Traffic artery coordination control real-time optimization method

Abstract

The invention relates to an arterial traffic coordination control real-time optimization method. The method comprises the following steps: S1, collecting traffic data and converting the traffic data into input data suitable for a graph neural network; s2, constructing a graph data structure, setting intersections as nodes, setting road sections as edges, and adjusting the weight of the weighted adjacent matrix at a fixed frequency according to the traffic data of the road and the historical traffic flow; s3, taking the graph data structure in the S2 as input, constructing a graph neural network model, and capturing spatial-temporal characteristics among nodes; and S4, formulating or adjusting an objective function according to the spatial-temporal characteristics, the actual traffic demands and the constraint conditions, solving the objective function by a genetic algorithm, carrying out continuous iteration on crossover and variation, and evaluating individuals according to the objective function and the constraint conditions in each iteration to obtain an optimal signal timing scheme. According to the method, spatial-temporal characteristics are effectively captured by adopting a graph neural network technology, signal timing optimization is carried out in combination with a genetic algorithm and reinforcement learning, and a signal timing scheme can be dynamically adjusted according to real-time traffic demands.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation, and particularly relates to a real-time optimization method for coordinated control of arterial roads. Background Art

[0002] With the acceleration of the urbanization process and the continuous growth of the motor vehicle ownership, the problem of traffic congestion has become increasingly serious, bringing many inconveniences to people's travel and the development of the city. As an important backbone of urban traffic, the operation efficiency of arterial roads directly affects the traffic conditions of the whole city. Most traditional traffic signal control methods are based on fixed timing plans or simple inductive control strategies, and it is difficult to adapt to the dynamic changes of traffic flow.

[0003] In recent years, intelligent transportation systems (ITS) have been widely studied and applied, aiming to improve traffic conditions through advanced information technology, communication technology and control technology. Among them, coordinated control of arterial roads, as an important part of ITS, is committed to realizing the smooth passage of traffic flow, reducing the number of vehicle stops and delay time by optimizing the signal timing of multiple adjacent intersections.

[0004] At the same time, significant progress has been made in artificial intelligence technology, especially the graph neural network (GNN) in the field of deep learning. The graph neural network can effectively process data with complex topological structures, such as the relationships between intersections and road segments in a traffic network. Applying the graph neural network to the coordinated control of arterial roads provides new ideas and methods for real-time optimization of traffic signals. By modeling and analyzing the nodes (intersections) and edges (road segments) in the traffic network, the graph neural network can learn the spatio-temporal characteristics of traffic flow, so as to more accurately predict traffic demand and adjust signal timing in real time to adapt to the changing traffic conditions.

[0005] Existing traffic arterial signal coordinated control methods have various defects:

[0006] 1. Limitations of fixed timing plans: Traditional arterial control methods rely on pre-set schedules, which are preset according to historical traffic data. However, traffic flow has significant dynamic change characteristics, and the traffic demands vary greatly at different times and under different weather conditions. Fixed timing plans cannot perceive the changes in traffic flow in real time, resulting in serious congestion at intersections during peak traffic periods, while there is a waste of signal resources during off-peak periods.

[0007] 2. Deficiencies of simple induction control: Although some induction control strategies can adjust signals based on vehicle detection information at the current intersection, this method often only considers the local situation of a single intersection and lacks global coordination of the entire traffic artery. Signal timing conflicts may occur between adjacent intersections, leading to the "oscillation" phenomenon of traffic flow in the artery and unable to achieve efficient coordinated traffic.

[0008] 3. Difficulty in dealing with complex traffic network structures: The urban traffic network is a complex topological structure, containing multiple intersections and road segments, with complex mutual influence relationships among them. Traditional control methods are difficult to effectively model and analyze such complex network structures and cannot fully utilize global information in the traffic network to optimize signal control.

[0009] 4. Lack of real-time performance and adaptability: Existing traffic control methods often cannot respond quickly when facing sudden traffic events (such as traffic accidents, large-scale events, etc.), and cannot adjust signal timing in real time to relieve traffic congestion. Their decision-making process is relatively slow and difficult to adapt to rapidly changing traffic conditions. Summary of the Invention

[0010] An embodiment of the present invention provides a real-time optimization method for coordinated control of traffic arteries, which comprehensively and accurately reflects traffic conditions through multi-source data collection and fusion.

[0011] To achieve the above object, the embodiment of the present invention provides the following technical solutions:

[0012] A real-time optimization method for coordinated control of traffic arteries, comprising the following steps:

[0013] S1. Collect traffic data and perform data preprocessing to convert traffic data into data suitable for input to a graph neural network;

[0014] S2. Construct a graph data structure, set intersections as nodes and road segments as edges, and then adjust the weights of the weighted adjacency matrix at a fixed frequency according to the traffic data and historical traffic volume of the road;

[0015] S3. Take the graph data structure in S2 as the input, select a graph convolutional network as the basic architecture, and construct a graph neural network model for the traffic artery to capture the spatio-temporal features between each node;

[0016] S4. Formulate and adjust the objective function according to the spatio-temporal features in combination with actual traffic demands and constraints, solve the objective function using a genetic algorithm, continuously iterate and update the population through crossover and mutation, evaluate individuals according to the objective function and constraints in each iteration, and finally obtain the optimal signal timing plan.

[0017] Further, the weight formula of the weighted adjacency matrix in S2 is

[0018]

[0019] Where ω ij is the weight between nodes i and j;

[0020] C ij is the road section capacity;

[0021] Q ij is the historical traffic volume of the road section.

[0022] Furthermore, the training method of the graph neural network in S3 includes dividing the traffic data contained in the graph data structure obtained in S2 into a training set, a validation set, and a test set according to a reasonable ratio;

[0023] Select a graph convolutional network as the basic architecture, and extract the features of nodes and their neighbor nodes through convolutional operations to obtain an initial graph neural model;

[0024] Input the training set data into the initial graph neural model, calculate the prediction result of the model through forward propagation, calculate the loss function according to the prediction result and the traffic flow data in the traffic data, calculate the gradient of the loss function with respect to the parameters of the graph neural network model using the backpropagation algorithm, and update the model parameters using the Adam optimizer to minimize the loss function;

[0025] Finally, use the test set data to evaluate the trained model to obtain the graph neural network model.

[0026] Furthermore, the optimization algorithm in S4 adopts a genetic algorithm, which includes mapping the traffic signal phase duration and traffic signal phase sequence to the chromosomes of the genetic algorithm, and comprehensively considering multiple indicators such as vehicle delay time, number of stops, queue length, and high-level feature influence factor to construct a comprehensive fitness function:

[0027]

[0028] Where T total_delay is the total delay time;

[0029] S total is the total number of stops;

[0030] Q total is the total queue length;

[0031] F advanced is the high-level feature influence factor;

[0032] ω5, ω6, ω7, ω8 are weight coefficients.

[0033] Furthermore, in the mutation and crossover operations of the genetic algorithm, a fine-tuning mechanism for the phase duration is introduced. By setting the range of phase duration changes and based on the maximum and minimum green light time constraints, the phase duration is randomly adjusted when generating new solutions to ensure that the new solutions not only meet the constraint conditions but also may bring performance improvements.

[0034] Furthermore, in each iteration, the individual with the highest fitness is selected as the elite individual. The elite individual does not participate in the crossover and mutation operations and is retained for the next generation to ensure that the algorithm does not lose the current optimal solution and at the same time speeds up the convergence rate.

[0035] Furthermore, the objective function of the S4 optimal signal timing scheme is optimized for the green wave band. The green wave band optimization adjusts the objective function through the following formula:

[0036] Z new =Z old -ΔT total_delay -ΔS total -ΔQ total

[0037] Where:

[0038] Z new is the objective function value after the green wave band optimization;

[0039] Z old is the objective function value before the green wave band optimization;

[0040] ΔT total_delay is the total reduced delay time obtained from the green wave band optimization;

[0041] ΔS total is the total reduced number of stops obtained from the green wave band optimization;

[0042] ΔQ total is the total reduced queue length obtained from the green wave band optimization.

[0043] Furthermore, the dynamic adjustment phase difference calculation method for the green wave band optimization includes:

[0044] First, calculate the average vehicle speed:

[0045]

[0046] where, v i is the speed of the i-th vehicle,

[0047] N is the number of vehicles within the statistical time period;

[0048] Then, calculate the time required for the vehicle to travel from the upstream intersection to the downstream intersection:

[0049]

[0050] Among them, L is the distance between adjacent intersections;

[0051] Then, considering traffic flow fluctuations and signal cycle constraints, the phase difference is dynamically adjusted. When it is detected that the queue at the downstream intersection is too long, the phase difference can be shortened to release vehicles in advance:

[0052]

[0053] Among them, Q queue is the queue length at the downstream intersection;

[0054] C phase is the lane passing capacity;

[0055] β is a correction coefficient, usually taking 0.1 - 0.3;

[0056] Finally, by dynamically adjusting the phase difference, the waiting time of vehicles at intersections can be reduced, thereby reducing the total delay time T total_delay :

[0057]

[0058] Among them, is the waiting time before adjustment at the i-th intersection;

[0059] is the waiting time after adjustment at the i-th intersection;

[0060] M is the number of intersections.

[0061] Furthermore, the coordinated optimization of the green wave band forms a coherent green wave band by coordinating the green light times of multiple intersections:

[0062]

[0063] Among them, M is the number of intersections on the main line;

[0064] T green,k is the green light time at the k-th intersection;

[0065] Q pass,k is the number of vehicles passing through the k-th intersection during the green light period.

[0066] Furthermore, a reward function is designed to optimize the update and iteration of the genetic algorithm. The positive rewards in the reward function include reducing vehicle delays, reducing queue lengths, and reducing the number of stops, while the negative rewards include increasing congestion, increasing the number of stops, and extending queue lengths;

[0067] R = r1·(-T delay ) + r2·(-S stop ) + r3·(-Qqueue )

[0068] Among them, r1, r2, and r3 are reward weights;

[0069] The reward function adopts Q-learning, iteratively updates the strategy, and realizes the dynamic optimization of signal timing;

[0070] Q(s,a)←Q(s,a)+α(r+γmax a' Q(s,a′)-Q(s,a))

[0071] Among them, α is the learning rate;

[0072] γ is the discount factor;

[0073] r is the reward;

[0074] s is the current state;

[0075] a is the current action;

[0076] s′ is the next state;

[0077] a′ is the next action.

[0078] The embodiments of the present invention have the following advantages:

[0079] A real-time optimization method for coordinated control of arterial roads of the present invention comprehensively and accurately reflects the traffic conditions through multi-source data collection and fusion; it adopts advanced data processing and graph neural network technologies to effectively capture the spatio-temporal characteristics of traffic data and improve the accuracy of model prediction; and combines genetic algorithms and reinforcement learning for signal timing optimization, which can dynamically adjust the signal timing scheme according to real-time traffic demands, significantly reduce vehicle delays, stop times, and queue lengths, improve the traffic efficiency of arterial roads, and provide an effective technical means for alleviating traffic congestion.

[0080] A real-time optimization method for coordinated control of arterial roads of the present invention has an emergency handling mechanism for special situations, which further enhances the stability and adaptability of the system and ensures good operation effects in complex and changeable traffic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, other implementation drawings can be obtained by extending according to the provided drawings without creative work.

[0082] The structures, proportions, sizes, etc. illustrated in this specification are only used to cooperate with the content disclosed in the specification for those familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention. Therefore, they do not have substantial technical significance. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the efficacy that the present invention can produce and the purpose that can be achieved, should still fall within the scope that can be covered by the technical content disclosed in the present invention.

[0083] Figure 1 It is a method flowchart of a real-time optimization method for arterial traffic signal coordination control provided by an embodiment of the present invention;

[0084] Figure 2 It is a method flowchart of the training method of the graph neural network model in the real-time optimization method for arterial traffic signal coordination control provided by an embodiment of the present invention;

[0085] Figure 3 It is a method flowchart of the genetic algorithm in the real-time optimization method for arterial traffic signal coordination control provided by an embodiment of the present invention;

[0086] Figure 4 It is a method flowchart of the green wave band optimization in the real-time optimization method for arterial traffic signal coordination control provided by an embodiment of the present invention. Specific Embodiments

[0087] The following specific embodiments illustrate the implementation manners of the present invention. Those familiar with this technology can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope protected by the present invention.

[0088] As Figure 1 shown, a real-time optimization method for arterial traffic signal coordination control includes the following steps:

[0089] S1. Collect traffic data, perform data preprocessing, and convert the traffic data into data suitable for input to the graph neural network.

[0090] S1-1. Data collection

[0091] The traffic data of this technology includes real-time traffic data and traffic inherent data. Among them, the real-time traffic data is the real-time traffic data of each intersection and road section in the arterial traffic, and it is collected at a fixed frequency. In this embodiment, it is preferably collected at a frequency of once per minute. The traffic inherent data is the inherent data of the intersection and road section itself, which is only entered initially. If there are changes in the intersection or road section later, the data is updated.

[0092] The real-time traffic data at intersections includes: real-time traffic volume at intersections, average vehicle speed at intersections, and signal phase status.

[0093] The inherent traffic data at intersections includes: inherent attributes such as intersection type (crossroads, T-junctions), intersection width, and intersection geographical location.

[0094] The real-time traffic data on road segments: includes traffic volume on road segments, queue length, and average vehicle speed.

[0095] The inherent traffic data on road segments includes: road segment length, number of lanes, road segment type (one-way road, two-way road, multi-lane road), road segment capacity, and road segment geographical location.

[0096] Among them, the geographical locations of intersections and road segments are obtained through electronic maps or manually input by humans.

[0097] The real-time traffic data is the data collected subsequently, which is different from the traffic historical data. Both are obtained from the traffic flow monitoring system. By docking with the authoritative data platform, corresponding weather data and holiday information can also be obtained, which can ensure the reliability, authenticity, and timeliness of the data. The real-time traffic data is pushed to the central system in real time through the network.

[0098] S1-2. Data preprocessing

[0099] The data preprocessing of this technology is used for data cleaning, removing noise and outliers; it is also used for data normalization to convert different types of data into a format suitable for input to the graph neural network. In this technology, the Kalman filtering technology is first used to construct the state equation and observation equation of traffic data to remove the noise data collected by sensors in the traffic data; then the 3σ principle is used to identify and remove outliers, and the time series prediction method based on the long short-term memory network (LSTM) is selected to fill in the missing data; finally, the min-max normalization method is used to convert the denoised traffic data into a format suitable for input to the graph neural network. The formula of the min-max normalization method is as follows:

[0100]

[0101] Among them, x is the original data;

[0102] x min is the minimum value in the dataset;

[0103] x max is the maximum value in the dataset.

[0104] S2. Construct a graph data structure. Set intersections as nodes, where the node attributes include traffic data of the intersections. Set road segments as edges, where the edge attributes include traffic data of the road segments. Then, adjust the weights of the weighted adjacency matrix at a fixed frequency according to the traffic data of the road segments and the historical traffic data of the road segments.

[0105] Among them, each intersection is set as a node in the graph, and the real-time traffic data and the inherent basic data of the intersection are both set as node attributes, which include information such as the real-time traffic volume of the intersection, the average vehicle speed in each direction of the intersection, the signal phase state of the intersection, the geographical location of the intersection, and the type of the intersection.

[0106] Each road segment is set as an edge in the graph, and the real-time traffic data and the inherent basic data of the road segment are both set as edge attributes, which include information such as the traffic volume of the road segment, the queue length, the average vehicle speed, the length of the road segment, the number of lanes, the type of the road segment, and the geographical location of the road segment.

[0107] In the time dimension, introduce timestamps for the real-time traffic data of intersections and road segments to construct a time series graph, and the timestamps can be accurate to minutes.

[0108] In the space dimension, in this technology, the weights of the weighted adjacency matrix are adjusted at a fixed frequency such as every minute or every hour by combining the traffic capacity of the road segment and the historical traffic volume of the road segment. In this embodiment, it is preferably adjusted every minute to be consistent with the acquisition frequency in the time dimension. The formula is as follows:

[0109]

[0110] Among them, ω ij is the weight between nodes i and j;

[0111] C ij is the traffic capacity of the road segment;

[0112] Q ij is the historical traffic volume of the road segment.

[0113] S3. Take the graph data structure in S2 as the input, select the graph convolutional network as the basic architecture, and construct a traffic trunk line graph neural network model to capture the spatio-temporal features between each node.

[0114] As Figure 2 shown, the specific training method of the graph neural network model in this technology is as follows:

[0115] S3-1. Data preparation:

[0116] Divide the historical data containing traffic flow information and corresponding signal timing plans under different time periods and different weather conditions in the graph data structure obtained in S2 into a training set, a validation set, and a test set according to a ratio of 6:2:2.

[0117] S3-2, Model Initialization:

[0118] Select a suitable graph neural network architecture: In this technology, the Graph Convolutional Network (GCN) is selected as the basic architecture. This architecture can effectively process graph-structured data and capture the complex relationships between nodes. The Graph Convolutional Network (GCN) can extract the features of nodes and their neighbor nodes through convolutional operations to obtain the initial graph neural model.

[0119]

[0120] Among them, is the adjacency matrix;

[0121] is the degree matrix;

[0122] H(l) is the node feature of the l-th layer;

[0123] W(l) is the weight matrix of the l-th layer;

[0124] σ is the activation function.

[0125] Initialize model parameters: Randomly initialize the weight parameters of the model and set hyperparameters such as the learning rate, number of layers, and number of hidden units. For example, the learning rate is set to 0.001, the number of hidden layers is 3, and the number of hidden units in each layer is 64.

[0126] S3-3, Training Process:

[0127] a. Forward propagation: Input the training set data into the initial graph neural model and calculate the prediction results of the model through forward propagation.

[0128] b. Calculate the loss function: Calculate the loss function based on the prediction results and the traffic flow data in the real-time traffic data. The Weighted Mean Squared Error (WMSE) is used to consider the traffic flow prediction errors at different times and different intersections, and a regularization term is introduced to prevent overfitting.

[0129]

[0130] Among them, ω i is the weight coefficient;

[0131] yi is the actual value;

[0132] is the predicted value;

[0133] λ is the regularization parameter;

[0134] θ is the model parameter.

[0135] c. Backpropagation and parameter update: The backpropagation algorithm is used to calculate the gradients of the loss function with respect to the parameters of the graph neural network model, and the Adam optimizer is used to update the model parameters to minimize the loss function.

[0136]

[0137] Among them, η is the learning rate.

[0138] Validation set evaluation: During the training process, the model is regularly evaluated using the validation set data, and hyperparameters are adjusted to avoid overfitting.

[0139] S3-4. Model evaluation:

[0140] The trained model is evaluated using the test set data, and the accuracy and performance metrics of the model in predicting traffic flow changes are calculated, such as accuracy, recall rate, root mean square error (RMSE), etc. According to the evaluation results, the model is further optimized and adjusted until the RMSE is less than a certain threshold (such as 10%), and the accuracy is greater than 90%, and finally a graph neural network model is obtained.

[0141] This technology takes the graph data structure obtained in S2 as the input, constructs a graph neural network model suitable for traffic trunk network modeling according to the graph data structure, and is used to learn the mutual relationship between nodes (intersections) in the traffic network and the spatio-temporal characteristics of traffic flow.

[0142] The spatio-temporal characteristics refer to the change rules of traffic data in time and space, that is, the spatio-temporal characteristics include time characteristics and space characteristics. Among them, the time characteristics are the periodic changes of traffic flow (such as morning and evening rush hours); the space characteristics are the mutual influence characteristics of traffic flow between adjacent intersections.

[0143] The extraction method of the spatio-temporal characteristics is as follows: The spatio-temporal characteristics in traffic data can be captured by multi-layer graph convolution or graph attention. Among them, the graph convolutional network (GCN) can extract the characteristics of nodes and their neighbor nodes through convolution operations, and the graph attention network (GAT) can dynamically adjust the weights between nodes through the attention mechanism.

[0144] The spatio-temporal features also include high-level features. Through multi-layer graph convolution or graph attention operations, the model can extract high-level features from the original data. High-level features refer to the abstract features extracted from the original traffic data by the graph neural network. These features can reflect the complex patterns and potential laws of the traffic network and can guide the formulation and adjustment of the objective function. Specifically, the time-dimensional features in the spatio-temporal features, such as the variation laws of traffic flow in different time periods like peak hours and off-peak hours, can guide the optimization algorithm to adopt different signal timing strategies in different time periods. During peak hours, focus on reducing the total delay time of vehicles, and the green light duration of the main roads can be appropriately extended; while during off-peak hours, the optimization of the total number of stops and the total queue length can be taken into account. The space-dimensional features, such as the distribution of congestion conditions on each road section, can enable the optimization algorithm to prioritize the adjustment of the signal timing at the intersections in the severely congested areas to relieve traffic pressure.

[0145] The complex patterns and potential laws of the traffic network reflected by the high-level features have a more profound guiding significance for the optimization algorithm. For example, the causal relationship of traffic flow between road sections reflected in the high-level features can help the optimization algorithm predict the impact on the traffic conditions of the traffic artery when adjusting the signal timing at a certain intersection, so as to make a more global and reasonable timing decision. Based on these spatio-temporal features, optimize and adjust the overall objective function in different situations to reduce the total vehicle delay time, the total number of stops, and the total queue length, and comprehensively improve the operation efficiency of the traffic artery. During peak hours, according to the time-dimensional features, increase the weight coefficient of reducing the total delay time in the objective function, so that the optimization algorithm focuses more on reducing delays; in severely congested road sections or intersections, according to the space-dimensional features, increase the weight coefficient of reducing the total queue length to relieve congestion first.

[0146] S4. Develop and adjust the objective function according to the spatio-temporal features in combination with the actual traffic demand and constraint conditions, solve the objective function using the genetic algorithm, continuously iterate and update the population through crossover and mutation, evaluate the individuals according to the objective function and constraint conditions in each iteration, and finally obtain the optimal signal timing plan.

[0147] According to the spatio-temporal features of the graph neural network model in S3, combine the actual traffic demand (reducing the total vehicle delay duration) and constraint conditions of the traffic artery (the constraint conditions are determined by traffic laws, traffic-related rules, and local regulations, such as the maximum and minimum green light times and phase sequences at intersections), and use the optimization algorithm to optimize the traffic signal timing plan to generate the optimal traffic signal timing plan.

[0148] S4-1. Set the objective function

[0149] All the components of the objective function are obtained based on the spatio-temporal features obtained from S3, combined with the actual traffic demands and constraints of the transportation arteries. The components of the objective function in this technology include reducing the total vehicle delay time, reducing the total number of stops, and reducing the total queue length. The objective function is formulated as follows:

[0150] minimize Z=ω1·T total_delay +ω2·S total +ω3·Q total +ω4·F advanced

[0151] where T total_delay is the total delay time;

[0152] S total is the total number of stops;

[0153] Q total is the total queue length;

[0154] F advanced is the spatio-temporal feature influence factor;

[0155] ω1, ω2, ω3, and ω4 are weight coefficients, which are adjusted according to the actual traffic demands.

[0156] S4-2. Solve the objective function:

[0157] In this technology, the genetic algorithm is used to solve the above objective function. Taking the initial signal timing plan as an individual of the population, through operations such as selection, crossover, and mutation, the population is continuously iteratively updated to find the optimal signal timing plan. In each iteration, the individuals are evaluated according to the objective function and constraints, and the individuals with higher fitness are selected to enter the next generation.

[0158] The weight coefficients ω1, ω2, ω3, ω4 in the objective function directly affect the evaluation process. If ω1 increases, it indicates that more attention is paid to reducing the total delay time in the evaluation of the genetic algorithm, and the algorithm will be more inclined to select individuals that can reduce the total delay time to enter the next generation. If ω2 increases, it means that more attention is paid to reducing the total number of stops in the genetic algorithm evaluation, and the algorithm will be more likely to choose individuals that can reduce the total number of stops to enter the next generation. If ω3 increases, it shows that more attention is paid to reducing the total queue length in the genetic algorithm evaluation, and the algorithm will tend to select individuals that can reduce the total queue length to enter the next generation. If ω4 increases, it means that more attention is paid to the influence factor of high-level features in the genetic algorithm evaluation, and the algorithm will be more inclined to select individuals that can enhance the influence factor of high-level features to enter the next generation. Then, the optimized signal timing plan is sent to the signal controllers at each intersection to achieve real-time adjustment and control of traffic signals. At the same time, the execution effect of signal control can also be monitored in real time, and the feedback information is returned to the data acquisition module to form a closed-loop control and continuously optimize the signal control strategy.

[0159] S4-2. Genetic Algorithm (GA)

[0160] As Figure 3 shown, the specific method of the genetic algorithm of this technology is as follows:

[0161] S4-2-1. Coding Scheme: Real number coding is adopted, and the traffic signal phase duration and traffic signal phase sequence in the signal timing plan are mapped to the chromosomes of the genetic algorithm.

[0162] S4-2-2. Constructing the Comprehensive Fitness Function: Considering multiple indicators such as vehicle delay time, number of stops, queue length, and influence factor of high-level features, a comprehensive fitness function is constructed.

[0163]

[0164] where T total_delay is the total delay time;

[0165] S total is the total number of stops;

[0166] Q total is the total queue length;

[0167] F advanced is the influence factor of high-level features;

[0168] ω5, ω6, ω7, ω8 are weight coefficients.

[0169] Among them, although the weight coefficients ω5, ω6, ω7, and ω8 have different forms from the weight coefficients ω1, ω2, ω3, and ω4 in the objective function, they have the same purpose, which is to balance the importance of different traffic indicators in the optimization process. There is no direct numerical correspondence between them, but they are interrelated under the overall goal of optimizing traffic signal timing. These weight coefficients can be adjusted according to spatio-temporal characteristics and actual traffic demands. For example, in areas with large traffic flow and severe congestion, ω5 can be appropriately increased to make the algorithm pay more attention to reducing the total delay time; in areas with relatively small traffic flow and more sensitive to the number of stops, ω6 can be increased; or by analyzing historical traffic data, statistically analyzing the influence degree of each indicator on traffic efficiency under different traffic conditions, and determining the weight coefficients based on this.

[0170] S4-2-3, Crossover and Mutation: Design an adaptive crossover probability and mutation probability, dynamically adjust them according to individual fitness, and maintain population diversity.

[0171] In the mutation and crossover operations of the genetic algorithm, a fine-tuning mechanism for phase durations is introduced. By setting the range of phase duration changes and based on the maximum and minimum green light time constraints, the phase duration is randomly adjusted when generating new solutions to ensure that the new solutions not only meet the constraint conditions but also may bring performance improvement. Specifically, it includes:

[0172] a. Crossover operation

[0173] In the crossover operation, this is a crossover strategy based on phase demands. Suppose there are two parent individuals P1 and P2, and their phase durations are T 1,i and T 2,i , where i represents the i-th phase. The crossover operation is as follows:

[0174] a1. Calculate the required time T demand,i for each phase:

[0175]

[0176] where, Q current,i is the current traffic volume of the i-th phase;

[0177] Q predicted is the predicted traffic volume of the i-th phase;

[0178] C phase is the phase passing capacity of the i-th phase.

[0179] a2. Adjust the phase duration after crossover according to the required time:

[0180]

[0181] Among them, α is the cross weight, and its value range is [0, 1]; when α is close to 0, the phase duration is mainly affected by T 2,i ; when α is close to 1, it is mainly affected by T 1,i ; when α is close to 0.5, the fine-tuning degree is relatively large.

[0182] β is the demand adjustment coefficient. The larger its value, the greater the amplitude of adjusting the phase duration according to the demand time, and the more obvious the fine-tuning.

[0183] b. Mutation operation

[0184] Assume that the phase duration of an individual P is T i , and the mutation operation process is as follows:

[0185] b1. Calculate the demand time T demand,i for each phase:

[0186]

[0187] b2. Adjust the phase duration according to the demand time:

[0188] T new,i = T i + γ·(T demand,i - T i ) + δ·N(0, σ)

[0189] Among them, γ is the demand adjustment coefficient;

[0190] δ is the mutation intensity;

[0191] N(0, σ) is a normal distribution random number with a mean of 0 and a standard deviation of σ.

[0192] b3. Ensure that the adjusted phase duration meets the constraint conditions:

[0193] T new,i = max(T min , min(T new,i , T max ))

[0194] Among them, T min is the minimum target traffic phase time;

[0195] T max is the maximum target traffic phase time.

[0196] S4-2-4. Elite Retention Strategy: Finally, retain the optimal individual in each generation to ensure that the algorithm converges to the global optimal solution. In each generation, select the individual with the highest fitness as the elite individual, retain the elite individual to the next generation, and do not participate in the crossover and mutation operations. Through the elite retention strategy, ensure that the algorithm does not lose the current optimal solution and at the same time accelerate the convergence speed.

[0197] S4-3. Signal Green Waveband Optimization:

[0198] Signal green waveband optimization is to form a green waveband by coordinating the signal timings of adjacent intersections, reducing the number of vehicle stops and delay time. The core of green waveband optimization is to dynamically adjust the phase difference and green light time between adjacent intersections to ensure that vehicles can pass through multiple intersections continuously without encountering red lights. The result of green waveband optimization directly affects the value of the objective function, specifically manifested as reducing the total delay time, total number of stops, and total queue length.

[0199] S4-3-1. Influence Factor of Green Waveband on Objective Function

[0200] The goal of green waveband optimization is to reduce the number of vehicle stops and delay time on the main line by coordinating the signal timings of multiple intersections, indirectly regulating the high-level feature influence factors, and thus optimizing the objective function. As Figure 4 shown, therefore, the influence factors of green waveband optimization on the objective function include the reduction amount of total delay time, the reduction amount of total number of stops, and the reduction amount of total queue length. The green waveband optimization adjusts the objective function through the following formula:

[0201] Z new =Z old -ΔT total_delay -ΔS total -ΔQ total

[0202] Where:

[0203] Z new is the value of the objective function after green waveband optimization;

[0204] Z old is the value of the objective function before green waveband optimization;

[0205] ΔT total_delay is the total delay time reduced by green waveband optimization;

[0206] ΔS total is the total number of stops reduced by green waveband optimization;

[0207] ΔQ total is the total queue length reduced by green waveband optimization.

[0208] S4-3-2. Dynamically Adjust the Phase Difference

[0209] Calculate the average driving speed of the vehicle:

[0210]

[0211] where \(v\) i is the speed of the \(i\)-th vehicle,

[0212] and \(N\) is the number of vehicles within the statistical time period.

[0213] Calculate the ideal phase difference: Calculate the time required for the vehicle to travel from the upstream intersection to the downstream intersection:

[0214]

[0215] where \(L\) is the distance between adjacent intersections.

[0216] Dynamically correct the phase difference: Considering traffic flow fluctuations and signal cycle constraints, dynamically adjust the phase difference. When it is detected that the queue at the downstream intersection is too long, the phase difference can be shortened to release vehicles in advance:

[0217]

[0218] where \(Q\) queue is the queue length at the downstream intersection;

[0219] \(C\) phase is the lane capacity;

[0220] \(\beta\) is the correction coefficient, usually taking values between 0.1 and 0.3.

[0221] By dynamically adjusting the phase difference, the waiting time of vehicles at intersections can be reduced, thereby reducing the total delay time \(T\) total_delay .

[0222]

[0223] where,

[0224] is the waiting time before adjustment at the \(i\)-th intersection;

[0225] is the waiting time after adjustment at the \(i\)-th intersection;

[0226] and \(M\) is the number of intersections.

[0227] S4-3-3, Optimization of the green wave band width

[0228] The green wave band width refers to the green light time window during which vehicles can continuously pass through multiple intersections within one signal cycle. The purpose of dynamically adjusting the green wave band width is to maximize the trunk line capacity.

[0229] Determine the reference green wave band:

[0230]

[0231] According to the traffic flow at the upstream intersection and the queue situation at the downstream intersection, extend or shorten the green light time to match the green wave band requirements. Green light extension formula (when the upstream traffic is dense):

[0232]

[0233] Green light shortening formula (when the downstream queue is too long):

[0234]

[0235] Among them, T green,i is the green light time of phase i;

[0236] Q in,i is the traffic flow at the upstream intersection of phase i;

[0237] Q queue,i is the queuing traffic flow at the downstream intersection of phase i;

[0238] Q threshold,i is the queuing threshold of phase i, taking 80% of the lane capacity;

[0239] C phase,i is the passing capacity of phase i;

[0240] γ is the adjustment coefficient. When the upstream traffic is dense, it takes 0.5 - 1.0; when the downstream queue is too long, it takes -1 - -0.5.

[0241] By dynamically adjusting the green light time, the queuing length and the number of stops of vehicles at the intersection can be reduced, thereby reducing the total queuing length Q total and the total number of stops S total . The formula is as follows:

[0242]

[0243] Among them, is the queuing length before adjustment at the i-th intersection;

[0244] is the queuing length after adjustment at the i-th intersection;

[0245] is the number of stops before adjustment at the i-th intersection;

[0246] is the number of stops after adjustment at the i-th intersection;

[0247] M is the number of intersections.

[0248] S4-3-4, Coordinated Optimization of Green Wave Bands

[0249] By coordinating the green light times of multiple intersections, a coherent green wave band is formed.

[0250]

[0251] Among them, M is the number of intersections on the arterial road;

[0252] T green,k is the green light time of the k-th intersection;

[0253] Q pass,k is the number of vehicles passing through the k-th intersection during the green light period.

[0254] Maximizing the total width B of the green wave band total is consistent with the objective function, both aiming to improve the operating efficiency of the arterial road. When B total increases, it means that more vehicles can continuously pass through multiple intersections during the green light time, reducing the total vehicle delay time, total number of stops, and total queue length, thus achieving the optimization of the objective function.

[0255] S4-4, Reinforcement Learning-Assisted Optimization - Reward and Punishment Mechanism

[0256] Environmental Modeling: The arterial road is regarded as a reinforcement learning environment. The state space includes the current traffic state, historical signal timings, etc., and the action space is the possible signal timing adjustment schemes.

[0257] Reward Function: Design a reward function. Positive rewards include reducing vehicle delays, reducing queue lengths, and reducing the number of stops, while negative rewards include increasing congestion, increasing the number of stops, and extending queue lengths.

[0258] R = r1·(-T delay ) + r2·(-S stop ) + r3·(-Q queue )

[0259] Among them, r1, r2, and r3 are reward weights.

[0260] Policy Optimization: Adopt Q-learning to iteratively update the policy and achieve dynamic optimization of signal timings.

[0261] Q(s,a) ← Q(s,a) + α(r + γmax a' Q(s′,a′) - Q(s,a))

[0262] Among them, α is the learning rate;

[0263] γ is the discount factor;

[0264] r is the reward;

[0265] s is the current state;

[0266] a is the current action;

[0267] s′ is the next state;

[0268] a′ is the next action.

[0269] In the Q - learning algorithm, the calculation result of the reward function R will affect the iterative update of the policy. When a certain signal timing adjustment plan (action a) is adopted, the reward r obtained according to the current traffic state (state s) will affect the update of the Q - value. If the reward r is positive, that is, the vehicle delay, queue length or number of stops is reduced, then the value of Q(s,a) will increase, indicating that this action is more conducive to achieving the optimization goal and is more likely to be adopted in subsequent policy selections, thus prompting the optimization algorithm to move in the direction of reducing the total delay time T delay , the total number of stops S stop and the total queue length Q queue of the objective function for signal timing adjustment. Conversely, if the reward r is negative, the value of Q(s,a) will decrease, and the algorithm will try to avoid selecting this action, enabling the optimization algorithm to continuously optimize the signal timing to meet the requirements of the objective function.

[0270] S5. Implementation and feedback mechanism

[0271] Real - time execution: The optimized signal timing plan is sent to each intersection signal controller through communication technology to achieve real - time adjustment.

[0272] Performance monitoring: Establish a real - time monitoring system to track the changes in traffic states after signal adjustment and evaluate the optimization effect.

[0273] Feedback loop: According to the monitoring results, regularly adjust the model parameters and optimization algorithm to form a closed - loop feedback mechanism and continuously improve the optimization effect.

[0274] Specific real - time case:

[0275] 1. Data collection stage:

[0276] Select multiple representative intersections on a certain urban traffic artery for data collection. Taking one intersection as an example, a traffic volume, vehicle speed, and vehicle queue length information are collected every minute through a radar - vision integrated machine and uploaded to the central system in real - time; engineers carefully check and input the geographical location, number of lanes, and signal phase configuration information of this intersection; at the same time, obtain the traffic flow historical data of the past week from a professional traffic flow monitoring system, connect with the meteorological data platform and the government holiday arrangement database to obtain real - time weather data and holiday information.

[0277] During the data collection process, a data quality monitoring mechanism is established to perform real-time verification on the collected data. If abnormal fluctuations or data beyond the reasonable range are found, the data review process is immediately initiated, and data confirmation is carried out through manual intervention or other backup collection devices to ensure the accuracy of the data.

[0278] 2. Data preprocessing stage:

[0279] For the collected raw data, the Kalman filtering technique is first used to remove noise. Taking traffic flow data as an example, a state equation and an observation equation are constructed according to its historical change pattern. After Kalman filtering, the 3σ principle is used to identify and remove outliers. If the vehicle speed data is missing in a certain period due to equipment failure, the time series prediction method based on LSTM is used for filling. During the training process of the LSTM model, reasonable training parameters such as the number of training epochs and batch size are set to improve the prediction accuracy. Finally, all data is subjected to min-max normalization processing to convert it into a format suitable for input to the graph neural network.

[0280] 3. Graph data structure construction and enhancement stage:

[0281] Each intersection is used as a node of the graph, and the road segment is used as an edge to construct the graph structure. For two adjacent intersections, the edge attributes such as traffic flow and queue length are determined according to the real-time traffic data of their road segments. Timestamps are introduced at one-minute time intervals to construct a time series graph. Every hour, the weights of the weighted adjacency matrix are recalculated according to the road segment passing capacity and historical traffic flow. For example, if the passing capacity of a certain road segment increases within a certain hour and the historical traffic flow is relatively stable, the weights between nodes change accordingly according to the formula, thereby enhancing the graph structure's ability to express traffic dynamic changes.

[0282] During the construction of the graph structure, special connection relationships between intersections, such as one-way road segments and tidal lanes, are considered, and special processing is carried out on the graph structure and the weight calculation method to more accurately reflect the actual traffic network. For one-way road segments, in the graph structure, they are represented as directed edges, that is, a directed connection is established from the starting intersection to the ending intersection to clarify the driving direction of vehicles and avoid incorrect representations of reverse connections. In weight calculation, the passing capacity of the road segment only considers the passing capacity of vehicles along the one-way driving direction, and the historical road segment traffic flow only counts the traffic flow data in this direction. For tidal lanes, according to the change of the passing direction in different periods, the direction of the edges in the graph structure is dynamically adjusted. After determining the actual passing direction of the tidal lane within a specific period, a directed edge is established in the same way as for one-way road segments. At the same time, the weight calculation needs to be updated in real time. According to the change of the passing capacity of the tidal lane in different periods, the value of the passing capacity of this phase is adjusted, and the weight is recalculated in combination with the historical traffic flow in the corresponding direction during this period.

[0283] 4. Construction and Optimization Stage of Graph Neural Network:

[0284] Select the Graph Convolutional Network (GCN) as the model architecture, randomly initialize the model parameters, set the learning rate to 0.001, the number of hidden layers to 3, and the number of hidden units in each layer to 64. Divide the processed graph data into training set, validation set, and test set according to the ratio of 6:2:2.

[0285] During the training process, input the training set data into the model for forward propagation, calculate the prediction results, use the weighted mean square error as the loss function to calculate the error, update the model parameters through the backpropagation algorithm and Adam optimizer, and evaluate the model performance using the validation set after each round of training to adjust the hyperparameters. After multiple rounds of training, when the root mean square error of the model on the test set is less than 10% and the accuracy is greater than 90%, it is considered that the model training is qualified. To further improve the generalization ability of the model, data augmentation techniques are adopted during the training process, such as performing random perturbations and rotations on the input data to increase the diversity of the data.

[0286] 5. Signal Timing Optimization and Adjustment Stage:

[0287] Use the genetic algorithm to optimize the signal timing plan, and perform real-number encoding on the phase duration and phase sequence. According to the actual traffic conditions of this traffic artery, determine the weight coefficients in the fitness function. In the genetic operation, dynamically adjust the crossover probability and mutation probability according to the individual fitness, and retain the optimal individual in each generation.

[0288] In terms of adjusting the phase duration, calculate the phase demand time based on the current and predicted traffic volumes, and then adjust the green light time. For example, if the current traffic volume of a certain phase is large and the predicted traffic volume will continue to increase in the future, calculate the demand time according to the formula, and appropriately extend the green light time of this phase under the constraints of the maximum and minimum green light times.

[0289] In terms of optimizing the signal green wave band, calculate the average vehicle driving speed, ideal phase difference, and dynamic correction phase difference, and adjust the width of the green wave band and the green light time according to the upstream and downstream traffic conditions. If it is found that the queue at the downstream intersection is too long, shorten the phase difference and appropriately shorten the green light time to relieve congestion.

[0290] At the same time, use reinforcement learning to assist in optimization. Take the real-time traffic state of the traffic artery as the environmental state, the possible signal timing adjustments as actions, give corresponding rewards according to the reward function, and continuously iterate and update the strategy through the Q-learning algorithm to achieve the dynamic optimization of signal timing. During the reinforcement learning process, set a reasonable learning rate decay strategy. As the number of training times increases, gradually reduce the learning rate to make the model converge more stably.

[0291] 6. Implementation and Feedback Mechanism Phase:

[0292] The optimized signal timing plan is sent to the signal controllers at each intersection through a wired network or 5G wireless communication technology to achieve real-time adjustment. Sensors and cameras are deployed at each intersection to collect traffic status data in real time, such as traffic flow, vehicle speed, queue length, etc., and a monitoring system is established to evaluate the optimization effect. The monitoring data is analyzed weekly. If it is found that the traffic congestion at a certain intersection is still relatively serious during a specific period, the relevant model parameters of this intersection are adjusted accordingly, such as adjusting the weight of the corresponding node of this intersection in the graph neural network, or optimizing the relevant parameters in the genetic algorithm and reinforcement learning, such as adjusting the crossover probability of the genetic algorithm and the reward weight of the reinforcement learning, to form a closed-loop feedback mechanism and continuously improve the operation efficiency of the traffic artery.

[0293] Although the present invention has been described in detail with general descriptions and specific embodiments above, based on the present invention, some modifications or improvements can be made, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.

Claims

1. A real-time optimization method for coordinated control of arterial roads, characterized in that, It includes the following steps: S1. Collect traffic data and perform data preprocessing to convert the traffic data into data suitable for input to the graph neural network; S2. Construct a graph data structure, set intersections as nodes and road segments as edges, and then adjust the weights of the weighted adjacency matrix at a fixed frequency according to the traffic data and historical traffic volume of the road; S3. Use the graph data structure in S2 as the input, select the graph convolutional network as the basic architecture, and construct a graph neural network model for the traffic artery to capture the spatio-temporal features between nodes; S4. Formulate and adjust the objective function according to the spatio-temporal features in combination with the actual traffic demand and constraints, use the genetic algorithm to solve the objective function, continuously iterate and update the population through crossover and mutation, and evaluate individuals according to the objective function and constraints in each iteration, and finally obtain the optimal signal timing plan.

2. The real-time optimization method for coordinated control of arterial roads according to claim 1, wherein: The weight formula of the weighted adjacency matrix in S2 is where ω ij is the weight between nodes i and j; C ij is the traffic capacity of the road section; Q ij is the traffic flow of the historical road section.

3. The real-time optimization method for coordinated control of arterial roads according to claim 1, characterized in that: The training method of the graph neural network in S3 includes dividing the traffic data contained in the graph data structure obtained in S2 into a training set, a validation set and a test set according to a reasonable ratio; Select the graph convolutional network as the basic architecture, extract the features of the node and its neighbor nodes through convolutional operations to obtain the initial graph neural model; Input the training set data into the initial graph neural model, calculate the prediction result of the model through forward propagation, calculate the loss function according to the prediction result and the traffic flow data in the traffic data, use the backpropagation algorithm to calculate the gradient of the loss function with respect to the parameters of the graph neural network model, and use the Adam optimizer to update the model parameters to minimize the loss function; Finally, use the test set data to evaluate the trained model to obtain the graph neural network model.

4. A real-time optimization method for coordinated control of arterial roads according to claim 1, characterized in that: The optimization algorithm in S4 adopts the genetic algorithm, which includes mapping the traffic signal phase duration and traffic signal phase order to the chromosomes of the genetic algorithm, comprehensively considering multiple indicators such as vehicle delay time, stop times, queue length, and advanced feature influence factors, and constructing a comprehensive fitness function: where T total_delay is the total delay time; S total is the total number of parking times; Q total is the total queue length; F advanced For the influence primer of advanced features; ω5, ω6, ω7, ω8 are weight coefficients.

5. A real-time optimization method for coordinated control of arterial roads according to claim 1 or 4, characterized in that: In the mutation and crossover operations of the genetic algorithm, a fine-tuning mechanism for the phase duration is introduced. By setting the range of phase duration changes, based on the maximum and minimum green light time constraints, the phase duration is randomly adjusted when generating new solutions to ensure that the new solutions not only meet the constraint conditions but also may bring performance improvement.

6. The real-time optimization method for coordinated control of arterial roads according to claim 5, characterized in that: In each iteration, select the individual with the highest fitness as the elite individual, and the elite individual does not participate in the crossover and mutation operations and is reserved for the next generation to ensure that the algorithm does not lose the current optimal solution and at the same time speeds up the convergence speed.

7. The real-time optimization method for coordinated control of arterial roads according to claim 1, characterized in that Optimize the objective function of the optimal signal timing plan in S4 for the green wave band, and the green wave band optimization adjusts the objective function through the following formula: Z new = Z old - ΔT total_delay - ΔS total - ΔQ total Where: Z new The objective function value after optimization for the green wave band; Z old is the objective function value before the optimization of the green wave band; ΔT total_delay is the reduced total delay time obtained by the green wave band optimization; ΔS total is the reduced total number of stops obtained by the green wave band optimization; ΔQ total The reduced total queue length obtained by green wave band optimization.

8. A real-time optimization method for coordinated control of arterial roads according to claim 7, characterized in that, The dynamic adjustment phase difference calculation method for the green wave band optimization includes: First calculate the average vehicle speed: where v i is the speed of the i-th vehicle, N is the number of vehicles in the statistical time period; Then calculate the time required for the vehicle to travel from the upstream intersection to the downstream intersection: Where L is the distance between adjacent intersections; Then consider the traffic flow fluctuation and signal cycle constraint, and dynamically adjust the phase difference. When it is detected that the queue at the downstream intersection is too long, the phase difference can be shortened to release the vehicle in advance: Among them, Q queue is the queue length at the downstream intersection; C phase is the lane capacity; β is a correction coefficient, usually taken as 0.1 - 0.3; Finally, by dynamically adjusting the phase difference, the waiting time of vehicles at intersections can be reduced, thereby reducing the total delay time T total_delay : Among them, is the waiting time before adjustment for the i-th intersection; The adjusted waiting time for the i-th intersection; M is the number of intersections.

9. A real-time optimization method for coordinated control of arterial roads according to claim 8, characterized in that, The coordinated optimization of the green wave band forms a coherent green wave band by coordinating the green light time of multiple intersections: Among them, M is the number of intersections on the main line; T green,k is the green light time for the k-th intersection; Q pass,k is the number of vehicles passing through the k-th intersection during the green light period.

10. A real-time optimization method for coordinated control of arterial roads according to claim 1, characterized in that, Design a reward function to optimize the update iteration of the genetic algorithm. The positive rewards in the reward function include reducing vehicle delay, reducing queue length, and reducing the number of stops, while the negative rewards include increasing congestion, increasing the number of stops, and extending the queue length; R = r1·(-T delay ) + r2·(-S stop ) + r3·(-Q queue ) Among them, r1, r2, and r3 are reward weights; The reward function adopts Q-learning and iteratively updates the strategy to achieve dynamic optimization of signal timing; Q(s,a) ← Q(s,a) + α(r + γ max a' Q(s′,a′) - Q(s,a)) Among them, α is the learning rate; γ is the discount factor; r is the reward; s is the current state; a is the current action; s ' For the next state; a ' For the next action.

Images