Traffic congestion prediction control method based on quality perception and multi-source data fusion

Abstract

The present application relates to the technical field of intelligent transportation system, and particularly relates to a traffic congestion prediction control method based on quality perception and multi-source data fusion. The method comprises the following steps: performing data enhancement on a plurality of traffic data streams according to a pre-trained generative adversarial network to obtain a plurality of enhanced data streams; performing feature fusion on the plurality of enhanced data streams according to the sensor capability of a traffic sensor corresponding to each enhanced data stream to obtain a target feature vector; processing the target feature vector according to a pre-trained spatio-temporal graph neural network to obtain congestion prediction information; and processing the congestion prediction information and the control demand of a user according to a pre-trained hybrid control model to obtain traffic control information. The present application can improve the prediction accuracy of a traffic state prediction scheme.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent transportation systems, specifically to a traffic congestion prediction and control method based on quality perception and multi-source data fusion. Background Technology

[0002] Urban traffic congestion is a significant challenge hindering urban operational efficiency and sustainable development. With the large-scale deployment of intelligent transportation infrastructure, the integrated application of traffic sensing data has become a crucial technological approach for urban traffic congestion management. Currently, mainstream solutions primarily rely on multi-source sensor data, including geomagnetic coils, video detectors, floating car GPS, and mobile signaling, to predict traffic conditions using deep learning methods such as spatiotemporal graph neural networks, and then optimize signal timing or provide route guidance accordingly.

[0003] However, existing technologies are mostly based on the assumptions of stable data quality and complete model input, which still have significant shortcomings in the complex traffic environment of real cities, resulting in the signal timing optimization or path guidance work carried out based on them not achieving the expected results. Summary of the Invention

[0004] The purpose of this invention is to provide a traffic congestion prediction and control method based on quality perception and multi-source data fusion, which can solve the technical problem of low prediction accuracy in existing traffic condition prediction schemes.

[0005] In a first aspect, one embodiment of the present invention provides a traffic congestion prediction and control method based on quality perception and multi-source data fusion, the method comprising: Data augmentation is performed on multiple traffic data streams using a pre-trained generative adversarial network to obtain multiple augmented data streams. The generative adversarial network uses the data quality of the multiple training data streams as training constraints. The multiple traffic data streams correspond one-to-one with multiple traffic sensors in the target road network. The data quality includes accuracy indicators, completeness indicators, timeliness indicators, and reliability indicators. Based on the sensor capabilities of the traffic sensor corresponding to each enhanced data stream, feature fusion is performed on multiple enhanced data streams to obtain a target feature vector. The sensor capabilities include spatial coverage indicators, temporal accuracy indicators, historical reliability indicators, and equipment operation indicators. The sensor capabilities and the fusion weights of the corresponding enhanced data streams during feature fusion show a positive correlation. The target feature vector is processed by a pre-trained spatiotemporal graph neural network to obtain congestion prediction information, wherein the congestion prediction information is used to indicate the queue length of each road segment in the target road network in a future time period. The congestion prediction information and user control requirements are processed by a pre-trained hybrid control model to obtain traffic control information. The hybrid control model includes a model predictive control module and a deep reinforcement learning module. The traffic control information is used to indicate control strategies that match the control requirements.

[0006] In some embodiments, the accuracy metric is determined based on the signal-to-noise ratio and historical prediction error of the corresponding data stream; the integrity metric is determined based on the actual number of received data points, the expected number of received data points, and the maximum duration of consecutive missing data in the corresponding data stream; the timeliness metric is determined based on the data latency of the corresponding data stream and the standard deviation of the time interval between adjacent data packets; and the reliability metric is determined based on the device failure rate and the ratio of abnormal data in the corresponding data stream.

[0007] In some embodiments, the loss function value corresponding to the network parameters of the generative adversarial network is obtained by weighted calculation of multiple data stream loss values. The multiple data stream loss values ​​correspond one-to-one with multiple training data streams, and the calculation weight of the data stream loss value is positively correlated with the data quality of the corresponding training data stream.

[0008] In some embodiments, the spatial coverage index is determined based on the degree of overlap between the sensing range of the corresponding traffic sensor and the target road network, the temporal accuracy index is determined based on the sampling frequency and communication delay of the corresponding traffic sensor, the historical reliability index is determined based on the historical prediction error of the corresponding traffic sensor, and the equipment operation index is determined based on the equipment failure rate and abnormal data ratio of the corresponding traffic sensor.

[0009] In some embodiments, the matrix element values ​​in the edge weight matrix corresponding to the spatiotemporal graph neural network are determined by the prior historical association strength and the real-time association strength between two corresponding road segments in the target road network. The real-time association strength is used to indicate the similarity of the feature components of the corresponding two road segments in the target feature vector.

[0010] In some embodiments, the graph convolution module corresponding to the spatiotemporal graph neural network includes a quality mask matrix corresponding to the target feature vector. The quality mask matrix is ​​used to indicate the sensor capability of at least one traffic sensor associated with each road segment in the target road network.

[0011] In some embodiments, the congestion prediction information is obtained by weighting historical regression information and network prediction information, wherein the historical regression information is used to indicate the queue length of each road segment in the target road network in the historical reference period associated with the future time period, and the network prediction information is used to indicate the queue length of each road segment in the target road network in the future time period predicted by the spatiotemporal graph neural network.

[0012] In some embodiments, the method further includes: Based on the network prediction information, gradient-weighted class activation mapping and attention weight backtracking are performed to obtain a congestion propagation heatmap, wherein the congestion propagation heatmap is used to indicate the contribution of each road segment in the target road network to the queuing length of other road segments in future time periods.

[0013] In some embodiments, the traffic control information is obtained by fusing first information output by the model prediction control module and second information output by the deep reinforcement learning module.

[0014] In some embodiments, the sum of the fusion weights of the first information and the second information is 1, and the fusion weight of the second information is positively correlated with the prediction confidence corresponding to the congestion prediction information.

[0015] Secondly, another embodiment of the present invention provides a traffic congestion prediction and control system based on quality perception and multi-source data fusion, the system comprising: The data augmentation module is used to augment multiple traffic data streams based on a pre-trained generative adversarial network to obtain multiple augmented data streams. The generative adversarial network uses the data quality of multiple training data streams as training constraints. The multiple traffic data streams correspond one-to-one with multiple traffic sensors in the target road network. The data quality includes accuracy indicators, completeness indicators, timeliness indicators, and reliability indicators. The feature fusion module is used to perform feature fusion on multiple enhanced data streams based on the sensor capabilities of the traffic sensor corresponding to each enhanced data stream, and obtain a target feature vector. The sensor capabilities include spatial coverage indicators, temporal accuracy indicators, historical reliability indicators, and equipment operation indicators. The sensor capabilities and the fusion weights of the corresponding enhanced data streams during feature fusion are positively correlated. The congestion prediction module is used to process the target feature vector according to the pre-trained spatiotemporal graph neural network to obtain congestion prediction information, wherein the congestion prediction information is used to indicate the queue length of each road segment in the target road network in the future time period. The traffic control module is used to process the congestion prediction information and the user's control needs according to a pre-trained hybrid control model to obtain traffic control information. The hybrid control model includes a model prediction control module and a deep reinforcement learning module. The traffic control information is used to indicate a control strategy that matches the control needs.

[0016] Thirdly, in another embodiment of the present invention, an electronic device is provided, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method described in the first aspect.

[0017] Fourthly, in another embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0018] The present invention has the following beneficial effects: This invention first uses a generative adversarial network to augment multiple traffic data streams, enhancing some low-quality traffic data streams in real-world traffic scenarios and ensuring the reliability of data processed in subsequent processes. Then, based on the sensor capabilities of each traffic sensor, feature fusion is performed on the augmented traffic data streams to amplify the feature influence of data streams from high-capacity sensors, improving the reliability of the fused target feature vector. Next, a spatiotemporal graph neural network is used to process the target feature vector to obtain congestion prediction information indicating the queue length of each road segment in the target road network in future time periods. Finally, a hybrid control model is used to process the congestion prediction information and control requirements. The model prediction control module in the hybrid control model ensures the application security of the output results, while the deep reinforcement learning module ensures the output efficiency and accuracy, ultimately improving the accuracy of traffic state prediction results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a traffic congestion prediction and control method based on quality perception and multi-source data fusion provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a traffic congestion prediction and control system based on quality perception and multi-source data fusion provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a traffic congestion prediction and control method based on quality perception and multi-source data fusion proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] The following description, in conjunction with the accompanying drawings, details a specific scheme for a traffic congestion prediction and control method based on quality perception and multi-source data fusion provided by this invention.

[0024] Analysis reveals that existing solutions for integrating traffic perception data largely rely on the assumptions of stable data quality and complete model input, which still have significant shortcomings in the complex traffic environments of real cities: First, data fusion lacks a quality perception mechanism, failing to uniformly quantify and assess uncertainties such as time asynchrony, spatial registration deviation, communication delay, and sensor performance degradation, resulting in a lack of robustness in downstream prediction and control modules when data quality fluctuates. Second, prediction models lack quality adaptive capabilities, unable to dynamically adjust information aggregation weights based on data credibility, making it difficult to guarantee prediction stability in scenarios with missing data or noise interference. Third, control strategies are decoupled from prediction uncertainty, failing to transform data quality scores into dynamic safety boundaries in control decisions, lacking a safety guarantee mechanism when perception credibility declines. Fourth, the "perception-prediction-control" chain is fragmented. Existing frameworks generally design data fusion, state prediction, and control decisions separately, lacking a closed-loop mechanism for continuously optimizing perception and prediction modules based on control effect feedback, and the prediction decision-making process lacks interpretability, making it difficult to trace the causes and propagation paths of congestion, and failing to provide reliable data for traffic managers.

[0025] Therefore, there is an urgent need for a unified method that can achieve multi-source data quality assessment, adaptive fusion and weighting, and directly apply data quality and predictive uncertainty to control decisions, so as to build an intelligent decision-making framework of "quality perception-dynamic prediction-optimized control" to improve the reliability, accuracy and timeliness of urban traffic congestion management.

[0026] To address the aforementioned problems, in one embodiment, the present invention provides a traffic congestion prediction and control method based on quality perception and multi-source data fusion, such as... Figure 1 As shown, the method includes: Step S1: Perform data augmentation on multiple traffic data streams based on the pre-trained generative adversarial network to obtain multiple augmented data streams.

[0027] The generative adversarial network uses the data quality of multiple training data streams as training constraints. Each of the multiple traffic data streams corresponds one-to-one with multiple traffic sensors in the target road network. The data quality includes accuracy indicators, completeness indicators, timeliness indicators, and reliability indicators.

[0028] In this invention, the target road network can be understood as any part of the control needs of users associated with the urban road network. For example, when the user's control needs are vehicle route planning needs, the target road network can be understood as the part of the urban road network corresponding to the convex hull formed by the path start point, path end point and surrounding key nodes of the vehicle route planning needs. Here, key nodes can be understood as road intersections, entrances / exits or road network structure points that are of great significance to route selection within a certain space or topological range, starting from the aforementioned path start point and path end point (such as road intersections within 1 kilometer around the path start point, road intersections within 1 kilometer around the path end point, road intersections within 1 kilometer around the midpoint of the line connecting the path start point and the path end point, etc.).

[0029] The aforementioned traffic data streams can be understood as real-time data streams from roadside fixed sensors (cameras, geomagnetic coils, millimeter-wave radar) and mobile sensors (GPS floating cars, mobile signaling data, intelligent connected vehicle terminals) accessed through standardized data interfaces.

[0030] In applications, protocol adaptation middleware can be designed to address the differences in communication protocols across different data sources, enabling seamless integration of various transmission protocols such as MQTT, Kafka, and HTTP / 2.

[0031] For example, the data access format can be standardized as follows: ,in, For sensor ID, For locally collected timestamps, For observations (flow rate, velocity, density), metadata includes metadata such as sensor type, location coordinates, and sampling frequency.

[0032] In addition, a global spatiotemporal stamp is automatically labeled for each data stream to establish a unified time reference (UTC time, accurate to milliseconds) and spatial reference (WGS-84 coordinate system projected to the local UTM coordinate system).

[0033] This invention defines the spacetime stamp function as follows: Where t is the local acquisition time, Three-dimensional geographic coordinates It uses a UTC timestamp, specifically employing the NTP protocol for time synchronization, with an accuracy better than 10ms.

[0034] It should be noted that the data augmentation process in this invention is specifically as follows: The traffic data stream reported in real time by traffic sensors is monitored. If the monitoring results indicate that the data quality of the traffic data stream meets the expected quality requirements, the current traffic data stream is directly used as an enhanced data stream. If the monitoring results indicate that the quality of the traffic data stream does not meet the expected quality requirements, a pre-trained generative adversarial network is used to enhance the current traffic data stream.

[0035] In this process, when the traffic data stream fails to meet the expected quality requirements due to missing data, the missing data in the traffic data stream is filled in by the generated data output by the generative adversarial network (which uses the closest historical data in the traffic data stream and its corresponding data quality). The traffic data stream after filling in the missing data is called the augmented data stream.

[0036] When the traffic data stream fails to meet the expected quality requirements due to external interference or equipment failure, the generated data output by the generative adversarial network is fused with the low-quality data in the traffic data stream. This process utilizes the generative adversarial network (which uses the low-resolution data in the traffic data stream and its corresponding data quality) to perform super-resolution enhancement on the low-resolution data in the traffic data stream. The fused traffic data stream is then called the enhanced data stream.

[0037] Specifically, this invention periodically (e.g., every 10 seconds or every 30 seconds, performing quality checks on several data points most frequently reported by traffic sensors) monitors the accuracy, completeness, timeliness, and reliability of each traffic data stream. If one or more of these indicators perform worse than expected, it determines that the data quality within the corresponding monitoring period does not meet the expected quality requirements, and accordingly performs data supplementation or over-resolution enhancement to ensure the reliability of the data stream sent to subsequent processes, providing a solid data foundation for subsequent data work.

[0038] Specifically, the accuracy metric is determined based on the signal-to-noise ratio of the corresponding data stream and historical prediction errors.

[0039] For example, accuracy metrics It can be represented as: in, , This represents the signal-to-noise ratio of the corresponding data stream; To correspond to the historical prediction errors of the data stream over the most recent N time steps, This represents the historical observation value of the corresponding sensor at the k-th recent time step. This represents the actual true value of the corresponding sensor at the k-th time step in the most recent period (obtained through multi-source cross-validation and posterior state estimation), used to characterize the long-term accuracy performance of the corresponding sensor. , Weighting coefficients (N=3, weighting coefficients) = =0.5, normalized mapping function (For linear mapping) The sum of the two is 1.

[0040] The integrity index is determined based on the actual number of data points received, the expected number of data points received, and the maximum duration of continuous missing data in the corresponding data stream.

[0041] For example, integrity metrics It can be represented as: in, This represents the actual number of data points received in the corresponding data stream. The expected number of data points to be received for the corresponding data stream. This represents the maximum duration of consecutive missing data in the corresponding data stream. To assess the time window, The minimum integrity threshold (the expected number of data points to be received if the standard sampling frequency of the sensor is 1 time / minute). =60, at this point, the minimum integrity threshold is set. =0.7, meaning that only data with a coverage rate of over 70% has analytical value. This is a complete threshold function, which takes the value 1 when the condition is met and takes the value 0 when the condition is not met.

[0042] The timeliness index is determined based on the data latency of the corresponding data stream and the standard deviation of the time interval between adjacent data packets.

[0043] For example, timeliness indicators It can be represented as: in, This refers to the data latency of the corresponding data stream (specifically indicating the latency from data generation to availability). Let λ be the standard deviation of the time interval between adjacent data packets in the corresponding data stream, and λ be the attenuation coefficient. The maximum data delay threshold is set (to ensure real-time signal timing). =5.0 seconds, decay coefficient =0.2), This is a time-sensitive threshold function, which takes a value of 1 when the condition is met and a value of 0 when the condition is not met.

[0044] The reliability index is determined based on the device failure rate and abnormal data ratio of the corresponding data stream.

[0045] For example, reliability metrics It can be represented as: in, The device failure rate corresponding to the data stream (sensor) Count the abnormal data (abnormal data can be located and identified through anomaly detection algorithms). This represents the ratio of abnormal data in the corresponding data stream. The duration of performance degradation is denoted as μ, which is obtained primarily through the system's real-time anomaly monitoring module. When the anomaly detection algorithm identifies a trend deviation in the data stream characteristics, the system timer is started to record the time span from the degradation trigger moment to the current calculation moment. μ is the degradation attenuation coefficient (its value varies adaptively with the sensor's hardware type; for example, for sensors prone to long-term performance deviations, the attenuation coefficient corresponding to their hardware characteristics is μ=0.01). This is a priori score for long-term operational stability.

[0046] In this invention, the loss function value corresponding to the network parameters of the generative adversarial network is obtained by weighted calculation of multiple data stream loss values. The multiple data stream loss values ​​correspond one-to-one with multiple training data streams, and the calculation weight of the data stream loss value is positively correlated with the data quality of the corresponding training data stream.

[0047] During the training phase of a generative adversarial network, its generator G is fed with a low-quality training data stream. Using its quality metric Q as input, augmented data is generated. The discriminator D simultaneously evaluates the authenticity and quality fit of the generated samples.

[0048] The generator G dynamically adjusts the loss weights of the generator network based on quality feedback: in and This is a quality-adaptive weighting function that automatically increases the weight of quality constraint terms when the input data quality is low, ensuring the controllability of the quality of the generated data. Indicates the content loss between the generated data and the ground truth samples during training. Indicates the data quality loss of the data generated during training.

[0049] The generator G constructs a spatiotemporal continuous field function based on implicit neural representation (INR). , where v is the traffic state vector (flow, speed, density).

[0050] Network parameters Optimize using the following quality-weighted loss function: Among them, weight The main role of high-quality observation points in training (e.g., in a target road network, there are two adjacent detection points—node A and node B, where node A is a high-precision video detector and node B is a coil subject to strong magnetic interference; the overall quality score of node A...) The overall quality score of node B In this case, compared to node B, the contribution / influence of the training data stream corresponding to node A will be greater during network parameter optimization. Indicates data stream loss value, hyperparameter Constrain the rate of change of traffic conditions to prevent non-physical jumps in generated data (set regularization coefficients). =0.1), For traffic situation observation vectors of sensing nodes, The rate of change of traffic conditions.

[0051] For super-resolution enhancement of low-resolution data, a conditional generative adversarial network (cGAN) can be used, with the quality index Q as the conditional vector, to generate a high-resolution spatiotemporal traffic feature map. Discriminator D simultaneously evaluates the authenticity and quality fit of the generated data: in, [.] represents the mathematical expectation operator. This indicates that the discriminator, given the quality condition vector Q, can analyze the real traffic feature map. The confidence level of identification, The enhanced feature map is reconstructed by the generator based on the random latent variable z and the quality condition Q. This is a quality consistency penalty coefficient, used to adjust the balance between adversarial losses and quality constraints. This refers to the quality assessment estimate obtained by the system from the posterior evaluation of the generated samples, and It is the target quality expectation vector produced based on step S1.

[0052] Based on the above settings, a four-dimensional quality assessment system corresponding to accuracy, completeness, timeliness, and reliability indicators is used to guide the enhancement of the quality of raw data, enabling the system to have adaptive capabilities when data quality fluctuates, thereby overcoming the fluctuations in application effects caused by fluctuations in raw data in existing technologies.

[0053] Step S2: Based on the sensor capabilities of the traffic sensor corresponding to each enhanced data stream, perform feature fusion on multiple enhanced data streams to obtain the target feature vector.

[0054] The sensor capabilities include spatial coverage indicators, temporal accuracy indicators, historical reliability indicators, and equipment operation indicators. The sensor capabilities and the corresponding enhanced data streams show a positive correlation in the fusion weights during feature fusion.

[0055] Specifically, the spatial coverage index is determined based on the degree of overlap between the sensing range of the corresponding traffic sensor and the target road network.

[0056] For example, spatial coverage indicators It can be represented as: in, Indicates the area of ​​the overlap between the sensing range of the corresponding traffic sensor and the target road network. Indicates the total area of ​​the target road network.

[0057] The time-domain accuracy index is determined based on the sampling frequency and communication delay of the corresponding traffic sensor.

[0058] For example, time-domain accuracy indicators It can be represented as: in, To correspond to the sampling frequency of the traffic sensor, This represents the highest sampling frequency achievable by traffic sensors in applications. To address the communication delays of traffic sensors, The maximum allowable communication delay for traffic sensors in application (set based on the control layer's bottom-line requirements for real-time performance). =2.0 seconds).

[0059] The historical reliability index is determined based on the historical prediction error of the corresponding traffic sensor.

[0060] For example, historical reliability indicators It can be represented as: in, This corresponds to the historical prediction error of the traffic sensor (specifically, its average prediction error within the most recent time window). This is the sensitivity coefficient. (Historical reliability index) This is used to quantify the accuracy stability of sensing nodes within the most recent time window. The average prediction error is converted into a confidence score in the [0,1] interval using an exponential mapping, and the system's sensitivity coefficient is set. =0.1, this parameter is set according to the road network fault tolerance preference, the larger the value, the more severe the penalty for error).

[0061] The equipment operation indicators are determined based on the equipment failure rate and abnormal data ratio of the corresponding traffic sensors.

[0062] For example, equipment operating indicators It can be represented as: .

[0063] It should be noted that the positive correlation between the sensor capability and the corresponding enhanced data stream in feature fusion should be understood as follows: during feature fusion, the enhanced data stream with high sensor capability will receive higher weight, while the contribution of the enhanced data stream with low sensor capability (faulty or degraded) to the target feature vector will be suppressed.

[0064] Specifically, this invention introduces a dynamic capability attenuation mechanism to achieve rapid isolation of faulty or degraded sensors with low capability values. Specifically, when a sudden drop in the quality index of a sensor is detected (e.g., ...), the mechanism will isolate the sensor from faulty or degraded low capability values. The system monitors the comprehensive quality score of the sensing nodes in real time. To determine whether to activate the uncertainty isolation procedure. Set the system's early warning threshold. =0.45, attenuation coefficient =0.5), which will automatically trigger the capability vector decay function, which can be expressed as: in, γ is the instantaneous capability vector (calculated from stage S1), and γ is the attenuation coefficient (usually taken as 0.1~0.5). This represents vector multiplication. The trigger moment (i.e. the moment when a sudden drop in the sensor's quality index is detected) is used to ensure that the capability value of a faulty or degraded sensor decays rapidly within 50ms, thus avoiding contamination of the feature fusion results by low-quality data.

[0065] The above feature fusion process is completed based on a fusion network with a multi-head spatiotemporal attention mechanism. For the target spatiotemporal location (x, y, t), the contribution weight of each sensor is dynamically calculated, and this contribution weight can be expressed as: in, For query vector, For key vectors, H stands for Attention Head Index. This represents vector concatenation. The dimension of the key vector.

[0066] Multi-head attention output is gated by capability vector: in, For the number of attention heads, This represents the total number of sensor nodes. This represents the output projection matrix of the h-th head. This represents the dynamic attention weight of sensor i in the h-th head. The value vector of sensor i in the h-th head. It is a sigmoid gated function. Let i be the capability vector of sensor i. The parameters of the gated network are used to achieve adaptive modulation of the attention output by the capability vector, thereby realizing that the fusion weights of sensor capabilities and corresponding enhanced data streams during feature fusion are positively correlated.

[0067] In practical applications, to further enhance the information richness of the target feature vector, multiple enhanced data streams can be fused to obtain fused features. These fused features can then be combined with weather conditions. (Rainfall intensity, visibility, temperature), special events External factors such as large-scale events, construction control, and accidents are used to construct a target feature vector, which can then be represented as: in, It is an aggregate quality index that characterizes the overall credibility of the current fusion result (i.e., fusion features).

[0068] Step S3: Process the target feature vector according to the pre-trained spatiotemporal graph neural network to obtain congestion prediction information.

[0069] The congestion prediction information is used to indicate the queue length of each road segment in the target road network in future time periods.

[0070] In this invention, the target road network is abstractly modeled as a hierarchical graph structure: in, These are regional level nodes, corresponding to traffic zones or functional clusters (such as a crossroads or T-junctions). These are road segment level nodes, corresponding to road segments with homogeneous traffic characteristics (such as multiple roads involved in an intersection). For lane layer nodes, corresponding to physical lanes or guide lane groups (such as the first segment (left turn lane) and the second segment (straight lane) in multiple road segments that make up a road); , These are the inclusion relationship edge sets for region-road segment and road segment-lane, respectively.

[0071] Information is transferred between layers through aggregation functions, constructing a three-layer data association structure of "region-road segment-lane".

[0072] In this invention, the time-varying graph structure corresponding to the spatiotemporal graph neural network can be represented as: Among them, node set Corresponding to road segment or region layer entities, node features Time-varying edge sets Representing dynamic traffic associations, edge weight matrix It is determined by both real-time traffic status similarity and historical patterns.

[0073] Specifically, the matrix element values ​​in the edge weight matrix corresponding to the spatiotemporal graph neural network are determined by the prior historical association strength and the real-time association strength between the corresponding two road segments in the target road network. The real-time association strength is used to indicate the similarity of the feature components of the corresponding two road segments in the target feature vector.

[0074] Based on the above settings, real-time association strength is used to track the real-time association between different road segments in the target road network, while historical association strength is used as a priori for empirical constraints to suppress the risk of data distortion caused by abnormal noise. The two work together to complete the adaptive dynamic update of the edge weight matrix, ensuring that the edge weight matrix accurately reflects the topological association between different road segments in the target road network.

[0075] For example, the matrix element values ​​in the edge weight matrix corresponding to a spatiotemporal graph neural network can be represented as: in, For cosine similarity, Assuming the historical correlation strength between road segment i and road segment j in the target road network as a priori, The real-time correlation strength between road segment i and road segment j in the target road network. This is the dynamic adjustment coefficient.

[0076] The edge weight matrix is ​​updated in real time to reflect traffic relationships through a graph-structured dynamic evolution network. in, , To evolve network parameters, the graph structure is dynamically adjusted based on the current traffic conditions and quality indicators, thereby achieving adaptive evolution of the topological relationships between different road segments in the target road network.

[0077] In addition, the graph convolution module corresponding to the spatiotemporal graph neural network includes a quality mask matrix corresponding to the target feature vector. The quality mask matrix is ​​used to indicate the sensor capability of at least one traffic sensor associated with each road segment in the target road network.

[0078] In the above settings, a quality mask matrix is ​​used to explicitly introduce data quality weights when aggregating neighbor information, so as to achieve adaptive suppression of low-quality node information.

[0079] For example, a graph convolutional layer that introduces a quality mask matrix can be called a quality-weighted graph convolutional layer (QA-GCN): in, To add a self-loop adjacency matrix, The degree matrix and the quality mask matrix are shown. .

[0080] The calculation process uses Chebyshev polynomial approximation to accelerate the computation: in, For the normalized Laplace matrix, Let K be a Chebyshev polynomial, where K is the order of the polynomial (usually K=3).

[0081] Spatiotemporal graph neural networks employ a variant of gated recurrent units—the quality-aware GRU (QA-GRU)—to capture multi-scale temporal patterns: in Let be the quality embedding vector at time t, and adjust the information flow through a gating mechanism.

[0082] Furthermore, a periodic attention mechanism is introduced for long-cycle patterns: in It is a typical periodic set. For periodic attention parameters.

[0083] Furthermore, the congestion prediction information is obtained by weighting historical regression information and network prediction information, wherein the historical regression information is used to indicate the queue length of each road segment in the target road network in the historical reference period associated with the future time period, and the network prediction information is used to indicate the queue length of each road segment in the target road network in the future time period as predicted by the spatiotemporal graph neural network.

[0084] Network prediction information can effectively capture complex spatiotemporal dependencies, but it is sensitive to low-quality data. By combining it with historical regression information, and taking advantage of the fact that historical regression information is not sensitive to abnormal fluctuations, the network prediction information can be constrained, thereby suppressing data bias caused by abnormal fluctuations.

[0085] For example, the process of weighted fusion of historical regression information and network prediction information can be represented as: in, Used to indicate congestion prediction information Used to indicate network prediction information Used to indicate historical regression information To predict confidence levels, when When the trust model is high, it predicts; when When the value is low, it reverts to the historical mean.

[0086] The above prediction confidence level can be expressed as: .

[0087] In application, to enhance the interpretability of network prediction information, the method further includes: Gradient-weighted class activation mapping (Grad-CAM) and attention weight backtracking are performed based on the network prediction information to obtain a congestion propagation heatmap, wherein the congestion propagation heatmap is used to indicate the contribution of each road segment in the target road network to the queuing length of other road segments in future time periods.

[0088] The above setup utilizes a congestion propagation heat map to trace the causes of congestion in each section of the target road network in reverse, facilitating subsequent congestion cause analysis and congestion management.

[0089] Specifically, the congestion propagation heatmap can be obtained using the following formula: in, L represents the calculated congestion cause attribution matrix; L represents the total number of layers in the spatiotemporal graph neural network (QA-STGNN); and H represents the total number of heads in each layer's multi-head attention mechanism. Representing the Dynamic attention weight mapping for the h-th head in layer; This indicates that the prediction loss function is applied to the first... The partial derivative (gradient) vector of the layer dynamic adjacency matrix A is used to characterize the sensitivity of feature jumps to the prediction results; the symbol ⊙ represents the Hadamard product (element-wise multiplication). The system engineering significance of this item is that, by coupling the attention distribution with the gradient response, it identifies which nodes in the road network significantly contribute to the accumulation of traffic pressure in the spatiotemporal dimensions.

[0090] Among them, heat map values Representation edge Contribution to the prediction results For the k-th prediction target (such as road segment speed, queue length). This represents the dynamic edge weight between node i and node j; To predict the output response gradient relative to the topological connectivity, Defined as gradient importance weight coefficients, they are usually obtained by predicting gradients using global average pooling.

[0091] It should be noted that the future time period can be represented by multiple discrete future moments to reduce the data volume of the input hybrid control model, thereby improving the output efficiency of traffic control information while ensuring the accuracy and reliability of the final output traffic control information.

[0092] In this invention, the future time period is specifically set as the next 15 minutes, and the corresponding discrete future moments are the next 5 minutes, the next 10 minutes, and the next 15 minutes.

[0093] In addition, to guide the use of congestion prediction information by the hybrid control model, this invention also feeds the prediction confidence index corresponding to the congestion prediction information into the hybrid control model for processing.

[0094] Step S4: Process the congestion prediction information and user control requirements according to the pre-trained hybrid control model to obtain traffic control information.

[0095] The hybrid control model includes a model predictive control module and a deep reinforcement learning module, and the traffic control information is used to indicate a control strategy that matches the control requirements.

[0096] In this context, the user's control requirements can be understood as the user's travel needs within the target road network. For example, when the user is a driver, their control requirements could be to minimize the time / distance travel from the origin to the destination (the control strategy matching these requirements would be the navigation path with the shortest time / distance travel). When the user is a traffic management department employee, their control requirements could be to minimize the cumulative queue length of each road segment in the target road network (the control strategy matching these requirements would be the temporary traffic light control strategy that minimizes the cumulative queue length).

[0097] Specifically, the traffic control information is obtained by fusing the first information output by the model prediction control module and the second information output by the deep reinforcement learning module.

[0098] The sum of the fusion weights of the first information and the second information is 1, and the fusion weight of the second information is positively correlated with the prediction confidence corresponding to the congestion prediction information.

[0099] In the model predictive control module, the control sequence for the next N steps is continuously optimized based on the traffic flow model, providing a basic control strategy. The corresponding objective function can be expressed as: Where u is the control input (phase duration), x is the state (queue length), f is the prediction model, w is the disturbance, and the constraints include minimum / maximum green light time (determined according to the user's control requirements), etc.

[0100] In the deep reinforcement learning module, the proximal policy optimization algorithm is used to train the policy network. The corresponding objective function can be expressed as: in, The importance sampling ratio, For generalized advantage estimation, This is the trimming parameter (usually set to 0.2).

[0101] The mixed output is: ,in, Refers to traffic control information. Refers to the first piece of information. Refers to the second piece of information, the weight of which is dynamically adjusted based on the prediction confidence level. When the value is high, trust DRL; when the value is low, rely on MPC.

[0102] Furthermore, the near-end policy optimization (PPO) algorithm is used to train the signal timing strategy. Action space The reward function is designed as follows, combining the phase sequence and green light duration for the next cycle: in, A phase switching penalty term is defined to suppress frequent invalid phase switching of traffic lights, ensuring smooth traffic flow. A phase state indication function is defined. ,in, This indicates the phase number of the current decision step. This is the phase number of the previous decision step; This is a binary indicator function; it takes a value of 1 when a phase transition occurs, and 0 otherwise. Throughput is the throughput reward, defined as... Where L is the total number of approach lanes at the intersection. This represents the total number of vehicles that actually left the intersection and crossed the stop line from the j-th entrance lane within the current decision-making time step; the last item is the quality perception penalty item, where, The confidence index for the output prediction is used when the uncertainty of the prediction model increases (i.e., At a lower level, An increase in the value of generates a larger negative reward (penalty), ensuring that a conservative control strategy is adopted when the prediction confidence is low.

[0103] And, each (Typically 5 seconds) Resolve for the optimal control sequence: Where u is the control input (phase duration) and x is the state (queue length). For the prediction model, w represents the perturbation. This is the cost to the end user.

[0104] In summary, this invention uses four-dimensional quality assessment (accuracy, completeness, timeliness, and reliability indicators) and a dynamic capability decay mechanism (capability vector decay function) to automatically identify faulty sensors and complete weight reconstruction in a short time. Under the same scenario, the MAPE decay is controlled within 20%, and the continuity of prediction service can be maintained without manual intervention. This can reduce system operation and maintenance costs by more than 60% (while existing methods such as ST-GCN and Graph WaveNet have a MAPE decay of 67-70% when 20% of sensors fail, requiring manual intervention to restart the model or change the data source).

[0105] The MPC-DRL hybrid architecture (i.e., hybrid control model) uses MPC to provide basic safety constraints, while DRL only makes minor adjustments, improving sample efficiency and shortening training time from several weeks. When the prediction confidence is lower than the threshold, it automatically switches to a conservative control strategy to avoid safety accidents such as signal timing conflicts (while existing MPC methods rely on accurate traffic flow models, and the control efficiency drops by 30-50% when the model is mismatched; the pure PPO method is unstable during training and requires millions of rounds of interaction to converge).

[0106] In one embodiment, the present invention also provides a traffic congestion prediction and control system 200 based on quality perception and multi-source data fusion, such as... Figure 2 As shown, the traffic congestion prediction and control system 200 based on quality perception and multi-source data fusion includes: The data augmentation module 201 is used to augment multiple traffic data streams according to a pre-trained generative adversarial network to obtain multiple augmented data streams. The generative adversarial network uses the data quality of multiple training data streams as training constraints. The multiple traffic data streams correspond one-to-one with multiple traffic sensors in the target road network. The data quality includes accuracy indicators, integrity indicators, timeliness indicators, and reliability indicators. The feature fusion module 202 is used to perform feature fusion on multiple enhanced data streams based on the sensor capabilities of the traffic sensor corresponding to each enhanced data stream to obtain a target feature vector. The sensor capabilities include spatial coverage indicators, temporal accuracy indicators, historical reliability indicators, and equipment operation indicators. The sensor capabilities and the fusion weights of the corresponding enhanced data streams during feature fusion are positively correlated. The congestion prediction module 203 is used to process the target feature vector according to the pre-trained spatiotemporal graph neural network to obtain congestion prediction information, wherein the congestion prediction information is used to indicate the queue length of each road segment in the target road network in a future time period. Traffic control module 204 is used to process the congestion prediction information and the user's control needs according to a pre-trained hybrid control model to obtain traffic control information. The hybrid control model includes a model prediction control module and a deep reinforcement learning module. The traffic control information is used to indicate a control strategy that matches the control needs.

[0107] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the traffic congestion prediction and control system based on quality perception and multi-source data fusion and the traffic congestion prediction and control method based on quality perception and multi-source data fusion provided in the above embodiments belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0108] This invention also provides an electronic device. Please refer to [link to relevant documentation]. Figure 3 The electronic device may include a processor 301, a memory 302, and a program 3021 stored in the memory 302 and capable of running on the processor 301.

[0109] When program 3021 is executed by processor 301, it can achieve the following: Figure 1 Any steps in the corresponding method embodiments and the achievement of the same beneficial effects will not be repeated here.

[0110] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by hardware related to program instructions, and the program can be stored in a readable medium.

[0111] This invention also provides a readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described functions. Figure 1 Any step in the corresponding method embodiment can achieve the same technical effect, and will not be repeated here to avoid repetition.

[0112] The computer-readable storage medium of this invention can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0113] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0114] The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0115] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or terminal. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0116] This invention also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to implement the traffic congestion prediction and control method based on quality perception and multi-source data fusion provided in the above embodiments.

[0117] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0118] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A traffic congestion prediction and control method based on quality perception and multi-source data fusion, characterized in that, The method includes: Data augmentation is performed on multiple traffic data streams using a pre-trained generative adversarial network to obtain multiple augmented data streams. The generative adversarial network uses the data quality of the multiple training data streams as training constraints. The multiple traffic data streams correspond one-to-one with multiple traffic sensors in the target road network. The data quality includes accuracy indicators, completeness indicators, timeliness indicators, and reliability indicators. Based on the sensor capabilities of the traffic sensor corresponding to each enhanced data stream, feature fusion is performed on multiple enhanced data streams to obtain a target feature vector. The sensor capabilities include spatial coverage indicators, temporal accuracy indicators, historical reliability indicators, and equipment operation indicators. The sensor capabilities and the fusion weights of the corresponding enhanced data streams during feature fusion show a positive correlation. The target feature vector is processed by a pre-trained spatiotemporal graph neural network to obtain congestion prediction information, wherein the congestion prediction information is used to indicate the queue length of each road segment in the target road network in a future time period. The congestion prediction information and user control requirements are processed according to the pre-trained hybrid control model to obtain traffic control information. The hybrid control model includes a model prediction control module and a deep reinforcement learning module. The traffic control information is used to indicate control strategies that match the control requirements. The aforementioned feature fusion process is completed using a fusion network based on a multi-head spatiotemporal attention mechanism, where the h-th head contains sensors... The dynamic attention weights can be expressed as: in, For the h-th head sensor The corresponding query vector, For the h-th head sensor The corresponding key vector, H stands for Attention Head Index. The dimension of the key vector; Multi-head attention output is gated by capability vector: in, For the number of attention heads, This represents the total number of traffic sensor nodes. This represents the output projection matrix of the h-th head. Sensor in the hth head The value vector, It is a sigmoid gated function. For sensors The capability vector, These are the parameters for the gating network.

2. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 1, characterized in that, The accuracy index is determined based on the signal-to-noise ratio and historical prediction error of the corresponding data stream; the integrity index is determined based on the actual number of received data points, the expected number of received data points, and the maximum duration of consecutive missing data in the corresponding data stream; the timeliness index is determined based on the data latency of the corresponding data stream and the standard deviation of the time interval between adjacent data packets; and the reliability index is determined based on the device failure rate and the ratio of abnormal data in the corresponding data stream.

3. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 1, characterized in that, The loss function value corresponding to the network parameters of the generative adversarial network is obtained by weighted calculation of multiple data stream loss values. Each of the multiple data stream loss values ​​corresponds one-to-one with a multiple training data stream, and the calculation weight of the data stream loss value is positively correlated with the data quality of the corresponding training data stream.

4. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 1, characterized in that, The spatial coverage index is determined based on the degree of overlap between the sensing range of the corresponding traffic sensor and the target road network. The temporal accuracy index is determined based on the sampling frequency and communication delay of the corresponding traffic sensor. The historical reliability index is determined based on the historical prediction error of the corresponding traffic sensor. The equipment operation index is determined based on the equipment failure rate and abnormal data ratio of the corresponding traffic sensor.

5. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 1, characterized in that, The matrix element values ​​in the edge weight matrix corresponding to the spatiotemporal graph neural network are determined by the prior historical association strength and the real-time association strength between the corresponding two road segments in the target road network. The real-time association strength is used to indicate the similarity of the feature components of the corresponding two road segments in the target feature vector.

6. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 1, characterized in that, The graph convolution module corresponding to the spatiotemporal graph neural network includes a quality mask matrix corresponding to the target feature vector. The quality mask matrix is ​​used to indicate the sensor capability of at least one traffic sensor associated with each road segment in the target road network.

7. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 1, characterized in that, The congestion prediction information is obtained by weighting historical regression information and network prediction information. The historical regression information is used to indicate the queue length of each road segment in the target road network in the historical reference period associated with the future time period, and the network prediction information is used to indicate the queue length of each road segment in the target road network in the future time period as predicted by the spatiotemporal graph neural network.

8. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 7, characterized in that, The method further includes: Based on the network prediction information, gradient-weighted class activation mapping and attention weight backtracking are performed to obtain a congestion propagation heatmap, wherein the congestion propagation heatmap is used to indicate the contribution of each road segment in the target road network to the queuing length of other road segments in future time periods.

9. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 1, characterized in that, The traffic control information is obtained by fusing the first information output by the model prediction control module and the second information output by the deep reinforcement learning module.

10. The traffic congestion prediction and control method based on quality perception and multi-source data fusion according to claim 9, characterized in that, The sum of the fusion weights of the first information and the second information is 1, and the fusion weight of the second information is positively correlated with the prediction confidence corresponding to the congestion prediction information.

Images