A highway intelligent operation traffic guide method and system

Abstract

The application discloses a highway intelligent operation traffic guide method and system, and belongs to the technical field of intelligent traffic. Real-time traffic flow state data of a target road network is acquired, and short-time traffic flow prediction results are generated. A multi-dimensional guide strategy library containing lane level, speed level, road right level and information level is constructed, and each strategy is associated with a strength parameter. An improved artificial bee colony algorithm is adopted, a strategy type selection and a strength parameter assignment are taken as joint decision variables, and a multi-objective optimization function is taken as a target to dynamically search an optimal strategy combination scheme. Control instructions are generated according to the optimal strategy combination scheme and are executed. Feedback data is collected to update decision parameters of an optimizer online. The application can realize dynamic and autonomous optimization of guide strategy combination, effectively improves road network passing efficiency, operation safety and travel time reliability, and is suitable for complex and changeable intelligent highway operation scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically to a method and system for intelligent traffic diversion in highway operation. Background Technology

[0002] Currently, with the rapid development of smart highway construction, various traffic diversion strategies such as variable speed limits, ramp control, and dynamic lane management have been widely applied in practical engineering. Existing technologies typically employ a fixed model of "preset strategy type + parameter optimization," meaning that the strategy combination type (such as a combination of dedicated lanes and speed limits, or a combination of speed limits and ramp control) is pre-fixed. Then, offline simulation of the preset objective function curve or the construction of a feedback control model are used to optimize the parameters in the strategy (such as speed limit values, green light duration, and number of lanes), thereby generating a traffic diversion scheme. This, to a certain extent, improves the traffic efficiency and operational safety of highways.

[0003] However, existing technical solutions have fixed combinations of strategy types, making it impossible to flexibly select or combine different diversion strategy types according to the dynamic changes in traffic scenarios, resulting in a single control method and insufficient adaptability. Secondly, the selection of strategy types and the assignment of intensity parameters are separated into two independent steps, making it difficult to achieve coordinated optimization between strategy types and execution intensity. Thirdly, the decision space is limited to the parameter space, making it impossible to dynamically optimize at the strategy type level, which limits the adaptability and optimization potential of diversion solutions. Fourthly, existing technologies mostly rely on offline simulation and preset function curves, making it difficult to achieve online real-time response to changes in traffic flow status and iterative evolution of strategies.

[0004] Therefore, how to propose a smart highway operation traffic diversion method and system to achieve online autonomous optimization of strategy combination schemes to adapt to complex and ever-changing traffic operation scenarios is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for intelligent traffic diversion in highway operation, which dynamically generates a combination of diversion strategies adapted to traffic scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: On the one hand, this invention proposes a traffic diversion method for intelligent highway operation, comprising the following steps: S1. Obtain real-time traffic flow status data of the target road network, and generate short-term traffic flow prediction results based on the real-time traffic flow status data; S2. Construct a multi-dimensional traffic redirection strategy library, which contains a variety of preset strategy types. Each strategy type is associated with an intensity parameter, which is used to quantify the degree of execution or control strength of the corresponding strategy type. S3. Using the real-time traffic flow status data and the short-term traffic flow prediction results as input, an improved artificial bee colony algorithm is adopted, with the selection of strategy type and the assignment of intensity parameters as joint decision variables, and with a preset multi-objective optimization function as the objective, the optimal strategy combination scheme is dynamically searched to obtain the optimal strategy combination scheme. S4. Generate control commands based on the optimal strategy combination scheme, and execute the control commands through the roadside dissemination device; S5. Collect traffic flow feedback data after the control command is executed, and use the feedback data to update the decision parameters of the optimizer online.

[0007] Preferably, S1 includes: S11. Real-time traffic flow parameters are collected at preset time intervals by roadside sensing devices deployed at each section of the target road network; S12. Perform spatiotemporal alignment and missing data interpolation on the real-time traffic flow parameters to form a time series data matrix; S13. Input the time series data matrix into a pre-trained multi-scale time series prediction model to generate short-term traffic flow prediction results for multiple future time windows.

[0008] Preferably, the multi-scale temporal prediction model includes a graph convolutional layer, an LSTM layer, and a multi-head attention mechanism layer connected in sequence; The graph convolutional layer performs graph convolution operations on the time-series data matrix based on the road network topology, extracts spatial correlation features between each cross section, and outputs a spatial feature sequence; the LSTM layer performs temporal modeling on the spatial feature sequence, extracts the time-series dependency features of each cross section, and generates a temporal feature sequence; the multi-head attention mechanism layer performs weighted fusion of the temporal feature sequence with the proximity features, periodic features, and trend features extracted based on the time-series data matrix to obtain the short-term traffic flow prediction result.

[0009] Preferably, the multi-dimensional traffic redirection strategy library includes: Lane-level control strategies include the number of dynamically activated lanes, the number of dynamically closed lanes, the type and location of dedicated lanes, and the mileage and time period for emergency lane borrowing; Speed-level control strategies include segmented variable speed limits, coordinated speed limits, and speed limit change gradients; Right-of-way management strategies include ramp signal timing parameters, toll station entrance control ratio, and service area traffic diversion ratio; Information-level control strategies include route guidance recommendation ratio, destination prompt intensity, and congestion warning release scope.

[0010] Preferably, S3 includes: S31. Randomly generate an initial population within a preset range of values ​​for joint decision variables, wherein each individual in the initial population corresponds to a combination of a strategy type selection vector and an intensity parameter vector; S32. Update the population using a population update guidance mechanism based on a historical good solution memory table; S33. The Pareto optimization framework is used to perform non-dominated sorting and crowding distance calculation on the updated population, and the Pareto optimal solution set is updated. S34. Select a predetermined number of optimal solutions from the updated Pareto optimal solution set and update the historical excellent solution memory table; S35. Determine whether to trigger the adaptive reconnaissance mechanism based on the state of the Pareto optimal solution set, and generate a new reconnaissance bee solution within the neighborhood range determined by the selected chaotic map, using the optimal solution in the updated historical good solution memory table as a reference point. S36. Add the new solution of the scout bee to the population, calculate the multi-objective function value of each solution in the updated Pareto optimal solution set, and record the solution with the best multi-objective function value; S37. Determine if the termination condition is met. If it is, proceed to S38; otherwise, return to S32 for the next iteration. S38. Output the solution corresponding to the recorded optimal multi-objective function value as the optimal strategy combination scheme.

[0011] Preferably, S32 includes: S321. Initialize the historical excellent solution memory table, set the storage length of the memory table, and store individuals in the initial population that meet the preset multi-objective function threshold as initial excellent solutions into the memory table; S322. The hired bee generates a new combination of strategy type selection vector and intensity parameter vector as a new solution based on each individual in the current population and within a preset neighborhood range through random perturbation. S323. Calculate the multi-objective function value of each new solution and the corresponding original population individual, and retain the individual with the better multi-objective function value among each new solution and the corresponding original individual to form the population after preliminary screening; S324. Extract the stored superior solutions from the historical superior solution memory table, calculate the multi-objective function value of the superior solutions, and perform a second greedy selection with the corresponding individuals in the population after the initial screening, retaining the individuals with better multi-objective function values, and completing the population update.

[0012] Preferably, S33 includes: S331. Set optimization goals; S332. Individuals in the unsorted population are compared pairwise according to the optimization objective, and a fast non-dominated sort is performed to classify the population levels; the unsorted population includes the updated population and the previous generation Pareto optimal solution set. S333. Calculate crowding distance for individuals within the same level; S334. Individuals are selected according to the rules of priority by level and priority by distance in the same level of crowding until the number of individuals is consistent with the initial population size, thus forming an updated Pareto optimal solution set.

[0013] Preferably, in S35: The convergence state is determined based on the state of the Pareto optimal solution set; the convergence state includes premature convergence risk, normal convergence, and good convergence. An adaptive detection mechanism based on chaotic mapping is triggered according to the convergence state, including: If the risk of premature convergence is identified, a large-scale chaos detection will be triggered. If the convergence is determined to be normal, then a medium-range chaos detection is triggered. If the convergence is determined to be good, a small-scale chaos detection is triggered. Based on the selected chaotic mapping type and search step size, a new solution for the scout bee is generated with the optimal solution in the historical good solution memory table as a reference point.

[0014] On the other hand, the present invention also proposes a smart highway operation traffic diversion system, comprising: The data acquisition module is used to acquire real-time traffic flow status data of the target road network and generate short-term traffic flow prediction results based on the real-time traffic flow status data. The strategy library construction module is used to build a multi-dimensional traffic redirection strategy library. The traffic redirection strategy library contains a variety of preset strategy types. Each strategy type is associated with a strength parameter, which is used to quantify the execution degree or control strength of the corresponding strategy type. The strategy search module is used to take the real-time traffic flow status data and the short-term traffic flow prediction results as input, adopt an improved artificial bee colony algorithm, use the selection of strategy type and the assignment of intensity parameters as joint decision variables, and use a preset multi-objective optimization function as the objective to dynamically search for the optimal strategy combination scheme. The control execution module is used to generate control commands based on the optimal strategy combination scheme and execute the control commands through the roadside publishing device. The feedback optimization module is used to collect traffic flow feedback data after the control command is executed, and to use the feedback data to update the decision parameters of the optimizer online.

[0015] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for intelligent traffic diversion in highway operation. Firstly, it accurately obtains the short-term evolution trend of road network traffic flow by fusing spatial graph convolution, long short-term memory networks, and multi-head attention mechanisms through a multi-scale temporal prediction model. Then, it constructs a multi-dimensional diversion strategy library including lane-level, speed-level, right-of-way-level, and information-level strategies. Strategy type selection and intensity parameter assignment are used as joint decision variables. An improved artificial bee colony algorithm is used as the optimizer. It guides the population to converge quickly through a historical good solution memory table, maintains a high-quality non-dominated solution set through a Pareto optimization framework, and dynamically escapes local optima based on an adaptive chaotic detection mechanism based on the convergence state. This achieves online search for the optimal strategy combination and closed-loop feedback updates. This invention can improve the dynamic adaptability and optimization efficiency of diversion strategies, autonomously discover and execute the optimal control combination with the best collaborative effect in complex and ever-changing traffic scenarios, effectively balancing traffic efficiency, operational safety, and travel time reliability, realizing a leap from passive response to proactive prevention, and from single strategy to full-domain collaboration. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 A flowchart of the method provided by the present invention; Figure 2 The system architecture diagram provided for this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] On the one hand, refer to Figure 1 This invention discloses a traffic diversion method for intelligent highway operation, comprising the following steps: S1. Obtain real-time traffic flow status data of the target road network, and generate short-term traffic flow prediction results based on the real-time traffic flow status data, including: S11. Real-time traffic flow parameters are collected at preset time intervals using roadside sensing devices deployed at various cross-sections of the target road network. These roadside sensing devices include microwave radar, checkpoint cameras, and road surface detectors. The collected real-time traffic flow parameters include the average vehicle speed at each cross-section. Traffic flow and lane occupancy ,in Indicates the section number. Indicates the current moment.

[0020] S12. Perform spatiotemporal alignment and data imputation on the collected real-time traffic flow parameters to form a complete time-series data matrix. ,in The total number of cross-sections. .

[0021] S13. Input the time series data matrix into a pre-trained multi-scale time series prediction model to generate short-term traffic flow prediction results for multiple future time windows.

[0022] The multi-scale temporal prediction model comprises a graph convolutional layer, a long short-term memory network layer, and a multi-head attention mechanism layer connected sequentially. The graph convolutional layer performs graph convolution operations on the time-series data matrix based on the road network topology, extracting spatial correlation features between cross-sections and outputting a spatial feature sequence. The long short-term memory network layer performs temporal modeling on the spatial feature sequence, extracting the temporal dependency features of each cross-section and generating a temporal feature sequence. The multi-head attention mechanism layer performs weighted fusion of the temporal feature sequence with neighboring features, periodic features, and trend features extracted from the time-series data matrix to obtain short-term traffic flow prediction results.

[0023] Specifically, the graph convolutional layer extracts spatial correlation features based on the road network topology. First, a road network topology adjacency matrix is ​​constructed. Road network topology adjacency matrix It is The square formation, Total number of cross sections. Matrix elements. Indicates cross-section and cross-section The connectivity between them. Defined as follows: If cross-section and cross-section If they are physically adjacent (including upstream and downstream relationships, connections between different lanes on the same cross-section, and merging / merging relationships between ramps and the mainline), then ;otherwise ; Typically, self-connection is configured. This ensures that the node's own information is also preserved.

[0024] Degree matrix It is A diagonal matrix whose diagonal elements , i.e., node The number of neighbors (including itself). The degree matrix is ​​used to normalize the adjacency matrix, so that graph convolution operations are not affected by differences in node degree. The normalized adjacency matrix is: .

[0025] The core operation of graph convolutional layers is to aggregate the features of neighboring nodes: ; in After L layers of graph convolution, the output spatial feature sequence is obtained. , For the hidden layer dimension.

[0026] The physical meaning of this formula is that the new feature vector of each cross section is obtained by weighted summation of the features of itself and its adjacent cross sections, and the weights are determined by the topological structure.

[0027] Long Short-Term Memory (LSTM) network layers perform temporal modeling of spatial feature sequences, extract time-series dependent features, and input the spatial feature sequences into the LSTM step by step. ; in For a moment The time-series feature output, This represents the dimension of the LSTM hidden layers. After iterating through all time steps, the temporal feature sequence is obtained. .

[0028] A multi-head attention mechanism layer fuses multi-scale temporal features. First, it starts from the original time-series data matrix... Extracting neighboring features Periodic characteristics and trend characteristics .

[0029] Let the current prediction target time be The prediction time window length is Time granularity is (e.g., 5 minutes). Definition: Proximity features : Take historical data from several consecutive time windows prior to the current moment to reflect short-term trends.

[0030] ; in This represents the number of adjacent time windows.

[0031] Periodic characteristics This method uses data from the same relative moment in one or more previous cycles to reflect daily periodic patterns. The cycle is one day, and within a day there are... If there is a time window, then: ; in For example, take the number of repetitions in a periodic cycle. That is, the same moment of the previous day and the previous two days.

[0032] Trend characteristics This method uses data from the same relative moment within one or more longer periods (such as a week) to reflect the periodic pattern. If a week is taken as the period, there are a total of [data points] within a week. If there is a time window, then: ; in The number of repetitions of the trend (e.g., taking) That is, the same moment of the previous week and the previous two weeks.

[0033] The above three features constitute three feature sequences, which, together with the temporal feature sequence output by the LSTM, form three feature sequences. Multi-head attention mechanism for shared input.

[0034] The input to a multi-head attention mechanism consists of three parts: Query Matrix Temporal feature sequences from LSTM output This indicates the query information that needs to be focused on at this time; Key matrix The attention weights are calculated using proximity features, periodic features, trend features derived from the explicit construction, and temporal features from the LSTM output. Value matrix These features are also used for weighted summation to obtain the fused output.

[0035] Specifically, the temporal feature sequence is first concatenated with three explicit features to form a comprehensive feature matrix. : ; Then, the query, key, and value matrix is ​​obtained through linear transformation: ; in This is a learnable weight matrix.

[0036] The attention mechanism calculates the weighted fusion result at each time step: ; in is the dimension of the key vector, used to scale the dot product to prevent gradient vanishing.

[0037] The multi-head attention mechanism computes multiple attention heads in parallel, concatenates their outputs, and then performs a linear transformation. ; in, .

[0038] Output of multi-head attention mechanism In order to generate the future The prediction results for each time window are taken. The portion corresponding to the future time step is mapped to the prediction dimension through a fully connected layer: ; Output As a short-term traffic flow forecast result, it includes the predicted vehicle speed at each cross-section within future time windows. Predicted flow and predicted market share .

[0039] The multi-head attention mechanism can dynamically learn the importance weights of features at different time scales for prediction, achieving adaptive weighted fusion rather than simple fixed weight averaging.

[0040] S2. Construct a multi-dimensional traffic redirection strategy library. The traffic redirection strategy library contains a variety of preset strategy types. Each strategy type is associated with a strength parameter, which is used to quantify the degree of execution or control of the corresponding strategy type.

[0041] The multi-dimensional traffic redirection strategy library includes: Lane-level control strategies include the number of dynamically activated lanes, the number of dynamically closed lanes, the type and location of dedicated lanes, and the mileage and time period for emergency lane borrowing; Speed-level control strategies include segmented variable speed limits, coordinated speed limits, and speed limit change gradients; Right-of-way management strategies include ramp signal timing parameters, toll station entrance control ratio, and service area traffic diversion ratio; Information-level control strategies include route guidance recommendation ratio, destination prompt intensity, and congestion warning release scope.

[0042] The intensity parameter associated with each strategy type is either a continuous value or a discrete hierarchical value. The value range is preset according to the physical constraints and traffic control specifications of the strategy type. Furthermore, there is a preset mapping relationship between the intensity parameter and the control parameters of the corresponding execution agency, as shown in Table 1. Table 1 Mapping Relationship between Strategy Type and Strength Parameter

[0043] S3. Using real-time traffic flow status data and short-term traffic flow prediction results as input, an improved artificial bee colony algorithm is employed. The selection of strategy type and the assignment of intensity parameters are used as joint decision variables. A preset multi-objective optimization function is used as the objective to dynamically search for the optimal strategy combination scheme, including: S31. Randomly generate an initial population within the preset range of values ​​for the joint decision variables. Each individual in the initial population corresponds to a combination of a strategy type selection vector and an intensity parameter vector.

[0044] First, determine the mathematical representation of the joint decision variables and the search space. The strategy type selection vector is represented as follows: ,in Indicates the first Whether a preset strategy is selected This represents the total number of policies in the policy library. The strength parameter vector is represented as follows: ,in The strength parameter value corresponding to the selected strategy. This represents the number of strategies selected. Each individual is a hybrid encoding vector composed of these two parts.

[0045] The default value range for the variable selected for each strategy type is: The value range of each intensity parameter is pre-defined based on the physical constraints of the strategy type and traffic control regulations. The initial population is generated using a uniform random sampling strategy. The initial population size is set to [value missing]. For the first in the population Each individual, its strategy type selection vector Each component in The intensity parameter vector is randomly selected from the values. Each component is randomly generated according to a uniform distribution within its corresponding value range. This process can be represented by the following formula: ; in, This indicates that 0 or 1 is generated randomly. Indicates in Uniformly distributed random numbers within an interval. This is how to generate... A random combination of strategies constitutes the initial population. Each individual .

[0046] S32. Update the population using a population update guidance mechanism based on a historical good solution memory table, including: S321. Initialize the historical excellent solution memory table, set the storage length of the memory table, and store individuals in the initial population that meet the preset multi-objective function threshold as initial excellent solutions into the memory table.

[0047] First, initialize a specific data structure in the computer memory as a memory table for historical good solutions. The memory table is designed as a fixed-length first-in-first-out queue, with its maximum storage length set to [value missing]. Storage length The setup needs to balance two factors: it must retain enough historical information to provide effective guidance, but it must also avoid storing too many outdated solutions, which would reduce the timeliness of the search. Typically... The initial population size can be taken as... 1 / 5 to 1 / 3.

[0048] Next, the initial population needs to be calculated. Each individual Multi-objective function values A multi-objective function is used to quantitatively evaluate the merits of a strategy combination scheme. Its core is to measure the scheme's comprehensive performance across three dimensions: traffic efficiency, operational safety, and travel time reliability. The smaller the multi-objective function value, the better the strategy combination scheme.

[0049] Then, set an initial multi-objective function screening threshold. This threshold can be set as the average of the multi-objective function values ​​of all individuals in the initial population, i.e.: ; All multi-objective function values ​​in the initial population Individuals are sorted in ascending order of their multi-objective function values, and then stored sequentially in a memory table. until full One slot. If the number of individuals meeting the criteria is less than... If the strategy combination scheme is not found, then all of them will be stored there, and the remaining slots will be left empty for the time being. Each entry in the memory table stores not only the strategy combination scheme, but also... It also stores the corresponding multi-objective function values. And a timestamp or iteration count marker for possible subsequent update or eviction strategies.

[0050] S322. The hired bee generates a new combination of strategy type selection vector and intensity parameter vector as a new solution based on each individual in the current population and within a preset neighborhood by randomly perturbing it.

[0051] During the hired bee phase, new solutions are generated through an adaptive learning mechanism. For the first solution in the current population... The number of hired bee individuals, in the first The strategy combination scheme in the next iteration is denoted as... New interpretation By superimposing a comprehensive perturbation vector on the current solution The generation, its mathematical expression is as follows .

[0052] The combined perturbation vector It consists of two core parts: a directional master learning term and an adaptive noise term. Its structure is as follows: ; The first term is the directional master learning term, which aims to guide the individual towards the historical optimal solution; the second term... This is an adaptive noise term used to maintain population diversity.

[0053] In the targeted main learning items, This represents the optimal strategy combination scheme with the best global multi-objective function value extracted from the historical good solution memory table up to the current iteration. (Matrix) It is an adaptive rotation and scaling matrix used to change the learning direction according to the distribution characteristics of the current population. The construction is based on the covariance matrix of the current population parameters, specifically defined as: ; here, This represents the covariance matrix calculated from the joint decision variables of all individuals in the current population, reflecting the correlation between variables and the shape of the population distribution. (Function) This indicates that performing a Jouleskiy decomposition on the matrix yields a lower triangular matrix, which geometrically represents a rotational transformation of the decision space aligned with the principal axes of the population distribution ellipsoid. Parameters It is a very small positive constant used to ensure the positive definiteness of the matrix. It is an identity matrix.

[0054] The entire perturbation process is subject to an adaptive learning rate. The learning rate is globally controlled. It decays with each iteration, and its variation is described by the following formula: ; in, and The minimum and maximum learning rates are defined, and $$\gamma$$ is the decay coefficient, which controls the decay rate. This is the preset maximum number of iterations. This design allows the algorithm to use a larger learning rate in the early stages of iteration to promote extensive exploration, and a smaller learning rate in the later stages to allow for fine-tuning.

[0055] The adaptive noise term follows a normal distribution with zero mean and a diagonal covariance matrix, i.e.: ; Noise intensity in each dimension It is dynamically adaptive and is determined by the following formula: ; In the formula, Is the current population in the 19th century? The standard deviation of a decision variable in a given dimension reflects the population diversity in that dimension. This is the preset search range length for this dimension parameter; The multi-objective function value for the current individual. This is the worst multi-objective function value in the current population. It is the global noise intensity adjustment coefficient.

[0056] This design ensures that noise intensity is directly proportional to the population's diversity in that dimension and inversely proportional to the individual's relative quality. When the population converges and diversity decreases, the noise automatically decreases to facilitate development; for individuals with poor multi-objective function values, stronger noise is applied to help them escape the current region.

[0057] To ensure the new solution lies within a valid decision space, boundary constraints must be applied after generation. This applies to the strategy type selection vector. The components in the middle will constrain the values ​​to... (by rounding to the nearest integer); for the intensity parameter vector The components in the middle adopt a reflection boundary strategy: if Less than the lower bound of the search Then it will be corrected to If it is greater than the upper search bound Then it is corrected to This reflective treatment can maintain population diversity near the boundary better than simple truncation.

[0058] S323. Calculate the multi-objective function value of each new solution and the corresponding original population individual, and retain the individual with the better multi-objective function value among each new solution and the corresponding original individual to form the population after preliminary screening.

[0059] For the first in the population Each individual, whose original solution is denoted as... The new interpretation is recorded as For each pair of solutions and Perform a greedy selection operation: compare the multi-objective function values ​​of the two. The solution with the smaller multi-objective function value is retained as the candidate solution for that position in the initial screening of the population. The selection rule can be formally represented as: ; For all in the population After each hired bee performs the above calculations and selections, it constitutes the initial screening population. This step, based on local search information, completed the first round of quality improvement for the population.

[0060] S324. Extract the stored superior solutions from the historical superior solution memory table, calculate the multi-objective function value of the superior solutions, and perform a second greedy selection with the corresponding individuals in the population after the initial screening. Retain the individuals with better multi-objective function values ​​to complete the population update.

[0061] Historical Excellent Solution Memory Table It is a fixed length The first-in-first-out queue continuously stores the optimal strategy combinations and their corresponding multi-objective function values ​​discovered in previous iterations of the algorithm.

[0062] From memory table Extract These excellent solutions constitute the set of historical excellent solutions. If the number of solutions actually stored in the memo is less than... Then, a cyclic extraction method is used until the required amount is reached. For each extracted historical excellent solution... Its multi-objective function value It can be read directly from the associated data in the memory table.

[0063] Perform a second greedy selection. Set the historical best solutions together. Compared with the initial screening population Individuals in the index Perform a one-to-one comparison. For each index position... Comparing historically superior solutions Multi-objective function values Compared with the initial screening solution Multi-objective function values The solution with the smaller multi-objective function value is retained as the final next generation population. Individuals in The selection process can be represented as: ; For all After completing the above operations at each index position, the final updated population of this iteration is obtained. This mechanism effectively guides the evolutionary direction of the population by directly injecting high-quality solution experience obtained from historical searches into the current population, thereby improving the convergence speed and enhancing the algorithm's robustness in escaping local optima.

[0064] S33. Using the Pareto optimization framework, non-dominated sorting and crowding distance calculations are performed on the updated population to update the Pareto optimal solution set, including: S331. Set optimization goals.

[0065] In intelligent traffic management for highway operations, the merits of different strategy combinations need to be comprehensively evaluated from multiple dimensions. This embodiment sets three optimization objectives: traffic efficiency objective... Operational safety objectives and travel time reliability target To balance traffic efficiency, operational safety, and travel time reliability. For any candidate solution Its merits and demerits depend on the objective function vector. The evaluation is conducted, and all three objectives are minimized objectives.

[0066] The multi-objective optimization function is: ; in, The preset weighting coefficients satisfy... .

[0067] Traffic efficiency target The overall traffic capacity of the road network is reflected in the ratio of total traffic volume to total travel time. ; In the formula, cross-section At any moment Predicted traffic flow (vehicles / hour). cross-section At any moment The predicted speed (km / h). The time window length (in hours). cross-section The length of the road segment (km). The total number of cross-sections. To predict the number of time windows.

[0068] Operational security objectives Traffic flow stability is reflected by two dimensions: speed differences between cross sections and flow fluctuations within a cross section. ; In the formula, This is a normalization coefficient used to balance the dimensional differences between velocity variations and flow fluctuations. cross-section The historical average traffic flow (vehicles / hour) reflects the normal traffic flow level of this section. The first penalty is the speed difference between upstream and downstream sections; the greater the difference, the more unstable the traffic flow and the higher the risk of accidents. The second penalty is the fluctuation of the section's traffic flow relative to the historical average level; the greater the fluctuation, the more turbulent the traffic flow and the worse the safety. The smaller the value (i.e., the larger the absolute value of the negative value), the higher the operational safety.

[0069] Travel time reliability target The stability of travel time is reflected in the relative deviation between predicted travel time and free-flow travel time. ; In the formula, cross-section Predicted travel time (in hours) for the specified route. This refers to the travel time (in hours) under free-flow conditions on this road segment. The smaller the value (i.e., the larger the absolute value of the negative value), the more stable and reliable the travel time is, and the more accurately travelers can estimate their arrival time.

[0070] S332. Compare individuals in the unsorted population pairwise according to the optimization objective, perform fast non-dominated sorting, and divide the population into levels; the unsorted population includes the updated population and the previous generation Pareto optimal solution set.

[0071] First, construct the mixed population to be sorted. From the current updated population Pareto optimal solution set retained from the previous generation It was formed by merging, that is Let the total number of individuals be... .

[0072] Then on Perform a fast non-dominated sort on all individuals. For any two individuals... and Compare their objective function vectors. If Not inferior to in any goal And is strictly superior to at least one objective. Then it is called Dominate Specifically, for the minimization problem, Dominate If and only if: and

[0073] Based on this dominance relationship, the fast non-dominated sorting algorithm calculates two quantities for each individual: a dominance count (the number of other individuals that dominate this individual) and a dominated set (the set of other individuals dominated by this individual). The algorithm first identifies all individuals with a dominance count of 0 and assigns them to the first non-dominated front. ; then regarding For each individual in the frontier, iterate through each individual in its dominated set, decrementing its domination count by 1. If the count reaches 0, move it to the next frontier. Repeat this process until all individuals have been assigned to a front. This ultimately yields a series of fronts stratified according to dominance relationships. ,in It is the Pareto optimal frontier, the highest level. And so on.

[0074] S333. Calculate crowding distance for individuals in the same class.

[0075] To measure the distribution density of individuals within the same non-dominated front and to provide a basis for diversity preservation in subsequent selection, the crowding distance for each individual needs to be calculated. For a given front... Suppose it contains Individual. First, for each objective function... Individuals within the frontier are categorized as follows: Sort by value in ascending order to obtain a sorted list of individuals. .

[0076] individual In the Partial congestion distance on each target Calculated according to the following rules: For individuals at both ends of the sorted list (i.e., extreme points on the target), their partial crowding distance is set to infinity:

[0077] For individuals in the middle of the list The crowding distance is partly the normalized difference between the target values ​​of adjacent individuals: ; in: and Individuals The first adjacent individuals after sorting One objective function value; and These are the current mixed populations. The Middle The maximum and minimum values ​​of the objective functions are used for normalization to eliminate the influence of different objective dimensions.

[0078] individual Overall crowding distance The sum of the partial crowding distances across all its targets: ; Crowded distance The larger the value, the lower the solution density around the individual, and the greater the individual's contribution to maintaining population diversity. For boundary individuals, since their crowding distance is set to infinity, they will be preferentially retained in subsequent selections, thus ensuring that boundary points of the Pareto front are not lost.

[0079] Crowding distance essentially measures the "sparseness" of an individual among its neighbors in the target space; by maximizing crowding distance, the algorithm can maintain a uniform distribution of solutions within the same non-dominated level, avoiding excessive concentration of the solution set in a certain local region. Boundary points (extreme solutions) are given infinitely crowded distances to ensure they are always preserved, which is crucial for maintaining the full coverage of the Pareto front.

[0080] S334. Individuals are selected according to the rules of priority by level and priority by distance in the same level of crowding until the number of individuals is consistent with the initial population size, thus forming an updated Pareto optimal solution set.

[0081] To form a new generation of Pareto optimal solution set It is necessary to start from mixed populations Selected from The optimal individual ( (This refers to the initial population size). The selection process follows these rules: individuals with higher non-dominant ranks are given priority, i.e., individuals from the first frontier are selected. Begin by adding them one by one. until a certain frontier The total number of individuals after being added exceeds If we add the cutting edge The total number of individuals is exactly equal to ,but It consists of all the individuals that have already joined. If the frontier is joined... The total number of individuals is greater than Then it is necessary to study the frontier Further screening is then conducted on individuals within the pool. At this point, based on the distance between the groups in the crowd, [the selection process is further refined]. Individuals in the crowd distance Sort in descending order and select the first few. Add individuals such that the total number of individuals is exactly . ,in .

[0082] The rationale for this selection strategy is that it first ensures convergence (by selecting solutions with higher non-dominant rank) and then ensures diversity (by selecting solutions with larger crowding distances within the same rank), thus making the final solution set both close to the Pareto front and evenly distributed within the target space.

[0083] Final selection Each individual constitutes the updated Pareto optimal solution set. This solution set simultaneously achieves convergence (close to the Pareto front) and diversity (uniformly distributed in the target space), providing a reserve of high-quality elite solutions for subsequent iterations.

[0084] S34. Select a preset number of optimal solutions from the updated Pareto optimal solution set and update the historical good solution memory table.

[0085] Updated solution set The number of solutions included and the initial population size Consistent. Each solution in the solution set It possesses two key attributes: its non-dominant frontier level and crowded distance .

[0086] To comprehensively evaluate the quality of each solution, a Pareto comprehensive evaluation value is defined. The calculation formula is as follows: ; in, Solution set The maximum value of all uncongested distances, parameters It is a coefficient used to adjust the weight of diversity contribution, and its value is usually between 0.1 and 0.5.

[0087] Calculate the solution set All solutions After determining the values, sort them in descending order. Then, select the top-ranked values. The solutions constitute a new generation of excellent solution set. .parameter This represents the number of excellent solutions that need to be selected in each generation, and is usually set to the initial population size. 5% to 20%.

[0088] Next, use Update historical excellent solution memory table The update process follows these rules: First, [the system will...] Add all solutions If added The total number of solutions exceeds If so, a replacement operation needs to be performed. During replacement, first... All solutions are sorted according to their multi-objective function values. Sort in ascending order, then keep the first few. The optimal solution for each multi-objective function value is selected. If multiple objective function values ​​are the same at the retention boundary, the most recently added solution is retained to ensure the timeliness of the memory table information.

[0089] Through the above steps, a historical excellent solution memory table is formed. It is dynamically updated after each iteration, always preserving the elite solutions with optimal multi-objective function values ​​and a certain degree of diversity discovered during the algorithm's search process, providing crucial empirical knowledge for guiding subsequent population updates.

[0090] S35. Determine whether to trigger the adaptive reconnaissance mechanism based on the state of the Pareto optimal solution set. Using the optimal solution in the updated historical good solution memory table as a reference point, generate a new reconnaissance bee solution within the neighborhood range determined by the selected chaotic map.

[0091] S351. Define two monitoring indicators to determine the convergence state based on the state of the Pareto optimal solution set: Monitor the update status of the Pareto optimal solution set and calculate the number of consecutive iterations without updates. ; Monitor the diversity index of the Pareto optimal solution set and calculate the average congestion distance. The formula for calculating the diversity of the unset is as follows: ; Based on the comparison of these two indicators with preset thresholds, the convergence status is classified into three categories: premature convergence risk, normal convergence, and good convergence. If the number of consecutive iterations without updates exceeds the threshold And the average crowding distance is less than the threshold. If so, it is judged as a risk of premature convergence; The crowding distance threshold characterizes a severe lack of solution set diversity; this state indicates that the algorithm may get stuck in a local optimum and the population diversity drops sharply.

[0092] If the number of consecutive iterations without updates is within a preset range and the average crowding distance is within the interval If the convergence is within the specified range, it is considered normal convergence. The crowded distance threshold is used to characterize the diversity of the solution set. This state indicates that the algorithm is steadily seeking optimization and the solution set maintains a certain distribution breadth.

[0093] If the number of consecutive iterations without updates is less than a preset threshold and the average crowding distance is greater than the threshold. If the convergence is good, it is considered a good convergence. This state indicates that the algorithm is actively discovering new non-dominated solutions, and the solution set is diverse, indicating that it is in an active period of global exploration.

[0094] S352. Trigger an adaptive detection mechanism based on chaotic mapping according to the convergence state, including: If premature convergence is detected as a risk, a large-scale chaos detection is triggered to force the algorithm out of its current local optimum by setting a larger step size coefficient. ; If convergence is determined to be normal, a medium-range chaos detection is triggered to balance global exploration and local development, with a medium step size coefficient set. ; If the convergence is determined to be good, a small-scale chaos detection is triggered to perform a fine-grained search of the current good region, using a small step size coefficient. .

[0095] S353. Based on the selected chaotic mapping type and search step size, and taking the optimal solution in the historical good solution memory table as the reference point, generate a new solution for the scout bee.

[0096] Deciphering the memory table from a historical perspective The optimal solution for the multi-objective function is selected as the reference point, denoted as . Subsequently, a chaotic sequence is generated using the Logistic chaotic mapping, with the iterative formula as follows: ; in As a control parameter, it is usually taken as At this point, the mapping is in a completely chaotic state, and the generated sequence is in the interval It has good traversal capabilities. Initial value exist The value is randomly generated within the range, avoiding the selection of fixed points (such as 0, 0.25, 0.5, 0.75, 1, etc.). Through iteration, a range of length is generated. chaotic vector ,in The dimension of the solution vector (i.e., the total length of the strategy type selection vector and the intensity parameter vector).

[0097] With reference point Based on, combined with step size coefficient and chaotic vectors Within the neighborhood defined by the upper and lower bounds of the specified decision space, a new solution for generating reconnaissance bees is generated. Its first The formula for generating the dimension parameter is: ; in and The first Upper and lower bounds of the dimension parameter. New solution generated. Boundary constraints need to be applied to ensure that the decision lies within a valid decision space. For the components of the strategy type selection vector, the values ​​should be rounded to the nearest integer. For the components in the intensity parameter vector, a reflection boundary strategy is used for constraint.

[0098] S36. Add the new solution from the scout bee to the population, calculate the multi-objective function value of each solution in the updated Pareto optimal solution set, and record the solution with the best multi-objective function value.

[0099] A new interpretation of reconnaissance bees Join the current population and participate in subsequent evolutionary operations. Then, analyze the updated Pareto optimal solution set. Each solution in Calculate its multi-objective function value After calculating the solution set After calculating the multi-objective function values ​​of all solutions, record the solution with the smallest multi-objective function value, denoted as . : ; This solution represents the optimal combination of strategies that achieves the best trade-off among the three competing objectives of traffic efficiency, operational safety, and travel time reliability in the current iteration. It will be used to determine the algorithm's termination condition and serve as a strong candidate for the final inversion result.

[0100] S37. Determine if the termination condition is met. If it is, proceed to S38; otherwise, return to S32 for the next iteration.

[0101] In this embodiment, the termination condition includes at least one of the following: The algorithm has reached its preset maximum number of iterations. The Pareto optimal solution set remains unchanged within a predetermined number of iterations. The recorded optimal multi-objective function value meets the preset accuracy requirements.

[0102] S38. Output the solution corresponding to the recorded optimal multi-objective function value as the optimal strategy combination scheme.

[0103] S4. Generate control commands based on the optimal strategy combination scheme, and execute the control commands through the roadside issuing device.

[0104] Based on the selected strategy type and its corresponding strength parameter value in the optimal strategy combination scheme, the strength parameter value is converted into the control parameters of the actuator according to a preset mapping relationship. For example, if the optimal strategy combination scheme includes "segmented variable speed limit" and the strength parameter... Then the speed limit value of the roadside variable message sign will be set to 70 km / h; if it includes "ramp signal timing" and the intensity parameter Then set the green light duration of the entrance ramp traffic lights to 25 seconds; if it includes "path guidance recommendation ratio" and the intensity parameter If the system is not in operation, guidance information will be pushed to 30% of the vehicles. Control commands are issued through variable message signs, entrance ramp traffic lights, lane indicator signs, and roadside broadcast units.

[0105] S5. Collect traffic flow feedback data after the control command is executed, and use the feedback data to update the decision parameters of the optimizer online.

[0106] After the control command is executed, the roadside sensing equipment continues to collect real-time traffic flow parameters at each cross-section, compares the actual results with the predicted results, and calculates the deviation of the strategy execution effect. This deviation is used as part of the reward signal to update and improve the parameters of the policy network in the artificial bee colony algorithm or the screening threshold of the historical good solution memory table, so that the optimizer gradually converges to the optimal decision mode for the current traffic scenario.

[0107] On the other hand, embodiments of the present invention also propose a smart highway operation traffic diversion system, such as... Figure 2 As shown, it includes: The data acquisition module is used to acquire real-time traffic flow status data of the target road network and generate short-term traffic flow prediction results based on the real-time traffic flow status data. The strategy library construction module is used to build a multi-dimensional traffic redirection strategy library. The traffic redirection strategy library contains a variety of preset strategy types. Each strategy type is associated with a strength parameter, which is used to quantify the degree of execution or control of the corresponding strategy type. The strategy search module takes real-time traffic flow status data and short-term traffic flow prediction results as input, uses an improved artificial bee colony algorithm, takes the selection of strategy type and the assignment of intensity parameters as joint decision variables, and takes a preset multi-objective optimization function as the objective to dynamically search for the optimal strategy combination scheme. The control execution module is used to generate control commands based on the optimal strategy combination scheme and execute the control commands through the roadside issuing device; The feedback optimization module is used to collect traffic flow feedback data after the control command is executed, and to use the feedback data to update the decision parameters of the optimizer online.

[0108] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0109] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent traffic diversion in highway operation, characterized in that, Includes the following steps: S1. Obtain real-time traffic flow status data of the target road network, and generate short-term traffic flow prediction results based on the real-time traffic flow status data; S2. Construct a multi-dimensional traffic redirection strategy library, which contains a variety of preset strategy types. Each strategy type is associated with an intensity parameter, which is used to quantify the degree of execution or control strength of the corresponding strategy type. S3. Using the real-time traffic flow status data and the short-term traffic flow prediction results as input, an improved artificial bee colony algorithm is adopted, with the selection of strategy type and the assignment of intensity parameters as joint decision variables, and with a preset multi-objective optimization function as the objective, the optimal strategy combination scheme is dynamically searched to obtain the optimal strategy combination scheme. S4. Generate control commands based on the optimal strategy combination scheme, and execute the control commands through the roadside dissemination device; S5. Collect traffic flow feedback data after the control command is executed, and use the feedback data to update the decision parameters of the optimizer online.

2. The intelligent traffic diversion method for highway operation according to claim 1, characterized in that, S1 includes: S11. Real-time traffic flow parameters are collected at preset time intervals by roadside sensing devices deployed at each section of the target road network; S12. Perform spatiotemporal alignment and missing data interpolation on the real-time traffic flow parameters to form a time series data matrix; S13. Input the time series data matrix into a pre-trained multi-scale time series prediction model to generate short-term traffic flow prediction results for multiple future time windows.

3. The intelligent traffic diversion method for highway operation according to claim 2, characterized in that, The multi-scale temporal prediction model includes a graph convolutional layer, an LSTM layer, and a multi-head attention mechanism layer connected in sequence. The graph convolutional layer performs graph convolution operations on the time-series data matrix based on the road network topology, extracts spatial correlation features between each cross section, and outputs a spatial feature sequence; the LSTM layer performs temporal modeling on the spatial feature sequence, extracts the time-series dependency features of each cross section, and generates a temporal feature sequence; the multi-head attention mechanism layer performs weighted fusion of the temporal feature sequence with the proximity features, periodic features, and trend features extracted based on the time-series data matrix to obtain the short-term traffic flow prediction result.

4. The intelligent traffic diversion method for highway operation according to claim 1, characterized in that, The multi-dimensional traffic redirection strategy library includes: Lane-level control strategies include the number of dynamically activated lanes, the number of dynamically closed lanes, the type and location of dedicated lanes, and the mileage and time period for emergency lane borrowing; Speed-level control strategies include segmented variable speed limits, coordinated speed limits, and speed limit change gradients; Right-of-way management strategies include ramp signal timing parameters, toll station entrance control ratio, and service area traffic diversion ratio; Information-level control strategies include route guidance recommendation ratio, destination prompt intensity, and congestion warning release scope.

5. A traffic diversion method for intelligent highway operation according to claim 1, characterized in that, S3 include: S31. Randomly generate an initial population within a preset range of values ​​for joint decision variables, wherein each individual in the initial population corresponds to a combination of a strategy type selection vector and an intensity parameter vector; S32. Update the population using a population update guidance mechanism based on a historical good solution memory table; S33. The Pareto optimization framework is used to perform non-dominated sorting and crowding distance calculation on the updated population, and the Pareto optimal solution set is updated. S34. Select a predetermined number of optimal solutions from the updated Pareto optimal solution set and update the historical excellent solution memory table; S35. Determine whether to trigger the adaptive reconnaissance mechanism based on the state of the Pareto optimal solution set, and generate a new reconnaissance bee solution within the neighborhood range determined by the selected chaotic map, using the optimal solution in the updated historical good solution memory table as a reference point. S36. Add the new solution of the scout bee to the population, calculate the multi-objective function value of each solution in the updated Pareto optimal solution set, and record the solution with the best multi-objective function value; S37. Determine if the termination condition is met. If it is, proceed to S38; otherwise, return to S32 for the next iteration. S38. Output the solution corresponding to the recorded optimal multi-objective function value as the optimal strategy combination scheme.

6. The intelligent traffic diversion method for highway operation according to claim 5, characterized in that, S32 includes: S321. Initialize the historical excellent solution memory table, set the storage length of the memory table, and store individuals in the initial population that meet the preset multi-objective function threshold as initial excellent solutions into the memory table; S322. The hired bee generates a new combination of strategy type selection vector and intensity parameter vector as a new solution based on each individual in the current population and within a preset neighborhood range through random perturbation. S323. Calculate the multi-objective function value of each new solution and the corresponding original population individual, and retain the individual with the better multi-objective function value among each new solution and the corresponding original individual to form the population after preliminary screening; S324. Extract the stored superior solutions from the historical superior solution memory table, calculate the multi-objective function value of the superior solutions, and perform a second greedy selection with the corresponding individuals in the population after the initial screening, retaining the individuals with better multi-objective function values, and completing the population update.

7. The intelligent traffic diversion method for highway operation according to claim 5, characterized in that, S33 includes: S331. Set optimization goals; S332. Individuals in the unsorted population are compared pairwise according to the optimization objective, and a fast non-dominated sort is performed to classify the population levels; the unsorted population includes the updated population and the previous generation Pareto optimal solution set. S333. Calculate crowding distance for individuals within the same level; S334. Individuals are selected according to the rules of priority by level and priority by distance in the same level of crowding until the number of individuals is consistent with the initial population size, thus forming an updated Pareto optimal solution set.

8. A traffic diversion method for intelligent highway operation according to claim 5, characterized in that, In S35: The convergence state is determined based on the state of the Pareto optimal solution set; the convergence state includes premature convergence risk, normal convergence, and good convergence. An adaptive detection mechanism based on chaotic mapping is triggered according to the convergence state, including: If the risk of premature convergence is identified, a large-scale chaos detection will be triggered. If the convergence is determined to be normal, then a medium-range chaos detection is triggered. If the convergence is determined to be good, a small-scale chaos detection is triggered. Based on the selected chaotic mapping type and search step size, a new solution for the scout bee is generated with the optimal solution in the historical good solution memory table as a reference point.

9. A smart highway traffic diversion system, characterized in that, include: The data acquisition module is used to acquire real-time traffic flow status data of the target road network and generate short-term traffic flow prediction results based on the real-time traffic flow status data. The strategy library construction module is used to build a multi-dimensional traffic redirection strategy library. The traffic redirection strategy library contains a variety of preset strategy types. Each strategy type is associated with a strength parameter, which is used to quantify the execution degree or control strength of the corresponding strategy type. The strategy search module is used to take the real-time traffic flow status data and the short-term traffic flow prediction results as input, adopt an improved artificial bee colony algorithm, use the selection of strategy type and the assignment of intensity parameters as joint decision variables, and use a preset multi-objective optimization function as the objective to dynamically search for the optimal strategy combination scheme. The control execution module is used to generate control commands based on the optimal strategy combination scheme and execute the control commands through the roadside publishing device. The feedback optimization module is used to collect traffic flow feedback data after the control command is executed, and to use the feedback data to update the decision parameters of the optimizer online.

Images