A traffic situation assessment and ramp cooperative control method

Abstract

This invention relates to the field of traffic control systems for road vehicles, and discloses a method for traffic situation assessment and ramp cooperative control. The method includes: collecting multi-source heterogeneous data from vehicles and roadside equipment via intelligent network; constructing road segment node features based on the multi-source data; dynamically calculating the inter-segment correlation weights using a dynamic graph attention network; and extracting traffic flow time-series features using a long short-term memory network; dynamically weighting and fusing spatial correlation features and time-series features with cross-modal bidirectional attention to obtain traffic situation assessment results; and using a model predictive control framework, continuously optimizing candidate signal control schemes based on a traffic flow simulation model to generate and implement optimal ramp control commands. This method solves the problems of single data source, inaccurate assessment, control lag, and traffic congestion in existing technologies, achieving the goals of accurate situation perception, real-time control optimization, improved traffic efficiency, and effective congestion relief.

Description

Technical Field

[0001] This invention relates to the field of traffic control system technology for road vehicles, and in particular to a method for traffic situation assessment and ramp cooperative control. Background Technology

[0002] With the acceleration of urbanization, traffic flow on expressways and highways continues to grow, exacerbating traffic congestion. Traditional traffic situation assessment methods rely on limited sensor data, making it difficult to comprehensively and accurately reflect traffic conditions. Meanwhile, existing ramp signal control strategies lack precise responses to real-time traffic situations, failing to effectively alleviate congestion. The development of intelligent connected vehicle technology offers new avenues for addressing these issues. By collecting real-time interaction data between vehicles and infrastructure, a more comprehensive understanding of traffic conditions can be achieved, supporting optimized traffic control. However, existing technologies based on intelligent connected vehicle data still have room for improvement in terms of innovation and effectiveness.

[0003] For example, Chinese patent CN119323879B discloses a multimodal traffic flow prediction method based on multi-source data feature fusion, providing the following technical solution: acquiring traffic data; dividing the traffic data into training and testing sets; constructing a traffic prediction model and training the model using the training set; and validating the model using the testing set. The traffic prediction model includes: a cloud map encoder for extracting features from cloud map data; a spatiotemporal encoder for extracting features from spatiotemporal data; a fusion module for fusing cloud map features and spatiotemporal features to obtain fused features; and the addition of the cloud map features, spatiotemporal features, and fused features, followed by inputting the result into a spatiotemporal decoder for prediction, to obtain the final prediction result. This invention integrates multi-dimensional and multi-faceted information, simplifies redundant information in the spatiotemporal feature extraction process, and achieves efficient and high-precision prediction, which can be widely applied in the field of traffic flow prediction technology. However, the aforementioned multimodal traffic flow prediction method based on multi-source data feature fusion can only achieve multimodal traffic flow prediction, but cannot achieve real-time ramp signal control, lacks dynamic traffic association mining and closed-loop optimization capabilities, and its model is slightly redundant and has a slow response. Summary of the Invention

[0004] This invention addresses the problems of single data source, inaccurate assessment, lagging control, and traffic congestion in existing technologies. It proposes a traffic situation assessment and ramp collaborative control method, which achieves the goals of accurate situation perception, real-time control optimization, improved traffic efficiency, and effective congestion relief.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A traffic situation assessment and ramp coordinated control method, comprising: Collect multi-source heterogeneous data from vehicles and roadside equipment through intelligent connected vehicles; Based on multi-source data, road segment node features are constructed. Dynamic graph attention network is used to dynamically calculate the association weights between road segments, and long short-term memory network is used to extract traffic flow time series features. By dynamically weighting and pooling spatial correlation features with time series features and fusing cross-modal bidirectional attention, traffic situation assessment results are obtained. Using a model predictive control framework, candidate signal control schemes are continuously optimized based on traffic flow simulation models to generate and implement optimal ramp control commands.

[0006] It integrates intelligent connected data collection, spatiotemporal feature fusion, and model predictive control to achieve a comprehensive and accurate assessment of traffic conditions. Based on this assessment, it optimizes ramp signal control, effectively improving the efficiency of expressways and alleviating traffic congestion.

[0007] As a preferred approach, real-time vehicle information is collected from onboard sensors. Traffic flow, vehicle density and occupancy, and vehicle point cloud data are collected by deploying roadside equipment. Real-time weather and road condition data are also integrated. During data transmission, intelligent network encryption and authentication mechanisms are employed, and multi-source heterogeneous data are aggregated to edge computing nodes or the cloud.

[0008] It ensures the diversity and real-time nature of data sources, guarantees the security of data transmission through encryption and authentication, and facilitates centralized processing of data at the edge or cloud, providing a high-quality and reliable data foundation for subsequent traffic situation assessment.

[0009] As a preferred approach, the preprocessed data of each road segment is constructed into initial feature vectors for nodes. These vectors are then linearly transformed into query vectors, key vectors, and value vectors using a learnable weight matrix. The original association scores between nodes are calculated, and a time decay factor based on data latency and an interaction intensity factor based on vehicle transfer rate within a preset time window are introduced for correction. The dynamic association weights are obtained by normalization using the softmax function. The value vectors of adjacent nodes are then weighted and summed according to the association weights to aggregate the spatial features of each road segment. The spatial features and association weights are updated at a preset period.

[0010] The dynamic graph attention network corrects the association weights in real time through time decay factors and interaction intensity factors, which can accurately capture the dynamic spatial relationships between road segments. The preset periodic update ensures the timeliness of spatial features and improves the accuracy and response speed of traffic state spatial feature extraction.

[0011] Preferably, using historical traffic flow data sequences as input, the Long Short-Term Memory Network controls the inflow of current input information through an input gate, controls the amount of information retained in the memory unit at the previous moment through a forget gate, and controls the amount of information output from the current memory unit to the hidden state through an output gate. It uses internal gating mechanisms and memory units to capture long-term dependencies and periodic change characteristics in traffic flow data.

[0012] Long Short-Term Memory (LSTM) networks effectively process time-series data through their gating mechanism, capturing long-term dependencies and periodic changes in traffic flow, improving the accuracy of time feature extraction, and providing reliable time-dimensional information for traffic condition prediction.

[0013] As a preferred approach, based on the scene perception factor calculated from the real-time congestion index and traffic fluctuation rate, the fusion weights of spatial and temporal features are dynamically allocated and weighted summation is performed to obtain preliminary fusion features. Prospective features are generated using path planning data, and a gating vector jointly determined by the spatiotemporal fusion features is adopted. The injection ratio of the prospective features is calculated through the sigmoid activation function and then weighted and fused with the features processed by cross-modal bidirectional attention.

[0014] Dynamic weight pooling adaptively adjusts feature weights based on real-time traffic scenarios, cross-modal bidirectional attention enables deep feature interaction, and forward-looking feature injection uses path planning data to predict future trends, making the fused features more comprehensive and accurate, and significantly improving the sensitivity and predictive ability of traffic situation assessment.

[0015] As a preferred approach, a hierarchical reinforcement learning architecture is constructed, comprising a macro-level traffic flow control layer and a micro-level vehicle scheduling layer. The macro-level layer takes traffic situation assessment results as state input and outputs a ramp traffic flow allocation scheme, with its reward function based on total traffic flow and congestion area. The micro-level layer takes the queue length and arrival rate of vehicles at a single ramp as state input and outputs a specific traffic light timing scheme, with its reward function based on average vehicle delay and queue length.

[0016] The hierarchical reinforcement learning architecture optimizes both macro-level overall traffic flow and micro-level local scheduling. The reward function is designed for different levels of objectives, achieving coordinated control from the overall to the details, and improving the pertinence of ramp signal control strategies and overall traffic efficiency.

[0017] As a preferred approach, in each control cycle, multiple sets of candidate control schemes for ramp signals are generated based on the current traffic conditions and the future multi-time period traffic conditions predicted by intelligent connected vehicle data. The implementation effect of each candidate scheme is simulated within the set prediction time domain using a traffic flow simulation model with parameters calibrated by intelligent connected vehicle data.

[0018] Rolling optimization generates candidate solutions based on real-time and predicted data, and simulates the effects of the solutions to ensure the foresight and adaptability of the control scheme, avoiding blind control, thereby improving control accuracy and traffic flow.

[0019] As a preferred approach, when training the control model for the new ramp, a transfer learning mechanism is adopted to transfer a preset proportion of network parameters from a pre-trained source model with similar traffic conditions to the target model for initialization, and to fine-tune the parameters using the data from the new ramp, thereby accelerating model convergence.

[0020] Transfer learning utilizes the knowledge of existing models to transfer and fine-tune parameters, significantly reducing the data and time required to train new models, accelerating model convergence, and improving the deployment efficiency and adaptability of models in different road segments or time periods.

[0021] As a preferred approach, spatial features are used as queries, and temporal features are used as keys and values. Attention weights are calculated and key time segment features are extracted. Temporal features are used as queries, and spatial features are used as keys and values. Attention weights are calculated and key spatial segment features are filtered. The two types of features after bidirectional attention processing are concatenated and compressed through a fully connected layer to obtain the final fused features.

[0022] The cross-modal bidirectional attention mechanism enables deep interaction and mutual enhancement of spatial and temporal features, extracts key spatiotemporal information, improves the quality of feature fusion, and makes traffic situation assessment results more accurate and reliable.

[0023] As a preferred approach, a cost function is constructed that measures the degree to which the system state deviates from the desired state and the magnitude of the control input. An optimization algorithm is used to find the optimal control sequence that minimizes the total cost in the prediction time domain. The first control command in the optimal control sequence is applied to the actual ramp signal control, and prediction and optimization are performed again based on the latest state in the next control cycle to achieve closed-loop feedback control in the rolling time domain.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0025] 1. This invention accurately predicts traffic condition changes by fusing multi-source heterogeneous data and utilizing dynamic graph attention networks to mine spatial relationships between road segments, combined with long short-term memory networks to capture temporal patterns. The signal control strategy based on hierarchical reinforcement learning can adjust ramp flow distribution and signal timing in real time, significantly alleviating congestion, improving overall road network efficiency, and reducing vehicle delays and queue lengths.

[0026] 2. This invention leverages the full-domain perception and real-time interaction in an intelligent connected environment to achieve a comprehensive and sensitive grasp of traffic conditions. It not only focuses on traffic flow and speed but also integrates multimodal data such as meteorological and vehicle behavior data to effectively identify abnormal states and potential risks. Furthermore, through a forward-looking traffic prediction and model predictive control framework, it can provide early warnings and implement collaborative control, thereby smoothing traffic fluctuations, reducing accident rates, and improving driving safety.

[0027] 3. This invention constructs a data-driven intelligent traffic management system. Based on rapid model deployment through transfer learning and adaptive optimization through hierarchical reinforcement learning, it has good scalability and scenario adaptability. Through simulation and rolling optimization, it can scientifically formulate control strategies, improve the efficiency of road network operation, provide real-time and accurate decision support for traffic management departments, and promote the development of intelligent transportation systems to a higher level. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of a traffic situation assessment and ramp coordinated control method according to the present invention.

[0029] Figure 2 This is a flowchart of one embodiment of the traffic situation assessment and ramp coordinated control method of the present invention. Detailed Implementation

[0030] See Figures 1-2 As shown, a traffic situation assessment and ramp coordinated control method includes: Collect multi-source heterogeneous data from vehicles and roadside equipment through intelligent connected vehicles; Based on multi-source data, road segment node features are constructed. Dynamic graph attention network is used to dynamically calculate the association weights between road segments, and long short-term memory network is used to extract traffic flow time series features. By dynamically weighting and pooling spatial correlation features with time series features and fusing cross-modal bidirectional attention, traffic situation assessment results are obtained. Using a model predictive control framework, candidate signal control schemes are continuously optimized based on traffic flow simulation models to generate and implement optimal ramp control commands.

[0031] This invention aims to provide a traffic situation assessment and ramp signal control method for expressways based on intelligent connected vehicle data. Through innovative data processing, model building, and control strategies, it achieves accurate assessment of traffic situation and optimizes ramp signal control strategies accordingly to improve expressway traffic efficiency and alleviate traffic congestion.

[0032] like Figure 1 In one embodiment shown, Figure 1This is an overall flowchart of a traffic situation assessment and ramp coordinated control method according to the present invention. The present invention aims to provide a traffic situation assessment and ramp signal control method for expressways based on intelligent connected vehicle data. Through innovative data processing, model building, and control strategies, it achieves accurate assessment of traffic situation and optimizes ramp signal control strategies accordingly to improve the traffic efficiency of expressways and alleviate traffic congestion.

[0033] First, the intelligent connected system collects multi-source heterogeneous data from vehicles and roadside equipment in real time. Vehicle sensors report real-time information such as vehicle position, speed, and acceleration; roadside equipment such as cameras and radar collect traffic flow, vehicle density, occupancy, and point cloud data; and environmental information such as weather and road conditions is also integrated. All data is encrypted and authenticated during transmission to ensure security, and is then aggregated at edge computing nodes or the cloud for unified processing.

[0034] In the feature extraction stage, a dynamic graph attention network is used to model the relationships between road segments. Based on the preprocessed feature vectors of road segment nodes, query, key, and value vectors are generated through a learnable weight matrix. The original association score is calculated, and a time decay factor and a vehicle transfer rate factor are introduced for dynamic correction. Finally, the dynamic association weights are obtained by softmax normalization, which are then aggregated to form the spatial features of road segments and updated periodically.

[0035] Meanwhile, by using long short-term memory networks to process historical traffic flow sequence data, and through the coordinated regulation of input gates, forget gates and output gates, long-term dependence and periodic change characteristics in traffic conditions are captured, and time series features are extracted.

[0036] Spatial and temporal features are initially fused using dynamic weighted pooling, with weights dynamically determined by scene-aware factors comprised of real-time congestion indices and traffic fluctuation rates. Further, a cross-modal bidirectional attention mechanism is designed: on one hand, using spatial features as the query, key temporal segments are extracted from temporal features; on the other hand, using temporal features as the query, key spatial road segment features matching temporal patterns are filtered. The features processed by bidirectional attention are then concatenated and compressed using a fully connected layer to form fused features.

[0037] Furthermore, by generating forward-looking features using vehicle routing data and dynamically adjusting their weighting ratio with the aforementioned fused features through a gating mechanism, the model's ability to predict future traffic conditions is enhanced.

[0038] The ramp signal control adopts a hierarchical reinforcement learning architecture, comprising a macro-level flow control layer and a micro-level vehicle scheduling layer. The macro-level layer formulates a flow allocation scheme based on the overall traffic situation assessment results, and the reward function integrates the total traffic flow and the area of ​​the congested area. The micro-level layer outputs specific signal timing based on the real-time queue length and vehicle arrival rate at the ramp entrance, and the reward function focuses on the average vehicle delay and queue length.

[0039] The control process introduces a model predictive control framework. In each control cycle, multiple candidate control schemes are generated based on the current state and short-term traffic forecast results. The effects of each scheme are simulated in the prediction time domain by a traffic flow simulation model calibrated with intelligent connected data. An optimization objective is constructed at the cost of system state deviation and control input. The optimal control sequence is solved and implemented in a rolling manner.

[0040] For model training at new ramp entrances, a transfer learning mechanism is employed. This mechanism transfers some parameters from previously trained models with similar traffic conditions for initialization, and then fine-tunes them using local data, significantly improving model convergence speed. The entire control process forms a closed loop, achieving continuous adaptive adjustment under dynamic environments through rolling optimization.

[0041] In another embodiment, the present invention specifically includes: 1. Traffic Situation Assessment Module 1.1 Spatial Feature Extraction: Dynamic Graph Attention Network (DGAT) In an intelligent connected environment, this algorithm deeply integrates heterogeneous data from multiple sources, including onboard sensors, roadside equipment, and traffic management systems. Onboard sensors, through V2X (Vehicle-to-Everything) communication technology, rapidly transmit real-time information such as vehicle position, speed, acceleration, and direction of travel to edge computing nodes or the cloud. Traffic flow and occupancy data collected by roadside equipment such as millimeter-wave radar and cameras are also uploaded via intelligent connected communication links. During transmission, the encryption and authentication mechanisms of intelligent connected systems ensure the security and integrity of the data.

[0042] In an intelligent connected environment, the spatial characteristics of traffic flow exhibit highly dynamic changes, making it difficult for traditional spatial feature extraction methods based on fixed topology to accurately capture their complex relationships. To further explore the spatiotemporal correlations of data, this algorithm introduces a Dynamic Graph Attention Network (DGAT), leveraging the real-time and interactive nature of vehicle-road cooperative data to dynamically characterize the relationships between road segments. By processing real-time vehicle-level interactive data, roadside equipment's full-domain perception data, and cross-segment collaborative decision-making data for each road segment, the algorithm dynamically learns the correlation weights between road segments, aggregating the spatial characteristics of each road segment. This provides a foundation for subsequent spatiotemporal feature fusion and traffic flow prediction, while also providing spatial correlation information for ramp signal control strategies. The specific process is as follows: Data collection and basic features: Collect multi-source data from intelligent connected vehicle environments, including real-time vehicle-level interaction data (location, speed, acceleration, steering intention, etc.), roadside equipment's comprehensive perception data (traffic density, headway, vehicle type distribution, etc.), and cross-segment collaborative decision-making data (signal timing, temporary traffic control, etc.). Preprocess this data, such as cleaning, integrating, and feature extraction, to construct an initial feature vector h for each road segment node. i .

[0043] Association weight calculation: First, the query vector, key vector, and value vector are obtained through linear transformation: Query vector Q i equal to h i Multiplied by W q Among them, W q h is a learnable query weight matrix. i This represents the initial characteristics of the i-th road segment node.

[0044] Key vector K j equal to h j Multiplied by W k W k Let j represent the key weight matrix that can be learned, where j represents the node numbers of other road segments associated with node i, and h represents the key weight matrix that can be learned. j This represents the initial feature vector of the j-th associated road segment node.

[0045] Value vector V j equal to h j Multiplied by W v W v is a learnable value weight matrix.

[0046] Calculate the original association score: score ij equals Q i Multiply by K j T Divide by the square root of d, where d is the feature dimension. This formula is used to measure the original degree of association between node i and node j.

[0047] Introducing a time decay factor λ t and interaction strength factor γ ij Adjust the original correlation score: adjusted_score ij equals score ij Multiply by λ t Multiply by γ ij Wherein, λ t Equals 0.8 t (t is the data delay in minutes), γ ijγ is determined based on the vehicle transfer rate from road segment i to j within 10 minutes. When the transfer rate exceeds 20%, γ is considered to be... ij When γ equals 1.2, it is below 5%. ij It equals 0.8.

[0048] The association scores are normalized using the softmax function to obtain the association weights between node i and node j: α. ij Equals softmax(adjusted_score) ij The adjusted_score equals e. ij Power of 1 Divide by ∑ k e's adjusted_score ik The power is given by the power of k, where k represents all nodes associated with node i, and e is the natural constant.

[0049] Spatial feature aggregation Based on the calculated association weights, the features of all nodes associated with node i are summed using a weighted average to obtain the final spatial features of node i: h i 'equals ∑ j α ij Multiply by V j equals ∑ j α ij Multiply by h j Multiply by W v .

[0050] The node features are updated every 2 seconds, and the associated weights are updated in real time based on new intelligent connected data to ensure that spatial features can reflect the dynamic changes in traffic conditions in a timely manner.

[0051] 1.2 Temporal Feature Extraction: Long Short-Term Memory (LSTM) Algorithm.

[0052] LSTM is used to capture the time-series features of traffic flow, with historical traffic flow data (such as flow rate and vehicle speed sequences over the past hour) as input. The LSTM unit internally controls the inflow, retention, and output of information through input gates, forget gates, and output gates. The specific calculations are as follows: H l+1 Equals σ multiplied by D ~-1 / 2 Multiply by A ~ Multiply by D ~-1 / 2 Multiply by H l W l ; i t Equals σ multiplied by ((W) ii Multiply by x t ) plus (W hi Multiply by ht-1 ) plus b i ); f t Equals σ multiplied by ((W) if Multiply by x t ) plus (W hf Multiply by h t-1 ) plus b f ); o t Equals σ multiplied by ((W) io Multiply by x t ) plus (W ho Multiply by h t-1 ) plus b o ); C ~ t Equals tanh multiplied by ((W) ic Multiply by x t ) plus (W hc Multiply by h t-1 ) plus b c ); C t equal to (f) t ⊙C t-1 ) plus (i t ⊙C ~ t ); h t equals o t ⊙tanh (C t ).

[0053] Among them, H l H represents the hidden layer output of layer l. l+1 Let σ represent the hidden layer output of layer l+1, σ represent the sigmoid activation function, tanh represent the hyperbolic tangent activation function, and D represent the hyperbolic tangent activation function. ~ Let A represent the normalization degree matrix. ~ W represents the normalized adjacency matrix. l Let x represent the weight matrix of the l-th layer. t It is the input at time t, h t-1 and C t-1 These are the hidden state and memory unit at time t-1, i t f t o t These are the input gate, forget gate, and output gate, respectively. (C) ~ t It is a candidate memory unit, C t It is the memory unit updated at time t, h tLet W be the hidden state at time t, W be the weight matrix, b be the bias vector, and ⊙ denote element-wise multiplication. In an intelligent connected environment, time series data has higher continuity and accuracy, and LSTM can better learn the changing patterns of traffic conditions over time.

[0054] 1.3 Spatiotemporal Feature Fusion: Dynamic Weight Pooling and Cross-Modal Bidirectional Interaction Dynamic weighted pooling: The weights of spatial and temporal features are dynamically adjusted based on the real-time traffic scenario. Spatial features are derived from the segment association features of the Dynamic Graph Attention Network (DGAT), and temporal features are derived from the temporal features of the LSTM. A scenario perception factor S is defined, quantified from the current traffic state. S equals α multiplied by [Congestion plus ((1 minus α) multiplied by Volatility)], where Congestion is the congestion index and Volatility is the traffic fluctuation rate. Spatial weights w are dynamically assigned based on S. s and time weight w t The sum of the two is 1. The fusion formula is F. pool equals [w s Multiply by Map(F) space )] plus [w t Multiply by Map(F) time )], where Map(·) is the feature mapping function.

[0055] Cross-modal bidirectional interaction: Design a cross-modal attention layer to achieve bidirectional interaction between spatial and temporal features. Space-to-temporal attention: Using spatial features as the query vector (Q) and temporal features as keys (K) and values ​​(V), calculate the key temporal fragments in the temporal features that are relevant to the current spatial context as F. time 'Equals Attention(F) space Multiplied by W q F time Multiplied by W k F time Multiplied by W v ).

[0056] Where α is the weighting coefficient of the congestion index (α∈[0,1]), Congestion is the congestion index, Volatility is the traffic fluctuation rate, and S is the scene perception factor; w s For spatial feature weights, w t The time feature weights are defined by Map(·), where F is the feature mapping function. pool F represents the characteristics after pooling and fusion. space For spatial characteristics, F time It is a time-related feature.

[0057] Temporal-spatial attention: Using temporal features as the query vector (Q) and spatial features as the key (K) and value (V), the system filters road segments within the spatial features that match the current temporal pattern. Specifically: F space 'Equals Attention(F) time Multiplied by W q ', F space Multiplied by W k ', F space Multiplied by W v ').

[0058] Finally, the features processed by bidirectional attention are concatenated and compressed through a fully connected layer: F fusion Equals FC(Concat(F) space ', F time ')).

[0059] Among them, F fusion For the final fused features, FC(•) is the fully connected layer processing function, and Concat(•) is the feature concatenation function. space 'For spatial features calibrated by temporal attention, F time 'Temporal features after spatial attention calibration.'

[0060] Forward-looking feature injection: Generating forward-looking features F using path planning data from intelligent connected vehicles. future The injection ratio is controlled using a gating mechanism. F fusion Equals σ(g) multiplied by F future Add [(1 minus σ(g)) multiplied by F] cross ].

[0061] Where g is the gating vector (determined by spatiotemporal features), σ is the sigmoid activation function, and F... fusion The final fusion feature F after injecting forward-looking features future As a forward-looking feature, F cross These are the features after bidirectional cross-attention processing.

[0062] This model combines multimodal data to fully explore the information contained in different types of data, achieving a comprehensive depiction of traffic conditions. Specifically, vehicle trajectory data accurately reflects the driving trajectory and behavioral characteristics of individual vehicles, traffic flow data intuitively presents the overall traffic busyness, and road weather data reveals the impact of weather conditions on traffic flow. The deep integration of multimodal data makes the model's understanding of traffic conditions more accurate. Furthermore, intelligent connected vehicle technology ensures the real-time acquisition and efficient fusion of multimodal data, providing the model with high-quality data input and significantly improving the model's sensitivity to traffic conditions.

[0063] Dynamic Graph Attention Network (DGAT) efficiently uncovers complex dynamic spatial relationships between road segments in traffic networks, overcoming the limitations of traditional models that only consider simple distance relationships. It constructs a dynamic graph structure with road segments as nodes and learns the association weights between nodes in real time through an attention mechanism, fully considering the mutual influence between different road segments. Whether the association is formed due to traffic flow shifts or temporary associations caused by traffic events, it can be accurately captured. This allows the model to better grasp the spatial propagation and diffusion patterns of traffic flow. Intelligent connected vehicle technology accelerates the real-time transmission of spatial information, enabling DGAT to respond more quickly to changes in spatial relationships, further improving the timeliness and accuracy of spatial feature extraction.

[0064] LSTM networks excel at handling long-term dependencies in time-series data, effectively capturing the trends and periodic characteristics of traffic data over time. Combining Dynamic Graph Attention Network (DGAT) with LSTM allows the model to simultaneously address the spatiotemporal coupling characteristics of traffic data. During peak hours, the model's traffic state prediction accuracy is approximately 30% higher than that of a single LSTM model, more accurately predicting changes in state parameters such as traffic flow and speed, providing a more reliable basis for ramp signal control. The rich and high-quality multimodal data in intelligent connected environments strongly supports the collaborative work of DGAT and LSTM, further enhancing the model's ability to learn spatiotemporal coupling characteristics.

[0065] 2. Ramp signal control module A ramp signal control algorithm based on hierarchical reinforcement learning and transfer learning: In the intelligent connected vehicle system, the state acquisition, action execution, and reward feedback of each layer of this algorithm are closely linked to the intelligent connected vehicle system. The macro-level traffic control layer can obtain the overall traffic situation output by the traffic situation assessment module in real time through the intelligent connected vehicle system, including the distribution of congested road sections and the overall traffic flow trend; the micro-level vehicle scheduling layer can accurately obtain information such as the real-time vehicle queuing situation and vehicle arrival rate of individual ramps with the help of the intelligent connected vehicle system.

[0066] 1) Hierarchical reinforcement learning part Macro-level traffic flow control layer: The state space represents the overall traffic situation output by the traffic situation assessment module, such as the distribution of congested road sections and the overall traffic flow trend. The action space represents the approximate traffic flow allocation scheme for ramps, such as dividing ramp traffic flow into high, medium, and low levels for control. Reward function R macro The main considerations are the optimization of overall traffic flow and the alleviation of congestion, specifically: R macro Equals [α1 multiplied by Q] total Divide by Q total-max Subtract [α2 multiplied by C] area Divide by C area-max ].

[0067] Among them, Q total Q represents the current total traffic flow on expressways. total-max This is the highest total traffic volume in history, C area C is the current area of ​​the congested region. area-max This represents the area of ​​the historically largest congested region. α1 and α2 are weighting coefficients, adjusted according to actual traffic demand, used to balance the importance of flow optimization and congestion mitigation. Intelligent connectivity ensures real-time and accurate monitoring of the overall traffic situation, making macro-control more targeted.

[0068] Micro-level vehicle scheduling layer: The state space contains real-time vehicle queuing information and arrival rate at a single ramp. The action space contains the specific timing scheme of the traffic lights within a single cycle, such as adjustments to the green and red light durations. Reward function R. micro Taking into account factors such as vehicle delays and queue lengths, the details are as follows: R micro Equals [β1 multiplied by 1 divided by D] avg Subtract [β2 multiplied by L] queue Divide by L queue-max Multiply by θ target ].

[0069] Among them, D avg L is the average delay time of vehicles at the ramp entrance. queue L is the current queue length of vehicles at the ramp entrance. queue-max This represents the historical maximum queue length at this ramp. β1 and β2 are weighting coefficients used to balance the goals of reducing vehicle delays and controlling queue length. Intelligent connectivity allows for more accurate acquisition of vehicle information at the micro-level, enabling traffic light timing adjustments to better meet actual needs.

[0070] Transfer learning component: When training the ramp signal control model for the current road segment, some parameters are transferred from models trained on other similar road segments or during different time periods. Assume the source model parameters are θ. source The target model parameter is θ targetThrough a fine-tuning mechanism, the parameters of the source model are transferred to the target model according to a certain proportion γ, that is: θ target Equals (1 minus γ) multiplied by θ target-init Add [γ multiplied by θ] source ].

[0071] Where, θ target These are the initial parameters of the target model. This method accelerates the convergence speed of the current road segment model, improving convergence speed by approximately 50% compared to traditional reinforcement learning algorithms, resulting in superior control performance. Intelligent connectivity facilitates rapid parameter transfer and sharing between models across different road segments and time periods, improving model training efficiency.

[0072] Collaborative strategies based on intelligent connected data predictive control: Intelligent connected vehicle technology provides comprehensive data support and efficient communication guarantees for predictive control, serving as the core driving force for achieving collaborative control. Real-time collected intelligent connected vehicle data includes individual vehicle information such as real-time location, speed, and acceleration; overall road segment information such as traffic flow, occupancy rate, and vehicle type distribution obtained from roadside sensors; and forward-looking information such as vehicle driving intentions and path planning obtained through vehicle-to-infrastructure (V2I) interaction. This data is transmitted in milliseconds via 5G / V2X communication technology, providing a precise, real-time, and comprehensive data source for predictive control.

[0073] State perception and prediction based on intelligent connected vehicle data A dynamic traffic condition perception model is constructed based on intelligent connected vehicle data. Through multi-source data fusion technology, scattered vehicle data and roadside data are integrated to form a real-time snapshot of traffic conditions on highways and expressways. For example, using real-time vehicle location and speed information, a clustering algorithm is used to generate the real-time traffic flow density distribution of road segments; combined with traffic flow data from roadside sensors, a dynamic relationship model of flow rate-degree-degree is established.

[0074] Building upon this foundation, a short-term traffic prediction model based on intelligent connected vehicle data is constructed. This model uses fused traffic state data as input and employs an improved temporal neural network (such as the spatiotemporal Transformer) to fully exploit the spatiotemporal correlation features inherent in the intelligent connected vehicle data. The model not only captures the periodicity and trends of traffic flow using historical traffic data but also predicts future traffic flow trends through vehicle route planning information. The model outputs key indicators such as traffic flow, average vehicle speed, and congestion index for each road segment within the next 5-30 minutes, with a prediction accuracy exceeding 85% during peak hours.

[0075] Model Predictive Control (MPC): Based on traffic prediction results driven by intelligent connected vehicle data, a model predictive control framework is constructed to achieve coordinated control of ramps and the main line. Within each control cycle, multiple sets of candidate ramp signal control schemes are generated based on the current traffic conditions and future prediction results (e.g., green light duration is adjusted in 5-second increments within the range of 30-90 seconds).

[0076] For each candidate control scheme, its impact on traffic conditions within the future prediction time domain N (e.g., 5 control cycles) is simulated using a traffic flow simulation model (based on intelligent connected vehicle data calibration parameters). Within the prediction time domain, for each prediction step k, the cost function J is calculated. k : J k (X k U k ) equals ∑ i=0 N-1 (||X k+i|k Subtract X ref || 2 Q In addition ||U k+i|k || 2 R ) plus ||X k+N|k Subtract X ref || 2 P .

[0077] Among them, X k+i|k The system state at time k+i is predicted based on information from time k, X. ref The desired system state (e.g., smooth traffic flow), U k+i|k Let be the control input at time k + i (i.e., the ramp signal control scheme), N be the prediction time domain length, ||·|| denote the weighted Euclidean norm, and Q, R, and P be weight matrices used to adjust the importance of different state variables and control inputs in the cost function. The optimal control sequence U* that minimizes the cost function is found using an optimization algorithm. U* equals arg min U ∑ k=1 N J k (X k U k ).

[0078] Applying optimal control schemes to actual ramp signal control improves overall traffic efficiency on expressways by approximately 25% compared to traditional cooperative control strategies. It also allows for the prediction of control scheme effects in advance, effectively alleviating traffic congestion. Intelligent connectivity facilitates smoother data interaction during predictive control, enhancing the optimization efficiency and implementation effectiveness of control schemes.

[0079] In another way Figure 2 In the embodiment shown, Figure 2 This is a flowchart illustrating one embodiment of the traffic situation assessment and ramp collaborative control method of the present invention. A pilot deployment is being conducted on a ring expressway in a certain city. High-definition cameras with 8 megapixels are installed every 500 meters, supporting AI intelligent analysis and capable of real-time identification of vehicle type, license plate, speed, and other information. This is paired with millimeter-wave radar, with a detection range of up to 300 meters, accurately acquiring vehicle distance, speed, and angle data. Simultaneously, lidar is deployed to achieve 360° all-around environmental perception. These roadside devices are all equipped with 5G communication modules, uploading collected data to edge computing nodes at a rate of no less than 100Mbps via NSA (Non-Standalone) networking. On the vehicle side, newly registered operating vehicles are required to install onboard terminals supporting PC5 direct communication, enabling information interaction between vehicles and roadside devices, and between vehicles themselves.

[0080] Taking the data collection and processing during the morning rush hour (7:00-9:00) on a weekday as an example, roadside equipment and vehicle-mounted terminals collected multi-source heterogeneous data, including vehicle position (latitude and longitude coordinates), speed (km / h), and acceleration (m / s²) uploaded by vehicle-mounted sensors. 2 Real-time information such as driving direction (angle value); video stream data collected by high-definition cameras, and traffic flow and vehicle type distribution data obtained after algorithm analysis; point cloud data such as vehicle distance, speed, and angle collected by millimeter-wave radar and lidar; and real-time weather conditions (such as temperature, humidity, and whether it is raining) and road slippage data accessed by meteorological departments through intelligent connected systems.

[0081] During the data cleaning phase, the DBSCAN algorithm was used to process vehicle speed data. Data points with speeds exceeding twice the speed limit of a road segment (e.g., speeds greater than 240 km / h in a 120 km / h speed limit) were identified as abnormal and deleted. For jump anomalies in vehicle location data (e.g., sudden changes in location exceeding 1 km between adjacent time points), linear interpolation was used for repair. In the data integration and feature extraction phase, vehicle trajectory data was aggregated by road segment using a 1-minute time window. The average vehicle speed, traffic flow, headway, and other features of each road segment were calculated. Meteorological data was encoded, with numerical data such as temperature and humidity normalized to the [0, 1] interval. Weather conditions (sunny, rainy, snowy, etc.) were converted into one-hot encodings, ultimately resulting in a fused dataset containing traffic flow features, individual vehicle features, and meteorological environmental features.

[0082] Feature extraction structure and data application: The city ring road is divided into 20 road segments as nodes in a graph structure. An adjacency matrix A is constructed based on the connectivity between road segments. For spatial feature extraction, vehicle trajectory data is encoded to convert vehicle latitude and longitude coordinates, speed, acceleration, and other information into 16-dimensional feature vectors. Traffic flow data is directly used as 4-dimensional features of the nodes (including current flow, average flow over the past 5 minutes, average flow over the past 10 minutes, and average flow over the past 15 minutes). Meteorological data is encoded to form 8-dimensional feature vectors. These feature vectors of different modalities are concatenated to form the initial feature matrix H0 of each node. Through three layers of graph convolution operations, with each layer having a kernel size of 3×3, spatial correlation features between road segments are extracted step by step.

[0083] The training set was selected from the morning rush hour (7:00-9:00) data of the city's expressways over a continuous week, including multi-source heterogeneous data under an intelligent connected environment. For a given moment, the node features h of road segment A... A and the node characteristics h of road segment B B (Including queue length, vehicle arrival rate, real-time vehicle speed on the road segment, and flow rate percentage), Q is obtained through linear transformation. A equal to h A Multiplied by W q K B equal to h B Multiplied by W k V B equal to h B Multiplied by W v Calculate the association weight α. AB equals softmax(Q) A Multiply by K B T Divide by the square root of 64 and then multiply by γ AB Multiply by λ t If 22% of vehicles travel from A to B within 10 minutes, then γ AB =1.2, λ t =1.0. Repeatedly calculate the association weights between all nodes, and then aggregate them using the spatial feature aggregation formula h. i 'equals ∑ j α ij Multiply by h j Multiply by W v The fusion characteristics of each node are obtained.

[0084] For temporal feature extraction, a 5-minute time step is used, and the spatial feature sequence extracted by the graph convolutional network is used as the input to the LSTM network. The LSTM network has two layers, each containing 128 memory units. At a certain time t, the input gate i...t Based on the current input x t (i.e., the spatial feature vector at the current time step) and the hidden state h from the previous time step t-1 Calculate the inflow of information; forget gate f t Determine the memory unit C of the previous moment t-1 Which information needs to be retained; output gate o t Control the current memory unit C t Output to hidden state h t This information allows us to capture patterns in traffic conditions over time.

[0085] The spatial feature sequence output by DGAT (updated every 2 seconds) and the temporal feature sequence output by LSTM (with a time step of 5 minutes) are input into a dynamic weight pooling module. The features are further calibrated through a cross-modal bidirectional attention layer, and finally, the fused features are processed through a fully connected layer to output traffic state predictions (average vehicle speed, traffic flow) for the next 15 minutes. During training, the Adam optimizer is used with a learning rate of 0.001, and historical data from the past month for the city's ring road are used for training. By continuously adjusting the model parameters, the prediction results are made as close as possible to the actual traffic conditions.

[0086] The macro-level traffic flow control layer learning method employs a Deep Q-Network (DQN) as the learning algorithm. A 3-layer fully connected neural network is constructed as the Q-value approximator. The input consists of overall traffic situation features (such as one-hot encoding of congested road segment distribution and normalized values ​​of overall traffic flow trends, totaling a 20-dimensional feature vector), and the output is the Q-value corresponding to three actions (high, medium, and low traffic allocation schemes). During training, the exploration rate εepsilon is set from 0.9 and gradually decays to 0.1 with training iterations to balance exploring new strategies and utilizing existing strategies. Every 100 interactions (i.e., 100 time steps), the target network parameters are updated to the current network parameters. The reward function R... macro Equals 0.6 multiplied by Q total Divide by Q total-max Subtract 0.4 from that.

[0087] If at a certain moment, the total traffic flow Q total Reaching the highest historical traffic volume Q total-max 80% of the area is congested, and the area of ​​the congested zone is C. area The area of ​​the largest historical congestion zone, C area-max If the reward is 30%, then the reward value R is calculated. macro Equals (0.6 multiplied by A) ~ ||0.8 Subtract (0.4 multiplied by A) ~ || 0.3 The value is 0.36. By continuously interacting with the environment, it learns the optimal macro-flow control strategy.

[0088] The micro-level vehicle scheduling layer learning method, also based on the DQN algorithm, constructs a 4-layer fully connected neural network. The input is the real-time state features of a single ramp (including vehicle queue length, vehicle arrival rate, current traffic light status, etc., a total of 12-dimensional feature vectors). The output is the Q-value corresponding to 10 actions with different combinations of green and red light durations. During training, a discount factor of 0.9 is set to represent the importance of future rewards. The target network parameters are updated every 50 interactions. The reward function R... macro Equals (0.7 divided by D) avg Subtract (0.3 multiplied by L) queue Divide by L queue-max If the average vehicle delay time D at a certain moment avg =30 seconds, current queue length L queue The longest queue in history, L queue-max If it is 50%, then the reward value R macro Equals (0.7 multiplied by A) ~ || 1 / 30 Subtract (0.3 multiplied by A) ~ ||0.5) is approximately equal to -0.07. By continuously optimizing the strategy, vehicle delays and queue lengths can be reduced.

[0089] Transfer learning application: When deploying a model at a new ramp, 80% of the network parameters are transferred from a pre-trained ramp model with similar traffic conditions in the vicinity as the initial parameters of the new model. Then, fine-tuning is performed using a small amount of data from the new ramp (such as data from the previous 2 hours), and the remaining 20% ​​of parameters are updated through the backpropagation algorithm, accelerating the convergence speed of the new model. Compared with traditional reinforcement learning algorithms, this allows the new model to achieve stable control effects in a short time.

[0090] All data collection and extraction in this invention are carried out under compliant and legal conditions.

Claims

1. A traffic situation assessment and ramp coordinated control method, characterized in that, include: Collect multi-source heterogeneous data from vehicles and roadside equipment through intelligent connected vehicles; Based on multi-source data, road segment node features are constructed. Dynamic graph attention network is used to dynamically calculate the association weights between road segments, and long short-term memory network is used to extract traffic flow time series features. By dynamically weighting and pooling spatial correlation features with time series features and fusing cross-modal bidirectional attention, traffic situation assessment results are obtained. Using a model predictive control framework, candidate signal control schemes are continuously optimized based on traffic flow simulation models to generate and implement optimal ramp control commands.

2. The traffic situation assessment and ramp coordinated control method according to claim 1, characterized in that, The collection of multi-source heterogeneous data from vehicles and roadside equipment via intelligent network connectivity specifically includes: The system collects real-time vehicle information uploaded by onboard sensors, collects traffic flow, vehicle density and occupancy, and vehicle point cloud data through roadside equipment, and accesses real-time weather and road condition data. During data transmission, it adopts intelligent network encryption and authentication mechanisms and aggregates multi-source heterogeneous data to edge computing nodes or the cloud.

3. The traffic situation assessment and ramp coordinated control method according to claim 1, characterized in that, The method of dynamically calculating the inter-segment association weights using a dynamic graph attention network specifically includes: The preprocessed data of each road segment is used to construct the initial feature vector of the node. Through a learnable weight matrix, it is linearly transformed into query vector, key vector and value vector. The original association score between nodes is calculated and corrected by introducing a time decay factor based on data delay time and an interaction intensity factor based on vehicle transfer rate within a preset time window. The dynamic association weights are obtained by normalization using the softmax function. The value vectors of adjacent nodes are weighted and summed according to the association weights to obtain the spatial features of each road segment. The spatial features and association weights are updated at a preset period.

4. A traffic situation assessment and ramp coordinated control method according to claim 2 or 3, characterized in that, The extraction of traffic flow time series features using Long Short-Term Memory Networks includes: Using historical traffic flow data sequences as input, the Long Short-Term Memory Network controls the inflow of current input information through an input gate, controls the amount of information retained in the memory unit at the previous time step through a forget gate, and controls the amount of information output from the current memory unit to the hidden state through an output gate. It uses internal gating mechanisms and memory units to capture long-term dependencies and periodic change characteristics in traffic flow data.

5. The traffic situation assessment and ramp coordinated control method according to claim 4, characterized in that, The dynamic weight pooling and cross-modal bidirectional attention fusion include: Based on the scene perception factor calculated from the real-time congestion index and traffic fluctuation rate, the fusion weights of spatial and temporal features are dynamically allocated and weighted summation is performed to obtain preliminary fusion features. Prospective features are generated using path planning data. A gating vector jointly determined by the spatiotemporal fusion features is adopted. The injection ratio of the prospective features is calculated through the sigmoid activation function and then weighted and fused with the features processed by cross-modal bidirectional attention.

6. The traffic situation assessment and ramp coordinated control method according to claim 5, characterized in that, The model-based predictive control framework specifically includes: A hierarchical reinforcement learning architecture is constructed, comprising a macro-level traffic flow control layer and a micro-level vehicle scheduling layer. The macro-level layer takes traffic situation assessment results as state input and outputs ramp flow allocation schemes. Its reward function is based on total traffic flow and congestion area. The micro-level layer takes the queue length and arrival rate of vehicles at a single ramp as state input and outputs specific traffic light timing schemes. Its reward function is based on average vehicle delay and queue length.

7. The traffic situation assessment and ramp coordinated control method according to claim 6, characterized in that, The rolling optimization includes: In each control cycle, based on the current traffic conditions and the future multi-period traffic conditions predicted by intelligent connected vehicle data, multiple sets of candidate control schemes for ramp signals are generated. Using a traffic flow simulation model with parameters calibrated by intelligent connected vehicle data, the implementation effect of each candidate scheme is simulated within the set prediction time domain.

8. A traffic situation assessment and ramp coordinated control method according to claim 1 or 6, characterized in that, The model-based predictive control framework further includes: When training the control model for the new ramp, a transfer learning mechanism is adopted to transfer a preset proportion of network parameters from the trained source model with similar traffic conditions to the target model for initialization. The parameters are then fine-tuned using the data from the new ramp, which accelerates the model convergence.

9. The traffic situation assessment and ramp coordinated control method according to claim 5, characterized in that, The preliminary fusion features specifically include: Using spatial features as queries and temporal features as keys and values, we calculate attention weights and extract key time segment features. Using time features as queries and spatial features as keys and values, we calculate attention weights and filter key spatial road segment features. The two types of features processed by bidirectional attention are concatenated and compressed through a fully connected layer to obtain the final fused features.

10. The traffic situation assessment and ramp coordinated control method according to claim 7, characterized in that, The rolling optimization also includes: A cost function is constructed, which measures the degree of deviation of the system state from the desired state and the magnitude of the control input. An optimization algorithm is used to find the optimal control sequence that minimizes the total cost in the prediction time domain. The first control command in the optimal control sequence is applied to the actual ramp signal control. In the next control cycle, prediction and optimization are performed again based on the latest state to achieve closed-loop feedback control in the rolling time domain.

Images