A method for actively regulating traffic on a highway

Abstract

This application discloses a method for proactive traffic control on highways, belonging to the fields of intelligent transportation systems and artificial intelligence. The method includes: collecting multi-source traffic data, preprocessing it to extract traffic operation status characteristics and predicting the situation; constructing a refined traffic simulation model, using historical data to calibrate key parameters, simulating different control strategies, and storing the simulation results in a structured form to form a simulation plan library containing traffic scenarios, control strategies, and effect indicators; inputting the current traffic operation status characteristics and situation prediction results into a large language model, and performing reasoning and matching with the plan library: if the scenario matches, the optimal strategy is selected; if it exceeds the plan library, the large language model directly infers and generates macro-control strategy instructions and triggers real-time simulation, supplementing the plan library with new scenarios and strategies; performing feasibility constraint verification on the obtained control strategies, and using the verified strategies as the final decision output. This invention, through the collaborative matching of a large language model and the simulation plan library, realizes dynamic perception of highway traffic operation status, congestion risk prediction, and intelligent generation of control strategies, effectively improving the scientific nature and real-time performance of management decisions.

Description

Technical Field

[0001] This application relates to the fields of intelligent transportation systems and artificial intelligence technology, and in particular to an active traffic control method for highways. Background Technology

[0002] With the continuous growth of highway traffic demand, the expansion of the road network, and a significant increase in vehicle ownership and travel frequency, the highway traffic system has been operating under high load for a long time. Against this backdrop, traffic flow is increasingly characterized by high intensity, strong time-varying nature, and high uncertainty. Traffic flow parameters fluctuate frequently in both time and space, and traffic operation states change rapidly. Traffic congestion has evolved from an occasional, localized phenomenon to a persistent bottleneck restricting highway operational efficiency and service levels. Especially during peak hours when traffic demand is highly concentrated, in areas with complex road network structures containing numerous interchanges or bottleneck sections, and in scenarios affected by traffic accidents, severe weather, or emergencies, traffic conditions are prone to rapid deterioration, and traditional traffic management methods are proving inadequate in dealing with these complex situations.

[0003] For a long time, highway traffic management has mainly relied on human experience and pre-set static control rules, with manual or semi-automatic systems monitoring and intervening in traffic conditions. This type of management can meet basic needs when traffic conditions are relatively stable and flow levels are low, but it is no longer sufficient to support refined, intelligent, and forward-looking traffic control objectives given the increasing intensity of traffic and the rapid evolution of traffic conditions.

[0004] From the perspective of existing technology, highway traffic control technology can be mainly divided into two categories: passive control methods based on threshold determination and auxiliary decision-making methods based on traffic simulation.

[0005] Passive control methods based on threshold judgment typically use single or a few combined indicators such as cross-sectional traffic flow, average speed, and lane occupancy as the basis for judgment. When the monitored data exceeds a pre-set threshold, the system triggers corresponding traffic control measures, such as implementing variable speed limits, entrance flow restrictions, or opening hard shoulders. This type of method has clear implementation logic, low deployment costs, and relatively fast response capabilities, thus it is widely used in actual highway traffic management. However, because its core relies on manually set fixed rules and static thresholds, it is difficult to fully reflect the continuous evolution characteristics of traffic flow and cannot dynamically model and predict trends in complex traffic scenarios. In scenarios with rapidly changing traffic conditions or multiple coupled factors, problems such as delayed activation of control measures, mismatched intensity, or unreasonable scope of action can easily occur, thus affecting the control effect and even exacerbating traffic congestion.

[0006] Another category is decision support methods based on traffic simulation. These methods construct macroscopic, mesoscopic, or microscopic traffic simulation models to simulate and extrapolate traffic operation processes under different traffic control strategies, thereby providing decision-making references for traffic managers. Among these, microscopic traffic simulation models can finely characterize vehicle following behavior, lane-changing decisions, and driver reaction characteristics, exhibiting high accuracy in assessing the impact of different traffic control strategies on traffic efficiency and safety. Therefore, this type of method has certain advantages in traffic planning analysis and the evaluation of control effects.

[0007] However, existing traffic simulation-based decision support methods generally suffer from problems such as complex model parameter calibration and lengthy computation processes. Their application often relies on large amounts of historical data and high computational resources, making it difficult to meet the real-time and rapid response requirements of highway operation and management. Furthermore, simulation results are typically presented in the form of numerical indicators, statistical reports, or visualizations, lacking semantic abstraction and comprehensive judgment capabilities. Manual interpretation and decision-making are still required, leading to a significant disconnect between simulation analysis results and actual traffic control decisions, hindering efficient reuse and automated application.

[0008] The above analysis reveals the following problems and shortcomings in existing highway traffic control technologies: First, the traffic control model is mainly based on passive response, lacking systematic analysis and forward-looking projection capabilities of traffic flow operation trends, making it difficult to implement timely and accurate proactive intervention; Second, the connection between refined traffic simulation analysis results and actual traffic control decisions is not close, and simulation results cannot directly support rapid decision-making and dynamic control, limiting their application effectiveness in actual highway operation and management. Summary of the Invention

[0009] This invention addresses the technical problems existing in the background art by proposing an active traffic control method for highways. Specifically, it involves an active traffic control method for highways based on matching a large language model with a simulation plan library. By constructing an integrated technical framework of "perception-deduction-decision", it realizes dynamic perception of highway traffic operation status, congestion risk prediction, and intelligent generation of control strategies, thereby improving the efficiency of highway traffic operation and the scientific nature and real-time performance of management decisions.

[0010] To solve the technical problem, the technical solution of the present invention is as follows:

[0011] An active traffic control method for highways, the method comprising the following steps:

[0012] S1: Collect multi-source traffic data, extract traffic operation status features after preprocessing, and predict traffic situation based on the features to obtain the situation prediction result;

[0013] S2: Construct a refined traffic simulation model, combine parameter sensitivity analysis and historical trajectory data, perform state-by-state calibration of the micro-key parameters of the simulation model, use the calibrated simulation model to simulate and deduce different control strategies, store the simulation results in a structured form, and construct a simulation plan library containing traffic scenarios, control strategies and effect indicators.

[0014] S3: Input the current traffic operation status characteristics and the situation prediction results into the large language model, and perform reasoning and matching in combination with the simulation plan library: If the current traffic scenario matches a scenario in the plan library, the large language model selects the corresponding optimal control strategy from the plan library; if the current scenario exceeds the coverage of the plan library, the large language model infers and generates macro-level traffic control strategy instructions, outputs them in a standardized structure, and supplements the new scenario and strategy to the plan library;

[0015] S4: Perform feasibility constraint verification on the control strategy obtained in step S3, and take the strategy that passes the verification as the final active traffic control decision output.

[0016] Furthermore, in step S1, the multi-source traffic data includes gantry traffic flow detection data, toll station flow data, roadside monitoring video data, and historical traffic operation data; the traffic operation status characteristics include OD flow, vehicle type ratio, road segment travel time, cross-sectional traffic flow, cross-sectional average driving speed, headway or spacing distribution, acceleration and deceleration distribution, number of lane-changing behaviors, and their changing trends over time.

[0017] Furthermore, in step S1, the extracted traffic operation status features are input into the situation prediction model, outputting the congestion probability within the future time window, and classifying them into different control and early warning levels according to the probability threshold.

[0018] Furthermore, in step S2, the calibration of the key parameters includes: extracting core parameters through sensitivity analysis, extracting benchmark values ​​of vehicle dynamic parameters using micro-trajectory data, and using an evolutionary algorithm to optimize and calibrate driving behavior parameters for free flow and congested flow states respectively.

[0019] Furthermore, in step S2, the different control strategies include: variable speed limit control, entrance flow restriction control, hard shoulder opening control, and combined control of the above control strategies; the simulation results include traffic operation status indicators, congestion characteristic indicators, congestion evolution indicators, and strategy effect indicators. The effect indicators include average speed, traffic flow, occupancy rate, queue length, congestion duration, traffic recovery time, speed improvement rate, queue reduction rate, and delay change rate.

[0020] Furthermore, in step S3, the reasoning and matching process of the large language model includes: guiding the large language model to conduct comparative analysis and causal induction of the effects of different simulation control strategies through a pre-set prompting process, and generating candidate traffic control strategies from the perspective of the current traffic status, future evolution trend and differences in control effects.

[0021] Furthermore, in step S3, if the current traffic scenario exceeds the coverage of the contingency plan database, the macro-level traffic control strategy instruction generated by the large language model specifically includes: the large language model does not rely on the effect indicators of historical contingency plans, but directly deduces and generates instructions containing the mainline dynamic speed limit value and the toll station entrance control method based on the current traffic operation status characteristics and situation prediction results, combined with traffic flow control rules.

[0022] Furthermore, in step S4, the feasibility constraint verification includes: matching verification with existing control rules in the traffic control knowledge base, constraint verification with the capabilities and communication conditions of traffic control equipment, and compliance verification with traffic safety regulations and operational safety requirements; the candidate strategy that is closest to the current scenario among those that pass the verification is determined as the final proactive traffic control decision, and the decision is output in a standardized structure form for the strategy issuance of the traffic management platform or to form a real-time closed-loop linkage with the refined traffic simulation system.

[0023] Furthermore, in step S1, the preprocessing of the multi-source traffic data includes: aligning the multi-source traffic perception data in the time dimension and matching it in the spatial dimension, eliminating the sampling frequency differences and spatial reference differences between different data sources, forming a unified data format, and removing outliers and missing data.

[0024] Furthermore, in step S1, the traffic situation prediction includes: using a situation prediction model to predict the traffic operation evolution trend over a future period of time, and outputting the corresponding continuous congestion probability value and warning level.

[0025] This application has the following advantages:

[0026] A closed-loop architecture encompassing "data perception, simulation analysis, and intelligent decision-making" has been constructed. This breaks through the limitations of traditional control methods that rely on human experience or a single rule base. By perceiving traffic conditions in real time through multi-source data and using refined micro-simulation as the core analysis engine, the system evaluates the effectiveness of control strategies in advance, forming an advanced "analysis-driven decision-making" model that significantly improves the scientific rigor and foresight of the strategies.

[0027] This innovative approach integrates a large language model with a simulation contingency plan library for policy reasoning. By combining structured simulation results with semantic understanding capabilities, the large language model can quickly retrieve and match historically optimal strategies, and can also directly infer and generate macro-control policy instructions for uncovered scenarios and issue them to the underlying implementation. This enables dynamic updates and adaptive expansion of the knowledge base, significantly improving the ability to cope with complex and unpredictable traffic scenarios.

[0028] It possesses a complete feasibility constraint verification and closed-loop feedback optimization mechanism. Candidate strategies must be verified through multiple dimensions such as equipment capabilities and safety standards to ensure that the decisions are executable and implementable. At the same time, the actual control effects are continuously fed back to the simulation layer, driving the optimization of model parameters and the correction of contingency plan effects, forming a virtuous cycle of self-learning and self-evolution. Under long-term operation, the accuracy and reliability of system decision-making are continuously improved.

[0029] In summary, this application has achieved a leap from "passive response" to "proactive simulation and intelligent decision-making" in highway traffic management, effectively improving the efficiency of congestion management and operational safety. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the active traffic control method for highways provided in an embodiment of the present invention;

[0032] Figure 2 A hierarchical architecture diagram of active traffic control methods for highways;

[0033] Figure 3 Build a flowchart for the simulation library;

[0034] Figure 4 This is a flowchart illustrating the process of obtaining decision support for proactive traffic control on highways, as described in an embodiment of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] Example 1:

[0037] like Figure 1 As shown, this application proposes an active traffic control method for highways, including: constructing a refined traffic simulation model based on the highway network structure and multi-source traffic data, calibrating traffic flow state simulation parameters, and predicting congestion risks; generating a structured simulation plan library by simulating the effects of different control strategies; and intelligently matching and generating the optimal control strategy in the plan library by integrating real-time simulation traffic parameters and situation prediction results through large language model reasoning, thereby supporting dynamic control.

[0038] This method first acquires multi-source traffic perception data from the highway operating environment, including gantry traffic flow detection data, toll station flow data, traffic parameters extracted from surveillance video data, and historical traffic operation data. After time alignment and spatial matching, the multi-source data is used to extract a set of core features characterizing the traffic operation status. These traffic operation status features include at least OD flow, vehicle type ratio, road segment travel time, cross-sectional traffic flow, cross-sectional average speed, headway / distance, acceleration / deceleration distribution, lane change frequency, and their trends over time, in order to comprehensively depict the macroscopic state of historical and current highway traffic operation.

[0039] The core of this application lies in introducing a refined traffic simulation model to achieve high-fidelity mapping of real-world traffic scenarios. Regarding simulation parameter calibration, core parameters are extracted through sensitivity analysis, baseline values ​​for vehicle dynamics parameters are extracted using microscopic trajectory data under two flow regimes, and an evolutionary algorithm is employed to decouple and independently optimize driving behavior parameters for both free-flow and congested flow states. This allows the simulation model to reproduce real traffic operation characteristics at the macroscopic level and approximate real vehicle interaction behavior at the microscopic level. The refined traffic simulation model, after parameter configuration, is used to construct traffic operation scenarios that match the current traffic conditions.

[0040] Based on the refined traffic simulation model, simulations are performed using historical traffic perception data. During this process, corresponding control simulation scenarios are constructed for various selectable traffic control methods to obtain traffic operation results under different control strategies. The simulation results include at least four major components: traffic operation status, congestion characteristic indicators, congestion evolution indicators, and strategy effectiveness indicators. These include average speed, flow rate, occupancy rate, queue length, congestion duration, recovery time, speed improvement rate, queue reduction rate, and delay change rate. These results are stored in a structured format, forming a reusable traffic control simulation scenario library.

[0041] Building upon this foundation, this application introduces a situation prediction algorithm that combines traffic operation state characteristics with a large language model to perform semantic reasoning analysis on the simulation results. Specifically, the current traffic operation state characteristics and a traffic control simulation scenario library are input into the large language model in the form of structured parameters. The model is guided by prompts to perform comparative analysis, causal induction, and strategy summarization of the simulation results. If there are situations outside the simulation scenario library, the large model, without relying on historical scenarios, directly infers and generates macro-control strategy instructions containing speed limits and flow restrictions based on traffic flow control rules. This enables the large model to generate a set of candidate traffic control strategies from the perspective of "current traffic state—future evolution trend—differences in control effect."

[0042] Finally, the generated candidate traffic control strategies undergo feasibility constraint verification to output the final proactive traffic control decision. This feasibility constraint verification includes at least a matching verification with existing control rules, equipment capacity constraints, and safety specifications in the traffic management knowledge base. By searching for similarities between the current traffic flow state and historical typical traffic states, the most effective control strategy closest to the current scenario is selected. The final verified control strategy is output in a standardized structure, which can be directly used for strategy distribution on the traffic management platform or to form a real-time closed-loop linkage with a refined traffic simulation system, enabling proactive intervention and continuous optimization of highway traffic operation status.

[0043] Specifically, this embodiment proposes an active traffic control method for highways, including the following steps:

[0044] Step S101: Multi-source traffic data collection and preprocessing;

[0045] In this embodiment, multi-source traffic sensing data is first collected from the highway operating environment. This multi-source traffic sensing data includes at least the following:

[0046] The gantry-type vehicle flow detection equipment collects cross-sectional traffic flow, average speed, and vehicle type identification data.

[0047] Vehicle entry and exit flow data collected at toll stations;

[0048] The data on vehicle trajectory, headway, lane occupancy, and lane-changing behavior are extracted from roadside surveillance videos using video analysis algorithms.

[0049] And traffic operation data stored in the historical traffic operation database.

[0050] The multi-source traffic perception data is aligned in the time dimension and matched in the spatial dimension to eliminate the differences in sampling frequency and spatial reference between different data sources. On this basis, a unified data format is formed, and outliers and missing data are removed to obtain a standardized traffic dataset that can be used for modeling.

[0051] Step S102: Traffic operation status feature extraction and congestion risk prediction;

[0052] Based on preprocessed traffic data, a core feature set is extracted to characterize the traffic operation status of highways. This core feature set includes at least the following:

[0053] OD traffic flow, vehicle type ratio, road segment travel time, cross-sectional traffic flow, cross-sectional average driving speed, headway or spacing distribution, vehicle acceleration and deceleration distribution, number of lane-changing behaviors and their trends over time.

[0054] In this embodiment, the extracted high-order spatiotemporal feature vector is input into the gradient boosting tree-based situation prediction model to predict the traffic operation evolution trend in the future period, and outputs the corresponding continuous congestion probability value and warning level.

[0055] Step S103: Construction and parameter calibration of refined traffic simulation model;

[0056] After obtaining traffic operation characteristics and congestion prediction results, a refined traffic simulation model matching the highway network structure is constructed. This simulation model includes a micro-level traffic simulation model to describe vehicle-level car-following behavior and lane-changing decision-making processes.

[0057] Based on toll station transaction data, gantry detection data, and vehicle behavior data extracted from video, key parameters in the simulation model are calibrated. These parameters include at least:

[0058] OD traffic, vehicle type ratio, expected speed distribution, car-following model parameters, lane-changing model parameters, and vehicle acceleration / deceleration characteristic parameters.

[0059] In this embodiment, core parameters are extracted through sensitivity analysis, physical calibration is performed using microscopic trajectory data under two flow states, and evolutionary algorithms are used to perform dual-state optimization calibration for free flow and congested flow states respectively. This enables the simulation model to reproduce the characteristics of real traffic flow at the macroscopic level and to approximate real vehicle interaction behavior at the microscopic level, thereby constructing a traffic simulation operation scenario that matches the current traffic state.

[0060] Step S104: Simulation and deduction of multi-strategy traffic control and construction of contingency plan library;

[0061] Based on a calibrated, refined traffic simulation model, corresponding simulation scenarios are constructed for various selectable traffic control measures. These traffic control measures include at least:

[0062] Variable speed limit control, entrance flow control, hard shoulder opening control, and a combination of the above control measures.

[0063] Simulations were conducted for different control strategies to obtain traffic operation results under each strategy. The simulation results include at least:

[0064] Traffic operation status indicators, congestion characteristic indicators, congestion evolution indicators, and strategy effectiveness indicators, wherein the indicators include average speed, traffic flow, occupancy rate, queue length, congestion duration, traffic recovery time, speed improvement rate, queue reduction rate, and delay change rate.

[0065] The simulation results are stored in a structured format to form a mapping relationship between traffic operation scenarios, control strategies, and effect indicators, thus constructing a reusable traffic control simulation plan library. For example... Figure 3 As shown, this demonstrates the closed-loop construction logic of the entire process, starting from multi-scenario parametric modeling, through congestion modeling, experimental configuration of control strategies, refined simulation and deduction, and finally realizing contingency plan iteration and library optimization.

[0066] Step S105: Reasoning and matching of regulation strategies based on large language models;

[0067] In this embodiment, the current traffic operation characteristics, congestion prediction results, and structured data from the traffic control simulation plan library are input into the large language model. Through pre-set prompts, the large language model is guided to compare and analyze the effects of different simulation control strategies and to summarize causal relationships.

[0068] If a similar scenario can be matched in the simulation plan library for the current traffic operation scenario, the large language model selects the control strategy that best matches the current traffic state and has the best control effect from the plan library.

[0069] If the current traffic operation scenario exceeds the coverage of the contingency plan library, the large language model does not rely on historical contingency plans, but directly infers and generates macro-level instructions that include the main line dynamic speed limit value and the toll station entrance control method, and triggers real-time simulation and deduction to form a new simulation contingency plan and supplement it to the contingency plan library.

[0070] Through the above reasoning process, the large language model outputs a set of candidate traffic control strategies.

[0071] Step S106: Feasibility verification of the control strategy and decision output;

[0072] The feasibility constraint verification of the candidate traffic control strategies includes at least the following:

[0073] Matching and verifying with existing control rules in the traffic management knowledge base;

[0074] Constraint verification of traffic control equipment capabilities and communication conditions;

[0075] Verification of compliance with traffic safety regulations and operational safety requirements.

[0076] After feasibility verification, and combining the similarity search results between the current traffic operation status and historical typical traffic scenarios, the traffic control strategy that is closest to the current scenario and has feasibility is selected as the final proactive traffic control decision, such as... Figure 4 As shown, this demonstrates the entire process from real-time traffic monitoring, through congestion warning, simulation library retrieval, large model inference, to finally outputting decision results and forming a closed-loop feedback.

[0077] like Figure 2 As shown, this embodiment also provides an active traffic management decision support system for highways. By constructing a full-link architecture of "data perception—simulation inference—intelligent decision-making—closed-loop feedback," it achieves the prediction, assessment, and dynamic control of traffic congestion. The system is based on multi-source traffic data, uses refined micro-simulation as its core inference engine, introduces a large language model for strategy reasoning and matching, and ultimately outputs control decisions with feasibility constraints. The simulation model and contingency plan library are optimized through feedback from actual control effects. The following describes the system's composition, logic, and data flow in a complete manner from four levels.

[0078] I. Data Perception Layer: Multi-source data acquisition and feature extraction (corresponding steps S101 and S102);

[0079] 1. Data acquisition and preprocessing (S101);

[0080] The system collects multi-source heterogeneous traffic perception data from the highway operating environment, including: gantry-type vehicle flow detection equipment: cross-sectional traffic flow, average speed, and vehicle type recognition data; toll stations: vehicle entry and exit flow data (OD information); roadside monitoring video: vehicle trajectories, headway, lane occupancy rate, and lane-changing behavior extracted by video analysis algorithms; and historical traffic operation database: stored historical traffic operation data.

[0081] All raw data enters the preprocessing module, where it is aligned in the time dimension (unified to minute-level time series) and matched in the spatial dimension (eliminating spatial reference differences between different detectors), outliers and missing data are removed, forming a standardized traffic dataset.

[0082] 2. Feature extraction and congestion risk prediction (S102);

[0083] Based on standardized data, a set of core temporal features characterizing traffic flow evolution is extracted, mainly including: multi-order lag features of traffic flow, rate of change of speed, and rolling smooth average speed features. The extracted spatiotemporal feature vectors are input into a machine learning-based situation prediction model, which outputs the probability value of continuous congestion within future time windows. This probability is then dynamically mapped to different control and warning levels (e.g., smooth flow, high alert, light congestion, severe congestion) based on probability thresholds. These prediction results will serve as the basis for subsequent simulations and decision-making.

[0084] Data flow: Preprocessed data → Feature extraction module → Congestion prediction module → Simulation and inference layer (for model calibration and scenario construction) and decision service layer (for state matching).

[0085] II. Simulation and Deduction Layer: Microscopic Simulation Modeling and Strategy Contingency Plan Library Construction (corresponding steps S103 and S104)

[0086] 1. Refined simulation model construction and parameter calibration (S103);

[0087] A microscopic traffic simulation model is built based on the actual road network structure to describe vehicle-level car-following behavior and lane-changing decisions. Highly sensitive parameters with a core impact on road network capacity are extracted through parameter sensitivity analysis. Benchmark values ​​of vehicle dynamics parameters are extracted from microscopic trajectory data under two flow regimes for physical calibration. An evolutionary algorithm is employed to decouple and independently optimize behavioral parameters such as driver reaction time and lane-changing aggression under both free-flow and congested flow conditions. This allows the calibrated simulation model to switch underlying microscopic parameters under different traffic conditions, approximating real vehicle interaction behavior and thus constructing a simulation operating scenario that matches the current traffic conditions.

[0088] 2. Multi-strategy simulation and contingency plan library construction (S104);

[0089] Based on the calibrated simulation model, simulation scenarios were constructed for various optional traffic control methods: variable speed limit control; entrance flow control; hard shoulder opening control; and combined control strategies.

[0090] Each strategy was simulated and analyzed to obtain operational results indicators, including: traffic operation status indicators: average speed, traffic flow, and occupancy rate; congestion characteristic indicators: queue length and congestion duration; congestion evolution indicators: congestion spread speed and traffic recovery time; and strategy effectiveness indicators: speed improvement rate, queue reduction rate, and delay change rate.

[0091] All simulation results are stored in a structured format, forming a mapping relationship between traffic operation scenarios, control strategies, and effect indicators—that is, a traffic control simulation plan library. This plan library supports rapid retrieval of similar historical cases based on scenario characteristics.

[0092] Data flow: Feature data from the data perception layer → simulation parameter calibration → simulation deduction → contingency plan library storage; at the same time, the contingency plan library will serve as a knowledge source for the decision service layer.

[0093] III. Decision Service Layer: Large Language Model Reasoning and Strategy Generation (corresponding to steps S105 and S106).

[0094] 1. Large Language Model Reasoning and Policy Matching (S105);

[0095] The current traffic operation characteristics, congestion prediction results, and structured data from the contingency plan database are input into the large language model. The model is guided by prompting engineering to perform comparative analysis and causal induction: if the current scenario can be matched with a similar scenario in the contingency plan database, the model selects the strategy that best matches the current state and has the optimal control effect; if the current traffic operation scenario is outside the coverage of the contingency plan database, the large language model, without relying on historical contingency plans, directly infers macro-instructions based on traffic flow control rules, including dynamic speed limits on main lines and toll station entrance control methods, and outputs candidate traffic control strategies in standardized JSON format. These strategies are then added to the database after real-time simulation verification.

[0096] 2. Strategy feasibility verification and decision output (S106);

[0097] Candidate strategies enter the feasibility constraint verification module, which verifies the following: Traffic control knowledge base rule matching: whether it conforms to existing control rules (e.g., maximum speed limit, maximum ramp adjustment rate); Equipment capability and communication constraints: whether currently available equipment supports the strategy, and whether communication latency meets real-time requirements; Traffic safety compliance: whether it meets safety requirements (e.g., opening the hard shoulder must ensure the availability of the emergency lane). After verification, combined with the similarity search results of historical typical traffic scenarios (extracted from historical cases), the strategy that is closest to the current scenario and has executable capabilities is selected as the final proactive traffic control decision. This decision will then be issued to the highway control system for execution.

[0098] Data flow: Real-time features from the perception layer + prediction results + contingency plan library → Large-scale model inference → Candidate strategies → Feasibility verification → Final decision output → Execution system. Simultaneously, newly generated simulation contingency plans are fed back to the contingency plan library for knowledge updates.

[0099] IV. Closed-loop feedback and system optimization;

[0100] After the actual control measures are implemented, the system continuously collects traffic operation data and feeds it back to the simulation and inference layer through the data perception layer. This data is used to: verify the control effect and correct the simulation model parameters; evaluate the effectiveness of strategies in the contingency plan library and update the effect indicators; and if a large deviation is found between the actual effect and the simulation expectation, trigger model recalibration or parameter optimization. This closed-loop mechanism enables the system to have self-learning capabilities, continuously improving prediction accuracy, simulation realism, and decision reliability.

[0101] Example 2:

[0102] To further clarify and fully illustrate the technical solution of the present invention, and to enable those skilled in the art to accurately understand and implement it, this embodiment provides specific implementation details and parameter examples for each step based on Embodiment 1.

[0103] 2.1 Specific methods for preprocessing multi-source traffic data;

[0104] In step S101, spatiotemporal alignment and outlier removal are performed on the multi-source traffic sensing data, specifically as follows:

[0105] (1) Time base alignment and outlier removal;

[0106] Let the original micro-transaction log dataset be... Since ETC transaction records are event-triggered data, the core feature vector extracted from a single record is... These represent the entry timestamp, gantry capture timestamp, entry toll station hexadecimal number, gantry hexadecimal number, and vehicle unique identifier, respectively.

[0107] Time alignment: Convert discrete timestamps uploaded from different devices to a standard time coordinate system to eliminate deviations caused by asynchronous device clocks.

[0108] Anomaly removal: If a record shows time reversal (i.e. (1) Missing key timestamps, or the same vehicle At the stabilization time threshold If multiple duplicate streams are generated, they are marked as invalid data segments and removed.

[0109] (2) Spatial topology matching;

[0110] Establish a mapping dictionary of "equipment number - physical section" for the highway network. Define space mapping function Discrete hexadecimal device numbers are mapped to a unified road network spatial reference system:

[0111]

[0112] This allows for the precise separation of through traffic on the main line from traffic merging into the toll station.

[0113] (3) Aggregation of macroscopic spatiotemporal features based on fixed time windows;

[0114] To eliminate the randomness of micro-traffic flow and align with the dynamic boundary control frequency of high-fidelity traffic simulation models (such as SUMO), a fixed time aggregation window step size is set. In this embodiment, the preferred option is... (i.e., 5 minutes) Mapped target monitoring section At this point, perform macro-feature aggregation:

[0115] Cross-sectional equivalent hourly flow rate :

[0116] Extract the unique set of vehicles passing through the cross section within this time window. .

[0117] Convert it to standard hourly flow rate (pcu / h):

[0118]

[0119] Cross-sectional space average velocity :

[0120] Calculate the arithmetic mean of the estimated interval speeds of all valid vehicles in this set:

[0121]

[0122] Through the above steps, the microscopic flow data was successfully reduced in dimensionality and transformed into a macroscopic spatiotemporal feature sequence suitable for large model inference and simulation input.

[0123] 2.2 Traffic state feature extraction and situation evolution prediction;

[0124] In step S102, this embodiment employs a machine learning prediction framework based on high-order spatiotemporal feature engineering and gradient boosting trees (XGBoost) to dynamically predict and classify short-term traffic conditions (e.g., T+10 minutes) and early warning levels. The specific implementation is as follows:

[0125] (1) High-order spatiotemporal sequence feature extraction;

[0126] To capture the spatiotemporal evolution memory effect of traffic flow, a system with a depth of [missing information] was constructed. (This embodiment is preferred) =3) sliding historical data cache. Based on the basic features (current time period) obtained from step S101 aggregation. average speed Mainline traffic Inbound traffic ), further derived to construct higher-order spatiotemporal feature vectors The core feature calculation model includes:

[0127] Multi-hysteresis features: Extracting state fragments from adjacent historical time windows, such as .

[0128] Temporal state change rate: Reflects the deterioration or recovery trend of traffic flow. Taking the speed change rate as an example, its calculation formula is:

[0129]

[0130] Rolling smoothing feature: Eliminates short-term system disturbances and calculates the sliding average velocity over the past three time periods. :

[0131]

[0132] Dynamic road network pressure index: Constructing the interactive coupling characteristics of speed and traffic flow, such as the operational pressure index. .

[0133] (2) Congestion probability prediction based on the XGBoost framework;

[0134] The constructed 26-dimensional spatiotemporal feature vector The data is input into a pre-trained XGBoost traffic situation prediction model. This model is... The ensemble of regression trees, through weighted summation and mapping using the Logistic function, outputs the continuous probability value of congestion occurring in the next time window (T+10 minutes). :

[0135]

[0136] in, Representing the The prediction scores of each decision tree. This model can accurately capture the complex patterns of nonlinear traffic flow evolution.

[0137] (3) Early warning level classification based on dynamic thresholds;

[0138] Based on the optimal model parameters obtained from offline training, the optimal congestion classification threshold for this road segment is extracted. (In this embodiment, the optimal threshold is calibrated) ). Continuous congestion probability. The mapping is to a discrete four-level control and early warning status, and the specific dynamic determination rules are as follows:

[0139] Level 1 (Smooth Traffic): Congestion Probability The road network is operating well and requires no intervention.

[0140] Level 2 (Pay Attention / Minor Congestion): The road network is in the critical period of phase transition, and the system begins to trigger low-level high-frequency monitoring;

[0141] Level 3 (Mild Congestion Warning): Traffic flow declined significantly, triggering the contingency plan database retrieval mechanism;

[0142] Level 4 (Severe Congestion Warning): Faced with the risk of physical queuing disruptions, the system will forcibly activate the Large Language Model (LLM) to conduct proactive traffic control decision-making simulations.

[0143] 2.3 Calibration of key parameters of the simulation model;

[0144] In step S103, sensitivity analysis and multi-source data fusion methods are used to calibrate the key parameters of the microscopic traffic simulation model (such as the SUMO model). The specific calibration process is as follows:

[0145] Parameter sensitivity analysis and dimensionality reduction based on Sobol sequences;

[0146] Nineteen candidate microscopic simulation parameters were selected, and a parameter matrix was generated using Sobol sequence sampling. The deviation between the macroscopic simulation speed and the measured speed was used as the output index. Feature dimensionality reduction was performed by calculating the global sensitivity index to eliminate redundant parameters with weak impact on traffic flow, and nine core parameters with significant impact on road network capacity were extracted as decision variables.

[0147] Dynamic parameter calibration based on trajectory data;

[0148] Among the nine core parameters mentioned above, vehicle dynamics parameters such as maximum acceleration and comfortable deceleration are extracted and directly calibrated using microscopic vehicle trajectory data from video analysis (free flow and congested flow). To eliminate trajectory extraction errors, a sliding time window is set to perform center smoothing on the instantaneous velocity sequence. Let the... The car is The smoothing speed at time step is Its actual acceleration The calculation is as follows:

[0149]

[0150] Based on the acceleration distribution characteristics of effective trajectory samples, specific quantiles are extracted as calibration benchmarks for corresponding dynamic parameters to ensure that the acceleration and deceleration performance of the simulated vehicle conforms to physical reality.

[0151] (3) Optimization of behavioral parameters based on evolutionary algorithms;

[0152] For the remaining seven behavioral characteristic parameters, including reaction time and lane-changing aggression, a differential evolution algorithm is used for optimization. Independent calibration is implemented for the two traffic states: "free flow" and "congested flow." A fitness function is constructed with the objective of minimizing the combined relative error between the simulation output and the measured data. ):

[0153]

[0154] Among them, the mean absolute percentage error of speed The calculation is as follows:

[0155]

[0156] In the formula, and The first Simulation speed and observation speed of each cross section; The mean absolute percentage error of the cross-sectional flow rate; and The weighting coefficients are 0.6 and 0.4 respectively in this embodiment. Through iterative optimization of the algorithm, the optimal parameter combinations for adapting to free flow and congested flow are obtained respectively, forming a dual-state calibration library of micro-parameters, as shown in Table 1.

[0157]

[0158] 2.4 Large Language Model Prompt Word Templates and Parameter Parsing;

[0159] In step S105, the large language model is guided to perform policy reasoning through retrieval enhancement generation and prompting engineering. This embodiment uses the following prompt word template (which can be adjusted according to actual needs):

[0160] System commands;

[0161] You are a top-tier expert system for active traffic control on highways. The following information provides current traffic state parameters, situation prediction results, and historical strategy effectiveness data from the simulation plan library. Based on this information, analyze the expected effects of different control strategies and recommend the optimal strategy.

[0162] If the current scenario is not in the contingency plan database, please deduce an optimal macro-level traffic control strategy based on your knowledge of traffic flow dynamics (including dynamic speed limits on the main road, lane restrictions / closures at tollbooth entrances, etc.). You must strictly adhere to the specified plain JSON format for outputting the decision results, and are prohibited from including any irrelevant Markdown tags or explanatory prefix text.

[0163] Current traffic conditions;

[0164] Cross-sectional average speed: {avg_speed} km / h

[0165] Traffic flow: {volume} veh

[0166] Inbound traffic: {entry_volume} veh

[0167] Speed ​​standard deviation: {speed_std} km / h

[0168] 25th percentile speed: {p25_speed} km / h

[0169] 75th percentile speed: {p75_speed} km / h

[0170] Is the current time period congested?: {is_congested} Dimensionless, value is 0 or 1

[0171] Situation forecast results;

[0172] Output the probability of congestion in the next 10 minutes: %

[0173] Output the specific speed prediction value for the next 10 minutes: km / h

[0174] The system automatically classifies traffic congestion into four warning levels based on probability: Severe Congestion Warning (probability ≥ 70%), Light Congestion Warning (probability 50%-70%), Attention (probability 30%-50%), and Smooth Traffic (probability < 30%).

[0175] A rolling forecast every 10 minutes forms a continuous forecast sequence: km / h

[0176] The effectiveness of candidate strategies in the simulation plan library;

[0177] {pre_library_results} / / This section is filled with historical scenarios similar to the current state and their adjustment effects.

[0178] 1. Compare and analyze the advantages and disadvantages of three strategies: variable speed limit, entrance flow restriction, and hard shoulder opening.

[0179] 2. Recommend one or a combination of strategies and explain your reasoning.

[0180] 3. If the current scenario is outside the scope of the contingency plan library, you must directly return the following strict JSON format structure, and are prohibited from including any other explanatory text:

[0181] {

[0182] "reasoning_chain": "Step-by-step analysis process (detailed explanation of the rationale for adopting this control strategy based on traffic flow theory)",

[0183] "toll_station_control":"Suggested locations of toll stations for traffic control and specific management methods (e.g., number of ETC lanes to be disabled)",

[0184] "speed_limit_recommendation": "Suggested dynamic speed limit for the main line (enter a pure number, such as 0 if there is no speed limit)",

[0185] "expected_speed_increase": "Based on historical plans, the expected increase in vehicle speed (km / h) due to this adjustment",

[0186] "expected_flow_increase": "Based on historical plans, the expected flow rate (pcu / h) increase from this adjustment is..."

[0187] }

[0188] 4. Recommended strategies must comply with safety regulations, with speed limits not lower than 60 km / h and ramp flow restriction rates not lower than 400veh / h.

[0189] The text output by the large language model is parsed using regular expressions to extract candidate strategies or simulation parameters for use in subsequent steps.

[0190] 2.5 The simulation plan library uses a structured storage format;

[0191] In step S104, the simulation results are stored in a structured format in the contingency plan database. This embodiment uses JSON format to store each contingency plan record, as shown in the example below:

[0192] {

[0193] "scenario_id":"PR20241021001",

[0194] "timestamp":"2026-2-2108:00:00",

[0195] "traffic_state":{

[0196] "type":"mild congestion",

[0197] "speed":45,

[0198] "flow": 1850,

[0199] "car_ratio":0.9,

[0200] "truck_ratio": 0.1

[0201] },

[0202] "strategy_applied":{

[0203] "type":"K65+000 Speed ​​Limit + Qujiang Toll Station Entrance Traffic Restriction",

[0204] "parameters":{

[0205] "F1speed_limit":70,

[0206] "F1time_minute":5

[0207] "F2Lane_Closed":1,

[0208] "F2time_minute":5

[0209] }

[0210] },

[0211] "simulation_results":{

[0212] "speed_improvement":15.2,

[0213] "flow improvement": 22.5,

[0214] },

[0215] "effectiveness_score": 0.86

[0216] }

[0217] The contingency plan database supports similarity retrieval based on traffic state feature vectors, using Euclidean distance as the metric.

[0218]

[0219] in The search returns the results with the smallest distance. 1 record.

[0220] 2.6 Example of a feasibility constraint verification rule base;

[0221] In step S106, the feasibility constraint verification is performed based on the rules in the traffic control knowledge base. Example rules for this embodiment are shown in Table 2.

[0222] Table 2

[0223]

[0224] The verification process is as follows: each candidate strategy is matched against the above rules. If any rule is violated, it is marked as "infeasible"; if all rules are passed, the candidate is retained. Finally, the strategy with the best performance in similar historical scenarios (i.e., the strategy that ranks first in similarity retrieval and passes the verification) is selected from multiple feasible strategies as the final output.

[0225] This embodiment provides specific algorithm details, parameter examples, and data structures involved in implementing the method of the present invention. Those skilled in the art, based on the description in this embodiment and combined with conventional programming and simulation tools, can implement the active traffic control method for highways described in this invention. It should be noted that the values, parameters, and rules in the embodiment are merely illustrative examples. In actual applications, they need to be recalibrated and adjusted according to specific road network characteristics and historical data, but these adjustments do not depart from the protection scope of this invention.

[0226] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0227] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for active traffic control on highways, characterized in that, The method includes the following steps: S1: Collect multi-source traffic data, extract traffic operation status features after preprocessing, and predict traffic situation based on the features to obtain the situation prediction result; S2: Construct a refined traffic simulation model, combine parameter sensitivity analysis and historical trajectory data, perform state-by-state calibration of the micro-key parameters of the simulation model, use the calibrated simulation model to simulate and deduce different control strategies, store the simulation results in a structured form, and construct a simulation plan library containing traffic scenarios, control strategies and effect indicators. S3: Input the current traffic operation status characteristics and the situation prediction results into the large language model, and perform reasoning and matching in combination with the simulation plan library: If the current traffic scenario matches a scenario in the plan library, the large language model selects the corresponding optimal control strategy from the plan library; if the current scenario exceeds the coverage of the plan library, the large language model infers and generates macro-level traffic control strategy instructions, outputs them in a standardized structure, and supplements the new scenario and strategy to the plan library; S4: Perform feasibility constraint verification on the control strategy obtained in step S3, and take the strategy that passes the verification as the final active traffic control decision output.

2. The active traffic control method for highways according to claim 1, characterized in that, In step S1, the multi-source traffic data includes gantry traffic flow detection data, toll station flow data, roadside monitoring video data, and historical traffic operation data; the traffic operation status characteristics include OD flow, vehicle type ratio, road segment travel time, cross-sectional traffic flow, cross-sectional average driving speed, headway or spacing distribution, acceleration and deceleration distribution, number of lane changing behaviors and their changing trends over time.

3. The active traffic control method for highways according to claim 2, characterized in that, In step S1, the extracted traffic operation status features are input into the situation prediction model, the congestion probability within the future time window is output, and different control and early warning levels are divided according to the probability threshold.

4. The active traffic control method for highways according to claim 1, characterized in that, In step S2, the calibration of the key parameters includes: extracting core simulation parameters through sensitivity analysis, extracting benchmark values ​​of vehicle dynamic parameters using micro-trajectory data, and using evolutionary algorithms to optimize and calibrate driving behavior parameters for free flow and congested flow states respectively.

5. The active traffic control method for highways according to claim 1, characterized in that, In step S2, the different control strategies include variable speed limit control, entrance flow control, hard shoulder opening control, and combined control of the above control strategies; the simulation results include traffic operation status indicators, congestion characteristic indicators, congestion evolution indicators, and strategy effect indicators. The effect indicators include average speed, traffic flow, occupancy rate, queue length, congestion duration, traffic recovery time, speed improvement rate, queue reduction rate, and delay change rate.

6. The active traffic control method for highways according to claim 1, characterized in that, In step S3, the reasoning and matching process of the large language model includes: guiding the large language model to compare and analyze the effects of different simulation control strategies and summarize the causes and effects through a pre-set prompting process, and generating candidate traffic control strategies from the perspective of the current traffic status, future evolution trend and the difference in control effects.

7. The active traffic control method for highways according to claim 6, characterized in that, In step S3, if the current traffic scenario exceeds the coverage of the contingency plan database, the macro-level traffic control strategy instruction generated by the large language model specifically includes: the large language model does not rely on the effect indicators of historical contingency plans, but directly deduces and generates instructions containing the mainline dynamic speed limit value and the toll station entrance control method based on the current traffic operation status characteristics and situation prediction results, combined with traffic flow control rules.

8. The active traffic control method for highways according to claim 1, characterized in that, In step S4, the feasibility constraint verification includes: matching verification with existing control rules in the traffic control knowledge base, constraint verification with the capabilities and communication conditions of traffic control equipment, and compliance verification with traffic safety regulations and operational safety requirements; the candidate strategy that is closest to the current scenario among those that pass the verification is determined as the final proactive traffic control decision, and the decision is output in a standardized structure form for the policy issuance of the traffic management platform or to form a real-time closed-loop linkage with the refined traffic simulation system.

9. The active traffic control method for highways according to claim 1, characterized in that, In step S1, the preprocessing of the multi-source traffic data includes: aligning the multi-source traffic perception data in the time dimension and matching it in the spatial dimension, eliminating the sampling frequency differences and spatial reference differences between different data sources, forming a unified data format, and removing outliers and missing data.

10. The active traffic control method for highways according to claim 1, characterized in that, In step S1, the traffic situation prediction includes using a situation prediction model to predict the traffic operation evolution trend over a future period of time, and outputting the corresponding continuous congestion probability value and warning level.

Images