Multi-intersection inter-cluster coupling-time sequence attention multi-target robust cooperative regulation and control method

Abstract

The invention belongs to the field of intelligent traffic control, particularly relates to a multi-intersection inter-cluster coupling-time sequence attention multi-target robust coordinated regulation and control method, and aims to improve the global coordination of multi-intersection traffic signal control. Comprising the following steps: S1, acquiring multi-dimensional traffic related data of multiple intersections through a sensor network, and obtaining a standardized enhanced data set after missing data enhancement and complementation are realized through preprocessing and a generative adversarial network; s2, processing the standardized enhanced data set by adopting a density peak value-adaptive K value clustering algorithm to obtain data clusters, core intersections in each cluster and key associated intersections among the clusters; s3, calculating the comprehensive priority weight of each intersection through an inter-cluster coupling-time sequence attention weight model based on the related data of the core intersection and the inter-cluster key association intersection; and S4, in combination with the comprehensive priority weight, an improved multi-target robust particle swarm optimization algorithm is adopted, a multi-target robust optimization function is constructed, constraint conditions are introduced, and a traffic signal lamp timing scheme of each intersection is adjusted.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control, and in particular to a robust collaborative control method for multi-intersection cluster coupling-temporal attention multi-objective control. Background Technology

[0002] With the continuous acceleration of urbanization, urban traffic flow has experienced explosive growth, and traffic congestion has become a core bottleneck restricting urban operational efficiency. Traffic lights, as key facilities for regulating traffic flow, directly determine regional traffic efficiency through the scientific and coordinated nature of their timing schemes. Most existing traffic signal control methods focus on local optimization at single intersections, failing to fully consider the mutual influence and constraints between traffic flows at multiple intersections in the urban road network. This results in the optimization effect of a single intersection failing to radiate globally, and may even exacerbate congestion at adjacent intersections, thus failing to meet the global coordinated control needs of complex road networks.

[0003] Existing multi-intersection collaborative control technologies still suffer from several technical shortcomings that urgently need to be addressed: First, the clustering methods are rigid, often employing the traditional K-means algorithm, which requires manual pre-setting of the number of clusters and random selection of initial cluster centers, resulting in poor stability and insufficient adaptability of the clustering results, making it difficult to dynamically match real-time changes in traffic flow. Second, the weight allocation logic is one-sided, relying solely on static indicators such as traffic flow and congestion status, without effectively integrating dynamic coupling relationships between intersections (such as traffic flow overflow and signal timing conflicts) and temporal correlation features, leading to biased judgments of control priorities and insufficient fulfillment of control needs at key related intersections. Third, the collaborative optimization objectives are singular, often focusing solely on traffic efficiency as the core objective, neglecting the avoidance of traffic flow conflicts between clusters and the anti-interference capabilities of the control scheme. When faced with interference such as traffic flow fluctuations, sensor noise, and sudden traffic events, the control effect is prone to significant decline. Fourth, the feedback correction mechanism is rudimentary, adjusting parameters based on only a single indicator, failing to form a closed-loop, intelligent correction system, resulting in insufficient dynamic adaptability of the system to changes in traffic flow. Summary of the Invention

[0004] The purpose of this invention is to provide a robust collaborative control method for multi-intersection cluster coupling-temporal attention multi-objective control, which aims to improve the global coordination, dynamic adaptability and anti-interference capability of traffic signal control at multiple intersections.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a multi-intersection cluster coupling-temporal attention multi-objective robust collaborative control method, comprising: S1, collecting multi-dimensional traffic-related data of multiple intersections through a sensor network, and obtaining a standardized augmented dataset after preprocessing and augmentation of missing data using a generative adversarial network; S2, processing the standardized augmented dataset using a density peak-adaptive K-value clustering algorithm to obtain data clusters and core intersections within each cluster and key inter-cluster related intersections; S3, calculating the comprehensive priority weight of each intersection based on the relevant data of core intersections and key inter-cluster related intersections using an inter-cluster coupling-temporal attention weight model; S4, combining the comprehensive priority weight, constructing a multi-objective robust optimization function and introducing constraints using an improved multi-objective robust particle swarm optimization algorithm to adjust the traffic light timing scheme of each intersection.

[0006] Multi-dimensional traffic-related data includes historical traffic data, real-time traffic data, inter-cluster coupling data, and time-series correlation data. Among them, traffic data includes traffic flow, vehicle speed, maximum queue length of intersection traffic direction, and environmental variables; inter-cluster coupling data includes traffic flow overflow coefficient and signal timing conflict degree of adjacent intersections; and time-series correlation data includes traffic flow transmission delay and traffic flow dynamic coupling coefficient of adjacent intersections.

[0007] The preprocessing in step S1 includes outlier detection, adaptive moving average denoising, and normalization.

[0008] In step S2, the density peak-adaptive K-value clustering algorithm is used to process the standardized augmented dataset, including: calculating the local density and relative distance of samples, selecting initial cluster centers through adaptive threshold segmentation; dynamically adjusting the number of clusters based on intra-cluster similarity and inter-cluster separation indices; calculating the coupling weighted distance between samples and cluster centers, dividing samples into corresponding data clusters and updating cluster centers until the cluster structure is stable; wherein, a sample is a single data unit in the standardized augmented dataset, corresponding to a multi-dimensional traffic-related data set of a single intersection in a single sampling.

[0009] The identification criteria for core intersections and key inter-cluster related intersections include the intersection's traffic flow influence, geographical location importance, temporal correlation strength, and inter-cluster coupling contribution.

[0010] The comprehensive priority weights of the inter-cluster coupling-temporal attention weight model include traffic flow weight, inter-cluster coupling weight, geographic location weight, and temporal attention weight; among them, the temporal attention weight is calculated based on the Transformer self-attention mechanism and is used to highlight the salience of the intersection's association in the temporal dimension.

[0011] In step S4, the optimization objectives of the multi-objective robust optimization function include the shortest global average travel time, the fewest global congested intersections, the minimum inter-cluster traffic flow conflict, and the maximum anti-interference margin of the control scheme; the constraints include signal timing constraints and inter-cluster coupling constraints.

[0012] The improved multi-objective robust particle swarm optimization algorithm includes: introducing adaptive inertia weights to balance global search and local convergence capabilities; adopting an elite retention strategy based on non-dominated sorting to avoid premature convergence; and incorporating robust constraint processing to perform disturbance simulation tests on the optimization results and eliminate inferior solutions.

[0013] Step S5 is included after step S4: collect real traffic data after the timing scheme is adjusted, calculate the reward value based on the reinforcement learning reward mechanism, dynamically adjust the comprehensive priority weight, data cluster structure and optimization algorithm parameters according to the reward value, and return to step S3 to re-execute the weight allocation and subsequent collaborative optimization control to achieve continuous iteration.

[0014] When a sudden traffic incident is detected, the weight coefficient of the anti-interference margin target of the control scheme is automatically increased, and an emergency timing scheme is quickly generated; sudden traffic incidents include traffic accidents, extreme weather, and traffic anomalies caused by large-scale events.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application provides a robust collaborative control method for multi-intersection cluster coupling and temporal attention, employing a multi-dimensional traffic-related data collection model to comprehensively cover historical and real-time traffic conditions, dynamic coupling relationships between intersections, and temporal correlation features, ensuring the integrity and relevance of data dimensions. A combined preprocessing process of outlier detection, adaptive moving average denoising, and normalization effectively filters data noise and improves data consistency. Simultaneously, a generative adversarial network is introduced to augment missing data, avoiding data distortion problems caused by traditional interpolation methods. The construction of a standardized augmented dataset provides high-quality and complete data support for subsequent clustering, weight calculation, and optimization control, ensuring the accuracy of decisions at each stage. A density peak-adaptive K-value clustering algorithm is used, selecting initial cluster centers through adaptive threshold segmentation, avoiding the poor adaptability problem caused by manually pre-setting the number of clusters in traditional clustering algorithms. The number of clusters is dynamically adjusted based on intra-cluster similarity and inter-cluster separation, enabling the data cluster structure to match the dynamic changes in traffic flow in real time. The coupling weighted distance is calculated using multi-dimensional features to ensure that the clustering results reflect the actual correlation strength between intersections. Meanwhile, the identification mechanism for core intersections and key intersections between clusters, based on traffic flow influence, geographical location importance, temporal correlation strength, and inter-cluster coupling contribution, significantly improves the identification accuracy of key control nodes, providing a scientific basis for subsequent precise control and solving the shortcomings of traditional fixed clustering and one-sided key intersection identification.

[0016] 2. This application provides a robust collaborative control method for multi-intersection cluster coupling-temporal attention multi-objectives. By constructing an inter-cluster coupling-temporal attention weight model, it overcomes the limitations of traditional weight allocation that relies solely on static indicators. It integrates traffic flow weight (reflecting traffic volume), inter-cluster coupling weight (reflecting dynamic interactions between intersections), geographic location weight (highlighting static importance), and temporal attention weight (capturing temporal correlation salience), achieving a comprehensive consideration of static features and dynamic correlations, as well as spatial and temporal dimensions. Specifically, the temporal attention weight design based on the Transformer self-attention mechanism can accurately uncover the correlation patterns of intersections across different time dimensions, further improving the accuracy of control priority judgment. The comprehensive priority weight calculated by this model ensures that key nodes such as core intersections and highly coupled intersections receive priority control, avoiding waste of control resources and significantly improving the allocation efficiency of control resources.

[0017] 3. The method provided by this invention constructs a multi-objective robust optimization function, simultaneously considering four objectives: shortest global average travel time, fewest global congested intersections, minimum inter-cluster traffic flow conflict, and maximum anti-interference margin of the control scheme. This overcomes the limitations of traditional single or dual-objective optimization, achieving a synergistic improvement in traffic efficiency, congestion mitigation, conflict avoidance, and anti-interference capability. Combined with an improved multi-objective robust particle swarm optimization algorithm, it balances global search and local convergence capabilities through adaptive inertial weights, avoids premature convergence based on an elite retention strategy using non-dominated sorting, and incorporates robustness constraints to eliminate inferior solutions with insufficient anti-interference capability. This ensures that the optimized traffic light timing scheme possesses both optimal comprehensive control effect and resistance to interference from traffic fluctuations and sensor noise. Furthermore, for emergency response mechanisms to sudden traffic events, by automatically increasing the weight coefficient of the anti-interference margin objective, emergency timing schemes can be quickly generated, significantly reducing the risk of control failure in emergency scenarios and further enhancing the stability and reliability of the scheme. Attached Figure Description

[0018] Figure 1 This is a flowchart of a multi-channel cluster coupling-temporal attention multi-objective robust collaborative control method provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 This application provides a multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method, including: S1. Collect multi-dimensional traffic-related data from multiple intersections through a sensor network. After preprocessing and augmentation of missing data using a generative adversarial network, a standardized augmented dataset is obtained.

[0021] By deploying a sensor network at various intersections, multi-dimensional traffic-related data is collected at sampling intervals of T. For example, the multi-dimensional traffic-related data includes historical traffic data, real-time traffic data, inter-cluster coupling data, and time-series correlation data. Specifically, the traffic data includes traffic flow, vehicle speed, maximum queue length in each direction of traffic at the intersection, and environmental variables; the inter-cluster coupling data includes traffic flow overflow coefficients and signal timing conflict degrees between adjacent intersections; and the time-series correlation data includes traffic flow propagation delays and dynamic coupling coefficients of traffic flow between adjacent intersections.

[0022] The preprocessing in step S1 includes outlier detection, adaptive moving average denoising, and normalization of the collected data based on the 3σ criterion. Then, a generative adversarial network (GAN) is used to augment and complete the missing data to obtain a standardized augmented dataset. The standardized augmented dataset contains N samples, each represented as... ,in Number the intersections. , (Total number of intersections participating in collaborative optimization). The first of the day Second sampling, For the number of sampling days, For the first The first intersection Heavenly Traffic flow from the second sampling For the first The first intersection Heavenly The vehicle speed at the next sampling point, For the first The first intersection Heavenly Maximum queue length per sampling For the first Environmental variables of the day, For the first Traffic flow propagation delay between each intersection and adjacent intersections, For the first The dynamic coupling coefficient of traffic flow between an intersection and its adjacent intersections. For the first Traffic overflow coefficient between each intersection and its adjacent intersections. For the first Signal timing conflict between an intersection and its adjacent intersections.

[0023] Step S1 involves collecting historical traffic data, real-time traffic data, inter-cluster coupling data, and time-series correlation data from multiple intersections using a multi-type sensor network, achieving comprehensive coverage of data dimensions. Traffic data includes traffic flow, vehicle speed, maximum queue length, and environmental variables (weather, holidays, etc.); inter-cluster coupling data includes traffic flow overflow coefficients and signal timing conflict degrees, reflecting the dynamic mutual influence between intersections; time-series correlation data includes traffic flow propagation delay and traffic flow dynamic coupling coefficients, capturing the time-series correlation characteristics between intersections. The collected data undergoes outlier detection (3σ criterion), adaptive moving average denoising, and normalization. To address data missingness issues, a generative adversarial network (GAN) is used for augmentation and completion, avoiding data distortion caused by traditional interpolation methods, and providing a high-quality, highly complete dataset for subsequent clustering and optimization.

[0024] S2. The density peak-adaptive K-value clustering algorithm is used to process the standardized augmented dataset to obtain data clusters and core intersections within each cluster, as well as key interconnected intersections between clusters.

[0025] Dynamic clustering of the normalized augmented dataset is achieved using the density peak-adaptive K-value clustering algorithm, resulting in K adaptively adjusted data clusters. For example, step S2, processing the normalized augmented dataset using the density peak-adaptive K-value clustering algorithm, includes: calculating the local density of samples based on the density peak clustering algorithm. and relative distance , build The decision graph uses adaptive threshold segmentation (based on the maximum inter-class variance method) to select initial cluster centers, avoiding poor cluster fit caused by manually setting the K value. The number of clusters K is dynamically adjusted based on intra-cluster similarity and inter-cluster separation indices; the intra-cluster similarity calculation formula is as follows: In the formula, For the first Clusters, The number of samples within the cluster. As the cluster center, The weighted distance is used. The formula for calculating the inter-cluster separation index is as follows: when Time-split cluster; when At that time, merge clusters. Among them, To preset the adaptation threshold, For example, a preset merging threshold is used. , .

[0026] Calculate the weighted coupling distance between the sample and the cluster center, divide the sample into the corresponding data cluster and update the cluster center until the cluster structure is stable; where the sample is a single data unit in the standardized augmented dataset, corresponding to a multi-dimensional traffic-related data set of a single intersection in a single sampling.

[0027] It should be understood that calculating the weighted coupling distance between a sample and a cluster center means calculating the weighted coupling distance between each sample and each cluster center. The formula for calculating the weighted coupling distance is as follows: in, These are the weighting coefficients for traffic flow, vehicle speed, queue length, traffic flow dynamic coupling coefficient, and traffic flow overflow coefficient, respectively. ), Samples The corresponding eigenvalues, Cluster centers The corresponding eigenvalues.

[0028] The sample is assigned to the data cluster corresponding to the nearest cluster center. After updating the weighted average of each data cluster, the two operation steps of dynamically adjusting the number of clusters K and calculating the coupling weighted distance between each sample and each cluster center based on the intra-cluster similarity and inter-cluster separation index are repeated until the cluster structure is stable.

[0029] For example, when the change in cluster center is less than a preset threshold When the cluster structure is stable, output the K data clusters at that time.

[0030] For example, the identification criteria for core intersections and key inter-cluster intersections include the intersection's traffic flow influence, geographical location importance, temporal correlation strength, and inter-cluster coupling contribution. Based on the traffic flow influence (weight 0.3), geographical location importance (weight 0.2), temporal correlation strength (weight 0.3), and inter-cluster coupling contribution (weight 0.2) of intersections within a cluster, core intersections within each cluster and key inter-cluster intersections are identified.

[0031] To address the shortcomings of traditional clustering algorithms, this application proposes a density peak-adaptive K-value clustering scheme. First, the local density and relative distance of samples are calculated using the density peak clustering algorithm to construct a decision map. Initial cluster centers are selected through adaptive threshold segmentation to ensure representativeness. Then, intra-cluster similarity and inter-cluster separation indices are defined, and the number of clusters K is dynamically adjusted (by splitting or merging clusters) to achieve dynamic adaptation of the clustering structure to traffic flow conditions. When calculating the weighted coupling distance between samples and cluster centers, multi-dimensional features such as traffic flow, vehicle speed, queue length, dynamic coupling coefficient of traffic flow, and traffic overflow coefficient are incorporated to highlight the impact of coupling correlation features on the clustering results. Finally, K stable data clusters are output, and core intersections within each cluster (playing a dominant role in intra-cluster traffic flow) and key inter-cluster intersections (playing a crucial role in inter-cluster traffic flow transmission) are identified, providing a clustering basis for precise coordinated control.

[0032] S3. Based on the relevant data of core intersections and key inter-cluster related intersections, calculate the comprehensive priority weight of each intersection through the inter-cluster coupling-temporal attention weight model.

[0033] By collecting real-time traffic status data, inter-cluster coupling data, and time-series correlation data of core intersections within each cluster and key inter-cluster related intersections, an inter-cluster coupling-time-series attention weight model is constructed to calculate the comprehensive priority weight of each intersection. The weight value is positively correlated with the degree of impact of the intersection on global traffic, the strength of inter-cluster coupling, and the significance of time-series correlation. Intersections with high weight values ​​are prioritized for regulation.

[0034] For example, the comprehensive priority weights of the inter-cluster coupling-temporal attention weight model include traffic flow weights (reflecting the traffic volume at the intersection), inter-cluster coupling weights (reflecting the dynamic coupling strength between the intersection and adjacent intersections, integrating the dynamic coupling coefficient of traffic flow, the traffic overflow coefficient, and the signal timing conflict degree), geographical location weights (reflecting the static importance of the intersection), and temporal attention weights. Among these, the temporal attention weights are calculated based on the Transformer self-attention mechanism to highlight the significant temporal correlation of intersections. By calculating the comprehensive priority weights of each intersection using the inter-cluster coupling-temporal attention weight model, the control priorities are clarified, ensuring that core intersections and key related intersections receive priority control, thereby improving the efficiency of control resource allocation.

[0035] For example, the specific calculation method for inter-cluster coupling-temporal attention weight allocation is as follows: Among them, W i For comprehensive priority weighting; , , , For example, the weighting coefficients are shown. =0.25、 =0.3、 =0.2、 =0.25. Traffic flow weighting, For inter-cluster coupling weights, For geographical location weight, Temporal attention weight; traffic flow weight in, For the first Real-time traffic flow at each intersection This represents the average real-time traffic flow at all intersections within the area.

[0036] Inter-cluster coupling weight in, For the first The number of adjacent intersections at each intersection. For the first The set of adjacent intersections of a given intersection. For the first The intersection and the first The dynamic coupling coefficient of traffic flow at each intersection. For the first The intersection and the first Traffic overflow coefficient at each intersection For the first The intersection and the first Signal timing conflict at each intersection.

[0037] Geographical location weight The weighted sum is calculated based on whether the intersection is a transportation hub (0.8-1.0 if yes, 0.3-0.7 otherwise) and whether it connects to a major traffic artery (0.8-1.0 if yes, 0.3-0.7 otherwise), with each weighting being 0.5. The value range is [0.3, 1.0]. Temporal attention weights are also considered. Computation based on Transformer's self-attention mechanism: in, The cosine similarity function is used. For the first Real-time feature vectors of each intersection, This is the regional traffic feature matrix for the previous sampling period, used to highlight the significant correlation of intersections in the temporal dimension.

[0038] S4. Combining comprehensive priority weights, an improved multi-objective robust particle swarm optimization algorithm is adopted to construct a multi-objective robust optimization function and introduce constraints to adjust the traffic signal timing schemes at each intersection.

[0039] For example, in step S4, the optimization objectives of the multi-objective robust optimization function include the shortest global average travel time, the fewest global congested intersections, the minimum inter-cluster traffic flow conflict, and the maximum anti-interference margin of the control scheme; the constraints include signal timing constraints (minimum green light duration, maximum cycle duration, phase switching interval) and inter-cluster coupling constraints.

[0040] The improved multi-objective robust particle swarm optimization algorithm includes: introducing adaptive inertia weights to balance global search and local convergence capabilities; adopting an elite retention strategy based on non-dominated sorting to avoid premature convergence; and incorporating robust constraint processing to perform disturbance simulation tests on the optimization results and eliminate inferior solutions.

[0041] As one possible implementation, the specific implementation process of the improved multi-objective robust particle swarm optimization algorithm is as follows: Optimization variable definition: based on the traffic light cycle at each intersection. , ( (seconds), green light duration for each phase , ( For phase numbering, Using seconds as the optimization variable, construct particle encoding vectors. ,in For the first The number of phases at each intersection; Constructing a multi-objective robust optimization function: in , , , For the target weight coefficient ( =0.35、 _2=0.25、 =0.2、 =0.2); The global average travel time (in seconds). , For timing scheme Next Queue length at each intersection For traffic speed, This represents the average waiting time. The total number of congested intersections (unit: number), with congestion determined by a road occupancy rate >60%. The number of traffic flow conflicts between clusters (unit: times / hour) is determined by the conflict judgment criteria: the overlap time of green light switching between adjacent intersections is >3 seconds or the traffic flow overflow is >th threshold. To ensure the interference immunity margin of the control scheme, , These are interference variables, including flow rate fluctuations of ±20% and sensor noise interference. The deviation of the objective function under disturbance.

[0042] Introducing adaptive inertia weights ,in This represents the current iteration number. The maximum number of iterations, , , , For the first Average fitness of generation of particles The historical average fitness To balance global search and local convergence (using fitness standard deviation), an elite retention strategy based on non-dominated sorting is adopted, retaining non-dominated solutions from each generation to form an external archive set. The optimal solution is then selected using crowding distance to avoid premature convergence. Robust constraint handling is incorporated, and disturbance simulation tests are performed on the particles to eliminate interference margins. The inferior solution.

[0043] Constraints: , For the first (Total phase switching interval at each intersection) ( Second), Second, ( (Seconds to avoid sudden changes in green light duration).

[0044] Step S4 constructs a multi-objective robust optimization function that aims to minimize the global average travel time, the number of globally congested intersections, inter-cluster traffic flow conflicts, and the interference resistance margin of the control scheme. This function balances traffic efficiency, congestion mitigation, conflict avoidance, and interference resistance. An improved multi-objective robust particle swarm optimization algorithm is used to solve for the optimal timing scheme. Adaptive inertia weights (dynamically adjusted based on fitness distribution) balance global search and local convergence. An elite retention strategy based on non-dominated ranking avoids premature convergence, and robustness constraints are incorporated to eliminate inferior solutions with insufficient interference resistance margin. Simultaneously, signal timing constraints (minimum green light duration, maximum cycle time, etc.) and inter-cluster coupling constraints are introduced to ensure the feasibility and coordination of the timing scheme.

[0045] Step S4 is followed by step S5: collecting real traffic data after the timing scheme is adjusted, calculating reward value based on reinforcement learning reward mechanism, dynamically correcting comprehensive priority weight, data cluster structure and optimization algorithm parameters according to reward value, and returning to step S3 to re-execute weight allocation and subsequent collaborative optimization control to achieve continuous iteration.

[0046] Step S5 collects real-time traffic data from each intersection after adjustments, such as travel time, congestion relief, number of inter-cluster conflicts, and control deviations under sudden disturbances. Based on the reinforcement learning reward mechanism (the reward value is positively correlated with the control effect and negatively correlated with the disturbance deviation), the comprehensive priority weight, cluster structure, and parameters of the collaborative optimization algorithm for each intersection are dynamically corrected. Then, the system returns to step S3 to re-execute the weight allocation and subsequent collaborative optimization control, thereby completing the dynamic closed-loop correction and iteration to ensure that the system maintains good control effect and stability in the long term.

[0047] When a sudden traffic incident is detected, the weight coefficient of the anti-interference margin target of the control scheme is automatically increased, and an emergency timing scheme is quickly generated; sudden traffic incidents include traffic accidents, extreme weather, and traffic anomalies caused by large-scale events.

[0048] As one possible implementation, the reinforcement learning reward mechanism in step S5 is designed as follows: Reward Value ,in =0.3、 =0.25、 =0.2、 =0.25\), For the first The reduction in average travel time per iteration Based on the standard travel time, To reduce the number of congested intersections, Based on the number of traffic jams, This represents the reduction in the number of inter-cluster conflicts. Based on the number of collisions, The objective function deviation under disturbance. The baseline objective function value is denoted as .

[0049] Based on reward value Adjust the overall priority weighting coefficient , , , ,when ( When the regulation effect is poor, increase the inter-cluster coupling weight coefficient. Reduce the geographic location weighting coefficient to 0.35. Up to 0.15; when ( When the control effect is good, the weight coefficients remain unchanged. Every 30 minutes, recalculate intra-cluster similarity and inter-cluster separation, adjust the cluster structure according to the rules, and update core intersections and key associated intersections. Dynamically adjust the learning factor of the particle swarm optimization algorithm based on the iterative convergence speed. , The convergence speed is too slow ( When the rate of change in fitness is <0.5%, adjust the learning factor. =2.5、 =1.5 to enhance global search; convergence speed is too fast ( When the rate of change in fitness is >5%, adjust the learning factor. , To enhance local convergence.

[0050] For example, embodiments of this application also provide a multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control system, including: The multi-dimensional data acquisition and enhancement preprocessing module, through a sensor network (flow sensor, speed sensor, queue length detector, environmental sensor, inter-cluster coupling monitoring module, and time-series correlation acquisition module) deployed at various intersections, collects historical traffic data, real-time traffic data, inter-cluster coupling data, and time-series correlation data from multiple intersections at sampling intervals of T. It then performs outlier detection, adaptive moving average denoising, normalization, and GAN enhancement and completion on the data, outputting a standardized enhanced dataset, thus fulfilling the content of step S1 in the aforementioned method.

[0051] The density peak-adaptive K-value clustering module receives the standardized augmented dataset and processes it using the density peak-adaptive K-value clustering algorithm. More specifically, it determines the initial cluster centers through density peak clustering, calculates the coupling weighted distance for clustering, and adaptively adjusts the number of clusters K based on intra-cluster similarity and inter-cluster separation. It outputs K stable data clusters and information on core intersections within clusters and key inter-cluster connections, thus realizing the content of step S2 in the above method.

[0052] The inter-cluster coupling-temporal attention weight allocation module collects traffic status data, inter-cluster coupling data, and temporal correlation data of each intersection in real time. It calculates the comprehensive priority weight of each intersection through the inter-cluster coupling-temporal attention weight model and outputs the control priority ranking, which is to realize the content of step S3 in the above method.

[0053] The multi-objective robust collaborative optimization control module combines comprehensive priority weights and cluster information to construct a multi-objective robust optimization function. It then employs an improved multi-objective robust particle swarm optimization algorithm to solve for the optimal timing scheme, which is equivalent to executing step S4 in the aforementioned method. The multi-objective robust collaborative optimization control module includes a built-in timing scheme conflict detection and robustness testing unit. It detects timing conflicts between clusters and triggers local adjustments, eliminating timing schemes with insufficient anti-interference margins.

[0054] The dynamic closed-loop correction and iteration module collects real traffic data after adjustments at each intersection in real time. Based on the reinforcement learning reward mechanism, it dynamically corrects the comprehensive priority weight, cluster structure, and collaborative optimization algorithm parameters, driving the system to return to the inter-cluster coupling-temporal attention weight allocation module for continuous iterative optimization, which is the content of step S5 in the above method.

[0055] The data storage and visualization module stores raw data, preprocessed data, clustering results, weight calculation data, timing schemes, feedback data, and iterative optimization process data, and provides data query, statistical analysis, and visualization display functions for control effects.

[0056] For example, this application also provides specific embodiments: The implementation process of the multi-intersection traffic signal coordinated control method involves multi-dimensional data collection and enhanced preprocessing: Thirty adjacent intersections in a core urban area were selected as the objects of coordinated optimization. Traffic flow sensors, speed sensors, queue length detectors, environmental sensors, inter-cluster coupling monitoring modules, and time-series correlation acquisition modules were deployed. Sampling intervals of T=5 minutes were used to collect historical traffic data, real-time traffic data, inter-cluster coupling data, and time-series correlation data from the past 30 days. Traffic data included traffic flow (vehicles / hour), vehicle speed (km / h), maximum queue length (meters), and environmental variables (weather: sunny / rainy / snowy, holidays: yes / no) for each intersection. Inter-cluster coupling data included traffic overflow coefficients of adjacent intersections (based on queue length). The data includes: intersection capacity calculation and signal timing conflict degree (calculated based on green light switching overlap time / 5-second threshold); time-series correlation data includes traffic flow transmission delay of adjacent intersections (calculated based on GPS trajectory data) and traffic flow dynamic coupling coefficient (calculated based on traffic flow change Pearson correlation coefficient); outliers are removed from the collected data using the 3σ criterion, and noise is denoised by adaptive moving average (window size 5), and normalized to the [0,1] interval using minimum-maximum normalization. GAN augmentation is used to complete missing data (generator is a 3-layer MLP, discriminator is a 2-layer MLP) to obtain a standardized augmented dataset D, which contains 30×30×288 samples (288 samples per intersection per day).

[0057] Density Peak-Adaptive K-Score Clustering Analysis: Initiate the density peak-adaptive K-score clustering process to calculate the local density of each sample. and relative distance , build The decision graph uses the Otsu's method to determine an adaptive threshold and selects five initial cluster centers; intra-cluster similarity is then calculated. =0.25, inter-cluster separation =0.8, =0.31 (between =0.3 and = between 0.7), the number of clusters K is kept at 5; set the coupling weighted distance weight coefficient. =0.25、 =0.2、 =0.2、 =0.2、 =0.15, calculate the weighted coupling distance between each sample and the cluster center, assign the sample to the corresponding cluster and update the cluster center, iterate until the change in the cluster center is <0.005, and output 5 stable data clusters; based on traffic flow influence (0.3), geographical location importance (0.2), temporal correlation strength (0.3), and inter-cluster coupling contribution (0.2), 2-3 core intersections are identified in each cluster, and 8 key inter-cluster correlation intersections are identified in the whole area.

[0058] Inter-cluster coupling - temporal attention weight allocation: Real-time traffic flow, inter-cluster coupling data, and temporal correlation data of each intersection are collected in real time, and weight coefficients are set. =0.25、 =0.3、 =0.2、 =0.25, calculate the comprehensive priority weight of each intersection; for example, if the real-time traffic flow of a core intersection is 1200 vehicles / hour and the average real-time traffic flow of the area is 800 vehicles / hour, then =1200 / 800=1.5; There are 4 adjacent intersections, with an average traffic flow dynamic coupling coefficient of 0.8, an average traffic flow overflow coefficient of 0.3, and an average signal timing conflict degree of 0.1. =(0.8×0.3×(1-0.1)) / 4=0.054; This intersection is a regional transportation hub and connects to major arterial roads. =0.8\cdot0.5+0.9×0.5=0.85; Calculate the temporal attention weights based on the Transformer self-attention mechanism. =0.12; Overall priority weight =0.25×1.5+0.3×0.054+0.2×0.85+0.25×0.12=0.375+0.0162+0.17+0.03=0.5912, ranking among the top 5 in regulatory priority.

[0059] Multi-objective robust collaborative optimization control: setting weight coefficients for the multi-objective robust optimization function =0.35、 =0.25、 =0.2、 =0.2, with the traffic light cycle (50-150 seconds) and green light duration (15-60 seconds) of each phase at each intersection as optimization variables; the parameters of the improved particle swarm optimization algorithm are set as follows: population size 60, maximum number of iterations 120, and initial inertia weight. =0.9, terminating inertia weight =0.4, initial value of learning factor =2.0、 =2.0; The search capability is adjusted by adaptive inertia weight, and elite solutions are selected based on non-dominated sorting and congestion distance. The particles are subjected to interference tests with traffic fluctuations of ±20%, and inferior solutions with anti-interference margin <0.6 are eliminated. The objective function is weighted by the comprehensive priority weight of each intersection to obtain the optimal timing scheme. The green light duration of the top 10 intersections in comprehensive priority is extended by 10-15 seconds compared to other intersections, and the phase switching interval of key related intersections between clusters is extended to 5 seconds to avoid conflicts.

[0060] Dynamic closed-loop correction and iteration: Real-time collection of actual traffic data at each intersection after adjustments, and calculation of reward values. =0.3×(200-130) / 200 + 0.25×(12-5) / 12 + 0.2×(18-7) / 18 + 0.25×(1-0.08)=0.3×0.35 + 0.25×0.583 + 0.2×0.611 + 0.25×0.92=0.105+0.146+0.122+0.23=0.603> =0.3, keeping the weight coefficient unchanged; recalculate intra-cluster similarity and inter-cluster separation every 30 minutes, with no adjustment to the cluster structure; the iteration convergence speed is normal (fitness change rate 2%), maintaining the learning factor. =2.0、 =2.0; Iteration optimization is completed every 15 minutes to continuously optimize the timing scheme.

[0061] By applying the methods and systems provided in the embodiments of this application, the global average travel time in the target area is reduced by 38%, the number of congested intersections is reduced by 42%, the number of traffic flow conflicts between clusters is reduced by 55%, the anti-interference margin of the control scheme reaches 0.82, and the control effect deviation is only 8% in the event of a sudden traffic accident. The overall traffic operation efficiency and stability of the area are significantly improved, effectively alleviating the traffic congestion problem.

[0062] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0063] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A robust collaborative control method for multi-objective inter-cluster coupling and temporal attention at multiple intersections, characterized in that: include: S1. Collect multi-dimensional traffic-related data from multiple intersections through a sensor network. After preprocessing and using a generative adversarial network to augment and complete missing data, a standardized augmented dataset is obtained. S2. The standardized augmented dataset is processed using the density peak-adaptive K-value clustering algorithm to obtain data clusters and core intersections within each cluster and key related intersections between clusters; S3. Based on the relevant data of core intersections and key related intersections between clusters, the comprehensive priority weight of each intersection is calculated using the inter-cluster coupling-temporal attention weight model; S4. Combining the comprehensive priority weight, an improved multi-objective robust particle swarm optimization algorithm is used to construct a multi-objective robust optimization function and introduce constraints to adjust the traffic light timing scheme of each intersection.

2. The multi-channel cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 1, characterized in that: The multi-dimensional traffic-related data includes historical traffic data, real-time traffic data, inter-cluster coupling data, and time-series correlation data. The traffic data includes traffic flow, vehicle speed, maximum queue length in the direction of traffic at intersections, and environmental variables. The inter-cluster coupling data includes traffic flow overflow coefficients and signal timing conflict degrees of adjacent intersections. The time-series correlation data includes traffic flow propagation delay and traffic flow dynamic coupling coefficients of adjacent intersections.

3. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 1, characterized in that, The preprocessing in step S1 includes outlier detection, adaptive moving average denoising, and normalization.

4. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 1, characterized in that, In step S2, the density peak-adaptive K-value clustering algorithm is used to process the standardized augmented dataset, including: calculating the local density and relative distance of samples, selecting initial cluster centers through adaptive threshold segmentation; dynamically adjusting the number of clusters based on intra-cluster similarity and inter-cluster separation indices; calculating the coupling weighted distance between samples and cluster centers, dividing samples into corresponding data clusters and updating cluster centers until the cluster structure is stable; wherein, the sample is a single data unit in the standardized augmented dataset, corresponding to a multi-dimensional traffic-related data set of a single intersection in a single sampling.

5. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 1, characterized in that, The identification criteria for core intersections and key inter-cluster related intersections include the intersection's traffic flow influence, geographical location importance, temporal correlation strength, and inter-cluster coupling contribution.

6. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 1, characterized in that, The comprehensive priority weights of the inter-cluster coupling-temporal attention weight model include traffic flow weight, inter-cluster coupling weight, geographic location weight, and temporal attention weight; wherein, the temporal attention weight is calculated based on the Transformer self-attention mechanism and is used to highlight the salience of the intersection's association in the temporal dimension.

7. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 1, characterized in that, In step S4, the optimization objectives of the multi-objective robust optimization function include the shortest global average travel time, the fewest global congested intersections, the minimum inter-cluster traffic flow conflict, and the maximum anti-interference margin of the control scheme; the constraints include signal timing constraints and inter-cluster coupling constraints.

8. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to any one of claims 1-7, characterized in that, The improved multi-objective robust particle swarm optimization algorithm includes: introducing adaptive inertia weights to balance global search and local convergence capabilities; adopting an elite retention strategy based on non-dominated sorting to avoid premature convergence; and incorporating robust constraint processing to perform disturbance simulation tests on the optimization results and eliminate inferior solutions.

9. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 1, characterized in that, Step S4 is followed by step S5: collecting real traffic data after the timing scheme is adjusted, calculating the reward value based on the reinforcement learning reward mechanism, dynamically correcting the comprehensive priority weight, data cluster structure and optimization algorithm parameters according to the reward value, and returning to step S3 to re-execute the weight allocation and subsequent collaborative optimization control to achieve continuous iteration.

10. The multi-inter-cluster coupling-temporal attention multi-objective robust collaborative control method according to claim 7, characterized in that, When a sudden traffic incident is detected, the weight coefficient of the anti-interference margin target of the control scheme is automatically increased, and an emergency timing scheme is quickly generated; the sudden traffic incident includes traffic accidents, extreme weather, and traffic anomalies caused by large-scale events.

Images