Traffic control subarea dynamic division method based on congestion evolution

Abstract

The invention belongs to the technical field of intelligent traffic information, and discloses a traffic control subarea dynamic division method based on congestion evolution, which comprises the following steps of: firstly, based on three types of similarity measurement of a space structure, traffic characteristics and functional attributes, adopting an improved Chameleon algorithm and combining with an optimal subarea quantity determination strategy, clustering road sections of a whole network, and determining the optimal subarea quantity; generating a plurality of initial homogeneous proton regions; thirdly, inputting the feature sequence of each initial subarea and the adjacent matrix into a space-time prediction framework, outputting a future congestion probability through model training, and constructing a prediction association graph according to the future congestion probability to provide space-time dependence information for subsequent dynamic division; and finally, the improved Chameleon algorithm is applied again on the basis of the prediction association graph, reconstruction and feedback iteration are carried out on the sub-regions, the homogeneity in the regions, the difference between adjacent regions and the boundary continuity are optimized, a sub-region division result and a corresponding macroscopic fundamental graph are finally output, and decision support can be provided for refined traffic control.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation information technology, specifically relating to a dynamic division method for traffic control sub-regions based on congestion evolution, which can be effectively applied to the dynamic division problem of traffic control sub-regions in large heterogeneous urban road network scenarios. Background Technology

[0002] With the acceleration of urbanization and the improvement of residents' living standards, the number of motor vehicles in cities is growing exponentially, and the traffic load and pressure on urban road networks are continuously rising. Problems such as traffic congestion, traffic accidents, air and noise pollution, and fuel consumption are becoming increasingly prominent. According to statistics from the World Health Organization, more than 1.25 million people die and 50 million are injured in traffic accidents each year, resulting in global economic losses of nearly US$1.85 trillion. Urban traffic congestion not only wastes a lot of public time but also significantly increases vehicle exhaust emissions, worsens urban air quality, and seriously restricts public health and socio-economic development. Under such a severe traffic situation, signal control has become a key means to improve the efficiency and safety of road networks. However, urban road networks are vast and complex, and traffic flows are random, dynamic, and diverse. Traditional overall signal optimization algorithms are often difficult to deploy on a large scale due to overly complex models or insufficient real-time data. To reduce the difficulty of control and improve system scalability, researchers have proposed dividing the urban road network into several traffic control sub-zones and achieving regional coordinated control through the collaborative optimization of intersections within these sub-zones.

[0003] Currently, traffic sub-zone division methods are mainly divided into two categories: static division and dynamic division. Static division maintains the sub-zone boundaries unchanged throughout the entire control cycle. Its planning is mostly based on fixed road network topology, road segment functions, or administrative regions, combined with empirically set flow thresholds for partitioning. Although this type of method has the advantages of simple implementation and low computational overhead, due to the fixed boundaries, it cannot respond in a timely manner to sudden changes and daily fluctuations in traffic flow, significantly reducing the effectiveness of coordinated control. Dynamic division methods adjust the sub-zone boundaries during operation through timed or threshold-based triggering mechanisms to adapt to flow fluctuations; however, most dynamic strategies rely only on a single traffic indicator (such as flow, speed, or queue length) or network structure (such as connectivity, centrality), making it difficult to comprehensively reflect the directional congestion transmission and functional correlation between road segments. Moreover, the update frequency needs to be manually set, and excessively frequent updates can lead to drastic changes in the shape of sub-zones, increasing computational and communication overhead, and potentially causing instability in the control strategy.

[0004] In existing technologies, static sub-zone division schemes keep the sub-zone boundaries unchanged throughout the entire control cycle, which cannot adapt to real-time fluctuations in traffic conditions and affects the effectiveness of coordinated control. Most dynamic division methods rely solely on traffic attributes or network structure, making it difficult to reflect directional congestion. They lack continuous adaptability to the spatiotemporal evolution of congestion, require manual setting of update frequency, and result in irregular sub-zone shapes due to boundary redistribution. The control strategy still needs to be optimized. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a dynamic sub-region division method for traffic control based on congestion evolution. This invention aims to solve the problems of existing methods relying on manually set update frequencies, exhibiting boundary jitter, and lacking adaptability. The method of this invention constructs a similarity metric through three-dimensional coupling of spatial structure, traffic flow characteristics, and functional requirements, and introduces an improved Chameleon algorithm to achieve initial sub-region division. Based on this, a closed-loop system of "prediction-error calculation-parameter correction" is established based on congestion evolution prediction and a dynamic error feedback mechanism to reconstruct and iterate the sub-regions, optimizing the homogeneity within the region, the differences between neighboring regions, and the continuity of boundaries. This achieves smooth evolution and structural optimization of sub-region boundaries, avoids frequent re-division, and thus significantly enhances robustness and suppresses directional congestion interference.

[0006] The present invention solves the above-mentioned technical problems through the following technical means:

[0007] This invention provides a method for dynamically dividing traffic control sub-zones based on congestion evolution, comprising the following steps:

[0008] S1. Obtain road network topology and traffic operation data;

[0009] S1.1 Obtain the topology information of the road network;

[0010] S1.2 Collect traffic operation data for each road segment;

[0011] S1.3 Clean the data collected in steps S1.1 and S1.1, and filter out and remove erroneous data;

[0012] S1.4 Import the cleaned data from step S1.3 into the geographic information system and assign network weights based on traffic characteristics;

[0013] S2. Initial sub-region division;

[0014] S3. Congestion Evolution Prediction;

[0015] S4. Dynamic sub-region partitioning.

[0016] Furthermore, in step S1.1, the topology information of the road network includes node locations, road segment attributes, and connectivity relationships.

[0017] Furthermore, in step S1.2, the traffic operation data for each road segment includes speed, traffic volume, and density every 3 minutes.

[0018] Furthermore, step S2 includes the following sub-steps:

[0019] S2.1 Similarity measurement is based on three types of coupling relationships: spatial structure, traffic characteristics, and functional attributes;

[0020] I. Calculate spatial structure similarity based on shortest path distance and Dice coefficient;

[0021]

[0022] In the formula, |S i |and|S j | represent S respectively i and S j The number of elements, S i and S j Let i and j represent all road segments adjacent to road segments i and j, respectively.

[0023] II. Calculating traffic feature similarity based on DTW distance using flow and speed time series;

[0024] X(i,j)=1-|x i -x j |

[0025] In the formula, x i x j These are the traffic feature weights for road segments i and j, respectively.

[0026] III. After obtaining the weights through multi-attribute decision-making and principal component analysis, calculate the consistency of road segment functional indicators, i.e., functional attribute similarity;

[0027] Y(i,j)=1-|y i -y j |

[0028] In the formula, y i y j These are the functional feature weights for road segments i and j, respectively.

[0029] IV. Using α, β, and γ obtained by the entropy weight method as fusion coefficients, a comprehensive similarity matrix is ​​synthesized;

[0030] ω i,j =α·Dice(i,j)+β·X(i,j)+γ·Y(i,j)

[0031] In the formula, α, β, and γ are the weights of spatial structure, traffic characteristics, and functional attributes, respectively;

[0032] S2.2 The improved Chameleon algorithm is used to perform the initial division of the road network;

[0033] The Chameleon algorithm is improved by merging adjacent road segments during the partitioning process to avoid the occurrence of non-adjacent road segments within the same sub-region. A k-nearest neighbor graph is constructed using the comprehensive similarity matrix as input and clusters are initialized. Cluster pairs are adaptively selected for merging by merging metrics of "relative connectivity" and "relative density" until a predetermined number of initial sub-regions are generated.

[0034] Among them, reverse redistribution is performed on small clusters;

[0035] S2.3 By comparing the modularity of subregion division with the average similarity within subregions under different numbers of subregions, the number of subregions corresponding to the optimal index is taken as the optimal number of subregion divisions.

[0036] Furthermore, step S3 includes the following sub-steps:

[0037] S3.1 Binarizes the congestion status of the road segment at each discrete time and constructs a matrix H;

[0038] Assuming there are L road segments in the study area and the historical observation time is T steps, at each discrete time t∈{1,…,T}, the congestion status of road segment l∈{1,…,L} is binarized to obtain matrix H;

[0039]

[0040] H = [h] lt ]∈{0,1} L×T

[0041] S3.2 uses a sliding window of length Δt to calculate the congestion trend of each road segment and generates a trend matrix T;

[0042] A sliding window of length Δt is introduced to accumulate and compare the congestion changes of each road segment before and after time t over Δt steps, thereby obtaining trend characteristics. t = Δt+1,…,T, and the trend characteristics of all road segments are organized as follows: Among them, h l,t-k This represents the state of l road segments at time tk;

[0043] S3.3 predicts the probability of congestion in the next K steps by jointly using alternating stacked one-dimensional temporal convolution and graph convolution networks based on trend subsequences and geographic topological adjacency matrices.

[0044] Node features are trend subsequences of length τ As input;

[0045] Geographic topological adjacency matrix Used to provide spatial dependency information;

[0046] By using one-dimensional temporal convolutional layers and graph convolutional layers, local short-term trends are captured and spatial adjacency information is aggregated to each node. After multiple layers are stacked alternately, the congestion probability vector of each road segment in the next K steps is output.

[0047] S3.4 utilizes real-time observation error feedback coefficients to dynamically scale input features, adjust adjacency weights, and decay the learning rate to achieve a prediction-error-correction closed loop;

[0048] When a new round of real-time observation Upon arrival, calculate the average absolute error e of the i-th road segment over the past K predictions. i ;

[0049]

[0050] Define the error feedback coefficient γ i Error feedback coefficient γ i Used for dynamic correction of input features and model parameters;

[0051]

[0052] First, the input features are linearly scaled: X' it =X it (1+αγ i ), α∈[0,1];X it Indicates the input features; α represents the scaling factor;

[0053] Then, the adjacency matrix is ​​adjusted according to exponential weights: β>0; A' ij A represents the elements in the adjusted adjacency matrix; ij Represents the elements in the original adjacency matrix; β represents the adjustment parameter; γ i γ represents the error feedback coefficient for road segment i; j This represents the error feedback coefficient for road segment j;

[0054] Finally, the global learning rate is decayed by summing the feedback coefficients of each road segment:

[0055]

[0056] In the formula, λ represents the decay coefficient; η0 represents the initial learning rate; η t This represents the adjusted global learning rate.

[0057] Furthermore, step S4 includes the following sub-steps:

[0058] S4.1 Congestion probability prediction based on the next K steps for each road segment and Calculate the Pearson correlation coefficients between road segments, truncate negative values, and construct a weighted correlation matrix after normalization.

[0059]

[0060]

[0061]

[0062] S4.2 with The Chameleon algorithm is used to adaptively merge high-density clusters as input, dynamically generating spatially coherent and boundary-interpretable sub-region partitioning loops.

[0063] Beneficial effects:

[0064] This invention proposes a congestion evolution-driven road network partitioning framework. Based on the principles of homogeneity, connectivity, and appropriate quantity and area, it accurately characterizes the correlation of road segments through three-dimensional similarity measurement of spatial structure, traffic flow characteristics, and functional requirements, effectively alleviating directional congestion. Furthermore, it predicts the spatiotemporal changes of congestion based on evolutionary patterns and adaptively adjusts the boundaries of sub-regions, avoiding boundary jitter and computational overhead caused by frequent updates, thus significantly improving real-time performance and flexibility. First, based on three similarity measures—spatial structure, traffic characteristics, and functional attributes—an improved Chameleon algorithm combined with an optimal sub-region number determination strategy is used to cluster the entire network of road segments, generating several initial homogeneous sub-regions. Next, the feature sequence and adjacency matrix of each initial sub-region are input into a spatiotemporal prediction framework. After model training, the future congestion probability is output, and a predictive correlation graph is constructed based on this, providing spatiotemporal dependency information for subsequent dynamic partitioning. Finally, based on the predictive correlation graph, the improved Chameleon algorithm is applied again to reconstruct and iterate the sub-regions, optimizing the homogeneity within the region, the differences between neighboring regions, and the continuity of boundaries. The final output is the sub-region partitioning result and the corresponding macroscopic basic map, which can provide decision support for refined traffic management.

[0065] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0066] Figure 1 This is a flowchart of a traffic control sub-zone dynamic division method based on congestion evolution according to the present invention. Detailed Implementation

[0067] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0068] like Figure 1 As shown, this invention provides a method for dynamically dividing traffic control sub-zones based on congestion evolution, including the following steps:

[0069] S1. Obtain road network topology and traffic operation data;

[0070] S1.1 Obtain spatial data of road nodes and edges from the OpenStreetMap platform, including attributes such as node coordinates, road segment name, road level and number of lanes, and unify the projected coordinate system;

[0071] S1.2 collects traffic flow, speed and density data for each road segment every 3 minutes using a variety of sensors such as induction coils, video detection and floating car trajectories;

[0072] S1.3 Clean the data collected in steps S1.1 and S1.1, and filter out and remove erroneous data;

[0073] Filter outliers based on reasonable upper and lower limits, remove records with missing rates higher than the set threshold, and fill isolated missing data points with interpolation or mean values ​​from adjacent time periods to ensure the continuity and reliability of spatiotemporal data.

[0074] S1.4 maps the processed traffic flow data to the corresponding road segments, sets the weights to a weighting function of traffic density, and performs visual verification to ensure that the weight distribution of main roads and branch roads reflects the real traffic pattern.

[0075] S2. Initial sub-region division;

[0076] S2.1 Similarity measurement is based on three types of coupling relationships: spatial structure, traffic characteristics, and functional attributes;

[0077] I. Calculate spatial structure similarity based on shortest path distance and Dice coefficient;

[0078]

[0079] In the formula, |S i |and|S j | represent S respectively i and Sj The number of elements, S i and S j Let i and j represent all road segments adjacent to road segments i and j, respectively.

[0080] II. Calculating traffic feature similarity based on DTW distance using flow and speed time series;

[0081] X(i,j)=1-|x i -x j |

[0082] In the formula, x i x j These are the traffic feature weights for road segments i and j, respectively.

[0083] III. After obtaining the weights through multi-attribute decision-making and principal component analysis, calculate the consistency of road segment functional indicators, i.e., functional attribute similarity;

[0084] Y(i,j)=1-|y i -y j |

[0085] In the formula, y i y j These are the functional feature weights for road segments i and j, respectively.

[0086] IV. Using α, β, and γ obtained by the entropy weight method as fusion coefficients, a comprehensive similarity matrix is ​​synthesized;

[0087] ω i,j =α·Dice(i,j)+β·X(i,j)+γ·Y(i,j)

[0088] In the formula, α, β, and γ are the weights of spatial structure, traffic characteristics, and functional attributes, respectively;

[0089] S2.2 The improved Chameleon algorithm is used to perform the initial division of the road network;

[0090] The Chameleon algorithm is improved by merging adjacent road segments during the partitioning process to avoid the occurrence of non-adjacent road segments within the same sub-region. A k-nearest neighbor graph is constructed using the comprehensive similarity matrix as input and clusters are initialized. Cluster pairs are adaptively selected for merging by merging metrics of "relative connectivity" and "relative density" until a predetermined number of initial sub-regions are generated.

[0091] Among them, excessively small clusters are reverse-reassigned to ensure that each sub-region is of moderate size and has good internal connectivity.

[0092] S2.3 By comparing the modularity of subregion division and the average similarity within subregions under different numbers of subregions, the optimal number of subregions and the division scheme are determined; the number of subregions corresponding to the optimal index is taken as the optimal number of subregions.

[0093] In this embodiment, the normalized total variance (NTV), the average Ncut Silhouette value, and the modularity (Q) are calculated under various partitioning conditions from k=2 to 9. Taking into account the inflection point and stability of the three indicators, k=6 is finally determined as the optimal scheme for the initial sub-region partitioning.

[0094] S3. Congestion Evolution Prediction;

[0095] S3.1 Binarizes the congestion status of the road segment at each discrete time and constructs a matrix H;

[0096] Assuming there are L road segments in the study area and the historical observation time is T steps, at each discrete time t∈{1,…,T}, the congestion status of road segment l∈{1,…,L} is binarized to obtain matrix H;

[0097]

[0098] H = [h] lt ]∈{0,1} L×T

[0099] S3.2 uses a sliding window of length Δt to calculate the congestion trend of each road segment and generates a trend matrix T;

[0100] A sliding window of length Δt is introduced to accumulate and compare the congestion changes of each road segment before and after time t over Δt steps, thereby obtaining trend characteristics. t = Δt+1,…,T, and the trend characteristics of all road segments are organized as follows: Among them, h l,t-k This represents the state of l road segments at time tk;

[0101] S3.3 Based on the above, a temporal graph convolutional network is used to jointly model the spatiotemporal correlation and temporal correlation of congestion propagation;

[0102] Based on trend subsequences and geographic topological adjacency matrices, the probability of congestion in the next K steps is jointly predicted by alternating stacked one-dimensional temporal convolutional and graph convolutional networks.

[0103] Node features are trend subsequences of length τ As input;

[0104] The graph structure is an adjacency matrix generated from geographic topology. Provides spatial dependency information;

[0105] By using one-dimensional temporal convolutional layers and graph convolutional layers, local short-term trends are captured and spatial adjacency information is aggregated to each node. After multiple layers are stacked alternately, the congestion probability vector of each road segment in the next K steps is output.

[0106] S3.4 utilizes real-time observation error feedback coefficients to dynamically scale input features, adjust adjacency weights, and decay the learning rate to achieve a prediction-error-correction closed loop;

[0107] When a new round of real-time observation Upon arrival, first calculate the average absolute error e of the i-th road segment over the past K predictions. i ;

[0108]

[0109] Define the error feedback coefficient γ i Error feedback coefficient γ i Used for dynamic correction of input features and model parameters;

[0110]

[0111] First, the input features are linearly scaled: X' it =X it (1+αγ i ), α∈[0,1];X it Indicates the input features; α represents the scaling factor;

[0112] Then, the adjacency matrix is ​​adjusted according to exponential weights: A' ij A represents the elements in the adjusted adjacency matrix; ij Represents the elements in the original adjacency matrix; β represents the adjustment parameter; γ i γ represents the error feedback coefficient for road segment i; j This represents the error feedback coefficient for road segment j;

[0113] Finally, the global learning rate is decayed by summing the feedback coefficients of each road segment:

[0114]

[0115] In the formula, λ represents the decay coefficient; η0 represents the initial learning rate; η t This represents the adjusted global learning rate.

[0116] S4. Dynamic sub-region partitioning mainly includes the following two parts:

[0117] S4.1 Congestion probability prediction based on the next K steps for each road segment and Calculate the Pearson correlation coefficients between road segments, truncate negative values, and construct a weighted correlation matrix after normalization.

[0118]

[0119] S4.2 Then on this association diagram, with The Chameleon algorithm is invoked as input to reconstruct sub-regions. The algorithm comprehensively considers the "relative connectivity" and "relative density" between clusters during each merging, and can adaptively identify high-density clusters. It also ensures that the resulting sub-regions are spatially coherent and have good interpretability of boundaries, thereby realizing a closed loop of dynamic sub-region partitioning based on congestion evolution.

[0120] It is hereby declared that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for dynamically dividing traffic control sub-regions based on congestion evolution, characterized in that, Includes the following steps: S1. Obtain road network topology and traffic operation data; S1.1 Obtain the topology information of the road network; S1.2 Collect traffic operation data for each road segment; S1.3 Clean the data collected in steps S1.1 and S1.1, and filter out and remove erroneous data; S1.4 Import the cleaned data from step S1.3 into the geographic information system and assign network weights based on traffic characteristics; S2. Initial sub-region division; S3. Congestion Evolution Prediction; S4. Dynamic sub-region partitioning.

2. The method for dynamically dividing traffic control sub-regions based on congestion evolution according to claim 1, characterized in that: In step S1.1, the topology information of the road network includes node locations, road segment attributes, and connectivity relationships.

3. The method for dynamically dividing traffic control sub-regions based on congestion evolution according to claim 2, characterized in that: In step S1.2, the traffic operation data for each road segment includes speed, traffic volume, and density every 3 minutes.

4. The method for dynamically dividing traffic control sub-regions based on congestion evolution according to claim 3, characterized in that, Step S2 includes the following sub-steps: S2.1 Similarity measurement is based on three types of coupling relationships: spatial structure, traffic characteristics, and functional attributes; I. Calculate spatial structure similarity based on shortest path distance and Dice coefficient; In the formula, |S i |and|S j | represent S respectively i and S j The number of elements, S i and S j Let i and j represent all road segments adjacent to road segments i and j, respectively. II. Calculating traffic feature similarity based on DTW distance using flow and speed time series; X(i,j)=1-|x i -x j | In the formula, x i x j These are the traffic feature weights for road segments i and j, respectively. III. After obtaining the weights through multi-attribute decision-making and principal component analysis, calculate the consistency of road segment functional indicators, i.e., functional attribute similarity; Y(i,j)=1-|y i -y j | In the formula, y i y j These are the functional feature weights for road segments i and j, respectively. IV. Using α, β, and γ obtained by the entropy weight method as fusion coefficients, a comprehensive similarity matrix is ​​synthesized; ω i,j =α·Dice(i,j)+β·X(i,j)+γ·Y(i,j) In the formula, α, β, and γ are the weights of spatial structure, traffic characteristics, and functional attributes, respectively; S2.2 The improved Chameleon algorithm is used to perform the initial division of the road network; The Chameleon algorithm is improved by merging adjacent road segments during the partitioning process to avoid the occurrence of non-adjacent road segments within the same sub-region. A k-nearest neighbor graph is constructed using the comprehensive similarity matrix as input and clusters are initialized. Cluster pairs are adaptively selected for merging by merging metrics of "relative connectivity" and "relative density" until a predetermined number of initial sub-regions are generated. Among them, reverse redistribution is performed on small clusters; S2.3 By comparing the modularity of subregion division with the average similarity within subregions under different numbers of subregions, the number of subregions corresponding to the optimal index is taken as the optimal number of subregion divisions.

5. The method for dynamically dividing traffic control sub-regions based on congestion evolution according to claim 4, characterized in that, Step S3 includes the following sub-steps: S3.1 Binarizes the congestion status of the road segment at each discrete time and constructs a matrix H; Assuming there are L road segments in the study area and the historical observation time is T steps, at each discrete time t∈{1,…,T}, the congestion status of road segment l∈{1,…,L} is binarized to obtain matrix H; H=[h lt ]∈{0,1} L×T S3.2 uses a sliding window of length Δt to calculate the congestion trend of each road segment and generates a trend matrix T; A sliding window of length Δt is introduced to accumulate and compare the congestion changes of each road segment before and after time t over Δt steps, thereby obtaining trend characteristics. And organize the trend characteristics of all road segments into Among them, h l,t-k This represents the state of l road segments at time tk; S3.3 predicts the probability of congestion in the next K steps by jointly using alternating stacked one-dimensional temporal convolution and graph convolution networks based on trend subsequences and geographic topological adjacency matrices. Node features are trend subsequences of length τ As input; Geographic topological adjacency matrix Used to provide spatial dependency information; By using one-dimensional temporal convolutional layers and graph convolutional layers, local short-term trends are captured and spatial adjacency information is aggregated to each node. After multiple layers are stacked alternately, the congestion probability vector of each road segment in the next K steps is output. S3.4 utilizes real-time observation error feedback coefficients to dynamically scale input features, adjust adjacency weights, and decay the learning rate to achieve a prediction-error-correction closed loop; When a new round of real-time observation Upon arrival, calculate the average absolute error e of the i-th road segment over the past K predictions. i ; Define the error feedback coefficient γ i Error feedback coefficient γ i Used for dynamic correction of input features and model parameters; First, the input features are linearly scaled: X' it =X it (1+αγ i ), α∈[0,1];X it Indicates the input features; α represents the scaling factor; Then, the adjacency matrix is ​​adjusted according to exponential weights: A' ij A represents the elements in the adjusted adjacency matrix; ij Represents the elements in the original adjacency matrix; β represents the adjustment parameter; γ i γ represents the error feedback coefficient for road segment i; j This represents the error feedback coefficient for road segment j; Finally, the global learning rate is decayed by summing the feedback coefficients of each road segment: In the formula, λ represents the decay coefficient; η0 represents the initial learning rate; η t This represents the adjusted global learning rate.

6. The method for dynamically dividing traffic control sub-regions based on congestion evolution according to claim 5, characterized in that, Step S4 includes the following sub-steps: S4.1 Congestion probability prediction based on the next K steps for each road segment and Calculate the Pearson correlation coefficients between road segments, truncate negative values, and construct a weighted correlation matrix after normalization. S4.2 with The Chameleon algorithm is used to adaptively merge high-density clusters as input, dynamically generating spatially coherent and boundary-interpretable sub-region partitioning loops.

Images