] Furthermore, the hyperparameters used during the training of the CNN-GRU prediction model are as follows: Window size = 18, batch size = 18, learning rate = 0.001, training epochs = 300; To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0062] The CNN-GRU prediction model proposed in this embodiment combines convolutional neural networks and gated recurrent units for feature extraction and prediction of traffic flow data. To enhance the accuracy of the model's predictions, hyperparameters are selected through experimental evaluation. For example, the window size when constructing the dataset using a sliding window is selected. Experiments were conducted using 6, 12, 18, 24, and 30 windows. According to the evaluation metrics, the model loss is minimized when the window size is 18. Therefore, this embodiment selects 18 as the window size for constructing the dataset. Similarly, this invention experimentally selected hyperparameters such as batch size (Batch_Size) = 18, learning rate (Learning_rate) = 0.001, and training epochs (Epochs) = 300.
[0063] Furthermore, the selection of evaluation indicators: To measure the predictive performance of the model, this embodiment uses three commonly used metrics: mean squared error (MSE), mean absolute percentage error (MAPE), and root mean square error (RMSE). MAE measures the magnitude of the average error; MAPE evaluates the error as a percentage and is suitable for comparing data of different magnitudes; RMSE is more sensitive to larger errors and emphasizes prediction accuracy.
[0064] MSE is used to measure the mean squared error between predicted and actual values. The calculation formula is:
[0065] Please refer to Figure 1. The map matching method based on Hidden Markov Models (HMM) effectively solves the problem of deviation between GPS trajectory data and the actual road network. Before matching, the trajectory points show a significant deviation from the road network, as shown in Figure 1(a). After processing by the HMM algorithm, as shown in Figure 1(b), the trajectory points are accurately projected onto the actual roads, eliminating unreasonable phenomena such as crossing buildings. This method significantly improves the spatial accuracy of trajectory data, providing a reliable foundation for subsequent traffic analysis.
[0066] Please refer to Figure 2-3. The temporal distribution pattern of traffic congestion based on GPS trajectory data of connected vehicles shows significant time-period heterogeneity and pattern differentiation. Specifically, it is manifested in the differentiated response mechanism between weekdays and rest days and the lag effect of rainfall. Rainfall reconstructs the spatiotemporal distribution pattern of traffic congestion by reducing road traffic efficiency, changing driving behavior and delaying the recovery process of traffic flow.
[0067] Please refer to Figure 4-5, which shows the distribution of road network congestion on rest days and weekdays under sunny and rainy conditions. The spatial distribution pattern of traffic congestion shows that rainfall reduces the overall capacity of the road network and decreases traffic efficiency. Based on 24-hour road segment congestion index analysis and visualized using ArcGIS, there are significant differences in the spatial distribution of congestion between weekdays and rest days, reflecting the interactive characteristics of traffic demand and the impact of rainfall under different date types.
[0068] Rainfall significantly alters the spatial pattern of urban road network congestion. Comparative analysis reveals that under sunny conditions, the road network operates relatively smoothly overall, with only localized areas experiencing mild to moderate congestion. However, rainy weather exhibits a clear characteristic of congestion spreading, manifested in three aspects: first, the congestion area expands, with a significant increase in the proportion of moderate to severe congestion sections; second, intersection efficiency decreases, with significantly longer queues and waiting times; and third, congestion on main roads and expressways intensifies.
[0069] Further in-depth analysis of peak hours reveals that weekday road networks experience an overall worsening of congestion under the influence of rainfall. By constructing a peak-hour road congestion index assessment model, the study found that rainfall leads to significantly different spatial distribution patterns of traffic pressure during morning and evening peak hours: morning peak congestion is mainly concentrated on main roads in the urban core area, while evening peak congestion is more characterized by persistent congestion on radial roads.
[0070] Please refer to Figure 6, which systematically evaluates the predictive performance of the CNN-GRU model. Two sets of experiments were set up for comparison: validation of the effectiveness of meteorological factors (CNN-GRU with / without considering meteorological factors) and comparison of model architectures (CNN, GRU, RF). The CNN-GRU incorporating meteorological features performed best in all scenarios, especially showing significant improvement in rainy weather scenarios (RMSE, MSE, and MAPE were optimized by 9.3%, 25.0%, and 12.2%, respectively, in the S2 scenario). In the most challenging S2 scenario, this model reduced the RMSE by 52.1%, 65.8%, and 74.1% compared to CNN, GRU, and RF alone, respectively.
[0071] Models that take meteorology into account have a higher degree of fit to the actual values and can more accurately capture congestion fluctuations during periods of rainfall.
[0072] This embodiment presents a traffic congestion distribution prediction system based on the aforementioned method for predicting traffic congestion distribution using connected vehicle GPS trajectory data. Specifically, it includes: Data processing module: used to receive raw trajectory datasets formed by multiple sets of GPS raw trajectory data of connected vehicles; to preprocess the raw trajectory datasets to obtain preprocessed datasets; The 3D tensor matrix construction and filling module is used to construct a 3D tensor matrix from the preprocessed dataset, identify the location of the original missing traffic data in the 3D tensor matrix and fill it in, and set the tensor used for filling as the optimal initial tensor; Tensor completion optimization model construction module: used to extract the three-dimensional principal component features of the optimal initial tensor using the truncated higher-order singular value decomposition method, and construct the core tensor based on the three-dimensional principal component features; Construct a tensor completion model and use the gradient descent algorithm to iteratively optimize the parameters to obtain a tensor completion optimized model. Traffic speed tensor completion module: Used to perform traffic speed tensor completion on each set of connected vehicle GPS raw trajectory data in the preprocessed dataset using the tensor completion optimization model to obtain the training set; CNN-GRU Prediction Model Training Module: Used to build a CNN-GRU prediction model. The model is trained using the dataset obtained from the Traffic Speed Tensor Completion Module to obtain the trained CNN-GRU prediction model. Prediction module: Used to predict the distribution of traffic congestion using the trained CNN-GRU prediction model.
Claims
1. A method for predicting traffic congestion distribution based on connected vehicle GPS trajectory data, characterized in that, Includes the following steps: Step 1: Collect multiple sets of raw GPS trajectory data from connected vehicles to form a raw trajectory dataset; preprocess the raw trajectory dataset to obtain a preprocessed dataset. Step 2: Construct a three-dimensional tensor matrix from the preprocessed dataset, identify the location of the original missing traffic data in the three-dimensional tensor matrix and fill it in, and set the tensor used for filling in the missing data as the optimal initial tensor. Step 3: Use the truncated higher-order singular value decomposition method to extract the three-dimensional principal component features of the optimal initial tensor, and construct the core tensor based on the three-dimensional principal component features; Construct a tensor completion model and use the gradient descent algorithm to iteratively optimize the parameters to obtain a tensor completion optimized model. Step 4: Use the tensor completion optimization model to perform traffic speed tensor completion on each set of connected vehicle GPS raw trajectory data in the preprocessed dataset to obtain the training set; Step 5: Build a CNN-GRU prediction model. Train the model using the dataset obtained in Step 4 to obtain the trained CNN-GRU prediction model. Step 6: Use the trained CNN-GRU prediction model to predict the distribution of traffic congestion.
2. The method for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles as described in claim 1, characterized in that, Step 1 includes the following sub-steps: Step 11: Collect multiple sets of raw GPS trajectory data from connected vehicles to form a raw trajectory dataset; perform abnormal data cleaning on the raw trajectory dataset, specifically including: S101. Delete duplicate data; S102. Delete data where signal interference causes abnormal positioning points; S103. Delete data where the speed exceeds the threshold; S104. Delete data where the position speed is continuously zero; S105. Delete data with consecutive missing values. Step 12: For the original trajectory dataset after anomaly data cleanup in Step 11, perform geographic coordinate transformation on each set of connected vehicle GPS original trajectory data, converting the GCJ-02 coordinate system to the WGS84 coordinate system, as shown in the following formula: in, , These represent the latitude and longitude coordinates of any set of raw GPS trajectory data for connected vehicles in the GCJ-02 coordinate system; , These represent the latitude and longitude coordinates of any set of raw GPS trajectory data for connected vehicles in the WGS-84 coordinate system; , These represent the random deviation values of latitude and longitude in different coordinate systems; , express , Transition value; ， ； Step 13: For each set of raw GPS trajectory data of connected vehicles that underwent geographic coordinate transformation in Step 12, perform time format conversion to form a preprocessed dataset; the time format conversion is as follows: in, Represents a timestamp; The time format is YYYY / M / D HH:MM:SS.
3. The method for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles as described in claim 2, characterized in that, Step 2 It includes the following steps: Step 21, set the road detection equipment to operate at fixed time intervals. t Traffic flow data was collected, and the number of data points collected per day was: The number of road sections is I ; The preprocessed dataset is constructed into a three-dimensional tensor matrix. M Specifically, it is expressed as: Step 22, using non-negative weight tensors Identify the locations of missing original traffic data, where the non-negative weight tensor As shown in the following formula: in, ; Step 23: Introduce spatiotemporal regularization constraints into the standard Tucker decomposition model, use a low-rank completion model to determine the tensor element used to fill in the original missing traffic data, use this tensor element to fill in the original missing traffic data, and designate this tensor element as the optimal initialization tensor, denoted as . ; Step 24: Use the truncated higher-order singular value decomposition method to extract the three-dimensional principal component features of the optimal initial tensor, specifically including temporal principal component features, spatial principal component features, and flow principal component features; Calculate the initial feature matrix of the three-dimensional principal component features , and . Step 25: Construct the graph Laplacian optimization strategy, specifically by defining a graph Laplacian matrix. , ; in, It is an adjacency matrix. It is a diagonal matrix; The core tensor is calculated using a space regularization constraint model, as shown in the following equation: in These are three-dimensional tensor matrices. M The factor matrix derived from decomposition; For the core tensor; * indicates element-wise product; It is a regularization term; For regularization parameters; Step 26: Introduce a time-dimension regularization term as a constraint, and perform iterative parameter optimization using the gradient descent algorithm to obtain the trained tensor completion optimization model.
4. The method for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles as described in claim 3, characterized in that, In step 26, the time dimension regularization term is specifically represented as follows: ;in, It is a time constraint matrix with a Toplitz structure, specifically as follows: 。 5. The method for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles as described in claim 4, characterized in that, The tensor completion optimization model in step 26 is as follows: in, A factor matrix representing the daily time intervals; Regularization term It is a time interval constraint within one day; It is the regularization parameter.
6. The method for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles as described in claim 5, characterized in that, The CNN-GRU prediction model described in step 5 includes 1D-CNN layers, GRU layers, and fully connected layers; The 1D-CNN layer is used to extract spatial features between adjacent road segments in the original GPS trajectory data of connected vehicles, and then generate a feature matrix; The GRU layer is used for time-series modeling, capturing the dynamic evolution of traffic states through a gating mechanism, and outputting high-level features. The fully connected layer is used to output traffic congestion distribution predictions based on high-level features.
7. The method for predicting traffic congestion distribution based on GPS trajectory data of connected vehicles as described in claim 6, characterized in that, The hyperparameters for training the CNN-GRU prediction model are selected as follows: Window size = 18, batch size = 18, learning rate = 0.001, training epochs = 300.
8. A traffic congestion distribution prediction module based on GPS trajectory data of connected vehicles, characterized in that, Specifically, it includes the following modules: Data processing module: used to receive raw trajectory datasets formed by multiple sets of GPS raw trajectory data of connected vehicles; to preprocess the raw trajectory datasets to obtain preprocessed datasets; The 3D tensor matrix construction and filling module is used to construct a 3D tensor matrix from the preprocessed dataset, identify the location of the original missing traffic data in the 3D tensor matrix and fill it in, and set the tensor used for filling as the optimal initial tensor; Tensor completion optimization model construction module: used to extract the three-dimensional principal component features of the optimal initial tensor using the truncated higher-order singular value decomposition method, and construct the core tensor based on the three-dimensional principal component features; Construct a tensor completion model and use the gradient descent algorithm to iteratively optimize the parameters to obtain a tensor completion optimized model. Traffic speed tensor completion module: Used to perform traffic speed tensor completion on each set of connected vehicle GPS raw trajectory data in the preprocessed dataset using the tensor completion optimization model to obtain the training set; CNN-GRU Prediction Model Training Module: Used to build a CNN-GRU prediction model. The model is trained using the dataset obtained from the Traffic Speed Tensor Completion Module to obtain the trained CNN-GRU prediction model. Prediction module: Used to predict the distribution of traffic congestion using the trained CNN-GRU prediction model.