A method and system for dynamically opening a hard shoulder of an expressway based on an NFGD model and a simulation platform

Abstract

The application discloses a kind of highway hard shoulder dynamic opening method and system based on NFGD model and simulation platform. Including: acquisition and using dynamic confidence mechanism weighted fusion multi-source real-time sensing data, obtain the traffic state characteristics after fusion;Based on the NFGD model including fuzzy-neural hybrid controller, multi-objective genetic algorithm controller and reinforcement learning controller, make hard shoulder dynamic control decision;With simulation platform real-time interaction, realize control instruction issuing, feedback acquisition, and online strategy evaluation.Compared with the traditional method based on fixed experience threshold, the application realizes adaptive modeling and opening discrimination by fusing fuzzy-neural hybrid controller, multi-objective genetic algorithm controller and reinforcement learning controller, improves the decision accuracy and scene adaptability.Meanwhile, combined with the prediction adjustment feedback step control structure, effectively suppresses the simulation time drift, enhances the timeliness of control strategy, and has high practical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of traffic management and intelligent transportation systems, and in particular to a method and system for dynamically opening a hard shoulder of an expressway based on an NFGD model and a simulation platform. BACKGROUND

[0002] With the continuous growth of traffic flow on expressways, traditional fixed hard shoulder management strategies are difficult to meet dynamic traffic demand. Existing technologies mainly rely on single sensors to collect limited dimensional static data, lack deep fusion and real-time analysis capabilities of multi-source heterogeneous data, resulting in insufficient accuracy and timeliness of traffic state evaluation. At the same time, existing hard shoulder opening strategies are mostly based on artificial rules or static thresholds, and cannot achieve high-precision synchronization with micro-traffic simulation (such as SUMO), lacking a closed-loop feedback mechanism between simulation results and real-time sensor data, limiting the dynamic adaptability of control strategies. At the decision-making level, traditional methods often only consider a single optimization goal (such as traffic efficiency), cannot intelligently coordinate multi-objective conflicts such as efficiency, safety, and disturbance minimization, and lack self-learning capabilities to adapt to complex traffic scenarios (such as bad weather or mixed traffic flow). In addition, the sensing, simulation, and control modules of existing systems lack coordination, and data interaction standards are not unified, making it difficult to achieve closed-loop optimization from state perception to strategy execution, resulting in delayed evaluation of hard shoulder opening effectiveness and the inability to continuously iterate control parameters based on historical data. Therefore, there is an urgent need for a dynamic opening scheme for expressway hard shoulders that integrates multi-source data fusion, high-precision simulation synchronization, and intelligent decision-making to improve road network traffic efficiency and safety management. SUMMARY

[0003] The present application relates to the field of traffic management and intelligent transportation systems, and in particular to a method and system for dynamically opening a hard shoulder of an expressway based on an NFGD model and a simulation platform.

[0004] Technical solution: In order to achieve the above-mentioned purpose of the application, the present application adopts the following technical solutions:

[0005] In a first aspect, the present application provides a method for dynamically opening a hard shoulder of an expressway based on an NFGD model and a simulation platform, comprising the following steps:

[0006] Collecting and weighting multi-source real-time sensor data using a dynamic confidence mechanism to obtain fused traffic state characteristics;

[0007] The hard shoulder dynamic control decision is made based on an NFGD (Neuro-FLC, NSGA-II, DDPG) model including a Neuro-FLC, a NSGA-II and a DDPG, comprising: inputting a fusion state vector into the Neuro-FLC to make a real-time open tendency judgment, outputting an open tendency score, and judging whether to trigger strategy optimization; when the strategy optimization is triggered, using the NSGA-II to collaboratively solve multiple conflict targets, outputting a set of Pareto optimal control strategies, and the control parameters including an open length, a vehicle type restriction, a speed limit value and an opening and closing period; using the perception state, the set of optimal control strategies and historical feedback to build a DDPG to output an action vector based on a deterministic policy gradient method, representing a currently recommended hard shoulder opening strategy combination, wherein the set of optimal control strategies is used as a basis for the action space of the DDPG; and the action instruction is applied to a simulation system or an actual induction system, the system collects current cycle feedback indicators, iteratively optimizes the strategy network and the value network, and forms a closed-loop adaptive control strategy.

[0008] The simulation platform is interacted with in real time to realize control instruction issuing, feedback collection and online strategy evaluation.

[0009] Further, the fused traffic state features include vehicle average speed, queue length, compression degree, large truck occupancy rate, traffic density and weather and accident risk information. The standardized weight w k (t) is calculated by using a dynamic confidence mechanism The standardized weight satisfies a normalization constraint; wherein K represents the number of sensors, represents the observation value of the kth sensor on the variable i, Var represents the variance of the observation value sequence within a certain time window, and σ 2 represents a normalization factor or a reference variance constant, used to adjust the tolerance degree to the fluctuation amplitude.

[0010] Further, in the Neuro-FLC, expert rules are built in the fuzzy reasoning layer and the fuzzification values of the input variables are calculated by using a Gaussian membership function; the neural network optimization layer performs online training on the fuzzy rule parameters through a three-layer feedforward network, and the training data comes from a benefit comparison set of opening strategies in historical simulation; when the output open tendency score is greater than a score threshold, or the output open tendency score is greater than the score threshold and the shoulder width and / or the traffic risk meet the set conditions, an opening instruction is triggered to perform control strategy optimization.

[0011] Further, in the collaborative solving of multiple conflict objectives using the multi-objective genetic algorithm, the NSGA-II non-dominated sorting genetic algorithm is used to generate a multi-objective control strategy, and the optimization objectives include maximizing average speed, minimizing lane-changing disturbance, and minimizing lane-changing disturbance; the algorithm uses mixed integer real coding to represent control parameters, and the chromosome structure includes the length of the open hard shoulder section, the type of vehicle allowed to pass, the speed limit value of the corresponding section, the open start time and the closing time; the algorithm first generates a Pareto optimal strategy set, and then selects the comprehensive optimal solution through the TOPSIS method as the reinforcement starting point of the reinforcement learning controller.

[0012] Further, the state space of the reinforcement learning controller includes the average speed of each lane, the queue length, the traffic density, the compression index, the truck occupancy rate, and the opening tendency score of the fuzzy-neural hybrid controller output, the action space is constructed based on the Pareto solution set output by the genetic algorithm, the action includes the length of the open hard shoulder section, the type of vehicle allowed to pass, the speed limit value of the corresponding section, the open start time and the closing time, and the reward function considers the average vehicle speed, the lane-changing disturbance intensity, and the accident risk factor estimated by simulation.

[0013] Further, different expert rules are set for different scenarios of main line hard shoulder opening, outflow ramp hard shoulder opening, and inflow ramp hard shoulder opening, and different models are trained.

[0014] Further, the simulation platform dynamically modifies simulation parameters by expanding lane attribute tags and speed limit configuration, uses a multi-thread and dynamic step adjustment mechanism to ensure that simulation time and actual time are synchronized, and controls the travel path of the simulation vehicle through OD point setting.

[0015] In a second aspect, the application provides a highway hard shoulder dynamic opening system based on an NFGD model and a simulation platform, which is used to implement the highway hard shoulder dynamic opening method based on the NFGD model and the simulation platform, and includes:

[0016] A data transmission and acquisition module is configured to acquire and fuse multiple-source real-time sensing data by using a dynamic confidence mechanism for weighted fusion, to obtain fused traffic state features.

[0017] The system control module is used for making dynamic control decisions of the hard shoulder based on the NFGD model including a fuzzy-neural hybrid controller, a multi-objective genetic algorithm controller and a reinforcement learning controller, and includes the following steps: inputting a fusion state vector into the fuzzy-neural hybrid controller to make a real-time open tendency judgment, outputting an open tendency score, and judging whether to trigger strategy optimization; when the strategy optimization is triggered, using the multi-objective genetic algorithm to collaboratively solve multiple conflict targets, outputting a set of Pareto optimal control strategies, and the control parameters including an open length, a vehicle type restriction, a speed limit value and an opening and closing time period; using the perception state, the set of optimal control strategies and historical feedback to build a reinforcement learning controller to output an action vector based on a deterministic policy gradient method, representing a current recommended hard shoulder opening strategy combination, wherein the set of optimal control strategies is used as a basis for the action space of the reinforcement learning controller; and the action instruction is used for a simulation system or an actual induction system, the system collects current cycle feedback indicators, iteratively optimizes a strategy network and a value network, and forms a closed-loop adaptive control strategy.

[0018] The simulation module is used for real-time interaction with a simulation platform, realizes control instruction issuing, feedback collection, and online strategy evaluation and response.

[0019] In a third aspect, the present application provides a computer system, including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the computer program is executed by the processor to implement the steps of the method.

[0020] In a fourth aspect, the present application provides a computer program product, including a computer program, wherein the computer program is executed by a processor to implement the steps of the method.

[0021] Advantages: The method and system for dynamic opening of a hard shoulder of an expressway based on an NFGD model and a simulation platform have the following technical advantages:

[0022] 1. Dynamic confidence mechanism improves fusion accuracy: By constructing a dynamic confidence weight model, the weighting coefficients are adaptively adjusted according to the historical stability and real-time reliability of various sensors, high-robustness fusion of traffic state features is realized, and the perception accuracy is significantly enhanced.

[0023] 2. Multi-model cooperation improves decision intelligence: An NFGD model is constructed, a fuzzy-neural hybrid controller, a NSGA-II multi-objective optimization controller and a DDPG reinforcement learning controller are integrated, real-time open tendency judgment, strategy generation and closed-loop optimization are realized respectively, an intelligent control chain is formed, and the adaptability and long-term stability of the control strategy are effectively improved.

[0024] 3. Simulation real-time linkage security response timeliness: Through the simulation platform to realize the online evaluation and real-time feedback of control strategy, combined with multi-thread and dynamic step regulation mechanism, to ensure the synchronization of simulation time and actual time, and improve the immediacy and verifiability of strategy response. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A flow chart of a highway hard shoulder dynamic opening method based on an NFGD model and a simulation platform.

[0026] Figure 2 A flow chart of a simulation module.

[0027] Figure 3 A schematic diagram of an integrated sensing system architecture.

[0028] Figure 4 A schematic diagram of the connection between network reliability and simulation.

[0029] Figure 5 A schematic diagram of a simulation road network of a part section of Jing-Shi highway.

[0030] Figure 6 A flow chart of multi-thread simulation operation.

[0031] Figure 7 A graph of the relationship between real time and simulation time in (a) single-thread and (b) multi-thread simulation.

[0032] Figure 8 A schematic diagram of a simulation trip road section unit separation.

[0033] Figure 9 A schematic diagram of a hard shoulder management strategy in an integrated sensing system.

[0034] Figure 10 An example diagram of a two-way four-lane main road simulation scene.

[0035] Figure 11 An example diagram of a two-way four-lane out-ramp simulation scene.

[0036] Figure 12 An example diagram of a two-way four-lane in-ramp simulation scene. DETAILED DESCRIPTION

[0037] The technical solutions and technical effects of the present application will be further described below in combination with the drawings and specific embodiments.

[0038] The embodiment of the application discloses a highway hard shoulder dynamic opening method based on an NFGD model and a simulation platform, which comprises the following steps: collecting and weighting and fusing multi-source real-time sensing data by using a dynamic confidence mechanism to obtain fused traffic state characteristics; performing hard shoulder dynamic control decision based on an NFGD model comprising a fuzzy-neural hybrid controller, a multi-objective genetic algorithm controller and a reinforcement learning controller; and interacting with the simulation platform in real time to realize control instruction issuing, feedback collection and online strategy evaluation. The process of performing hard shoulder dynamic control decision based on the NFGD model comprises the following steps: inputting the fused state vector into the fuzzy-neural hybrid controller to perform real-time opening tendency judgment, outputting an opening tendency score, and judging whether a strategy optimization is triggered; when the strategy optimization is triggered, using the multi-objective genetic algorithm to collaboratively solve multiple conflict targets, outputting a set of Pareto optimal control strategies, and the control parameters comprising an opening length, a vehicle type restriction, a speed limit value and an opening and closing time period; using the perception state, the set of optimal control strategies and the historical feedback to construct a reinforcement learning controller to output an action vector based on a deterministic policy gradient method, which represents a currently recommended hard shoulder opening strategy combination, wherein the set of optimal control strategies serves as a basis for the action space of the reinforcement learning controller; and the action instruction is applied to a simulation system or an actual induction system, the system collects current period feedback indicators, iteratively optimizes the strategy network and the value network, and forms a closed-loop adaptive control strategy.

[0039] Correspondingly, a highway hard shoulder dynamic opening system based on the NFGD model and the simulation platform for realizing the above method mainly comprises the following core modules: a data transmission and collection module, which is used for collecting and weighting and fusing multi-source real-time sensing data by using a dynamic confidence mechanism to obtain fused traffic state characteristics; a system control module, which is used for performing hard shoulder dynamic control decision based on an NFGD model comprising a fuzzy-neural hybrid controller, a multi-objective genetic algorithm controller and a reinforcement learning controller; and a simulation module, which is used for interacting with the simulation platform in real time to realize control instruction issuing, feedback collection and online strategy evaluation.

[0040] In Figure 1 , multi-source data is collected by sensors, transmitted, communicated and analyzed, and system control and simulation verification are performed to determine whether to perform hard shoulder opening and generate an opening strategy.

[0041] In Figure 2 , after a static road network is constructed and a strategy label is extended by the simulation platform, the Neuro-FLC and the NSGA-II intelligent decision unit are accessed after strategy synchronization and time coordination are realized, the DDPG finally realizes closed-loop optimization, and the TRACI executes the control strategy.

[0042] The following describes the details of the highway hard shoulder dynamic opening method and system based on the NFGD model and the simulation platform in the embodiment of the application.

[0043] 1. Architecture of the road network sensing system

[0044] The embodiment determines the appropriate data transmission communication mode considering the characteristics of the sensors connected to the system and the characteristics of the perceived targets. The data processing process after the data transmission to the system includes data preprocessing, data persistence, data mining, structured data storage, and effective information output. According to the data processing flow, the sensing system module integrated with SUMO simulation is as shown in Figure 3 .

[0045] (1) Multi-source data transmission and collection module

[0046] Contains hardware sensor groups (coil detectors, video detectors, microwave radars) and software sensor groups (vehicle-mounted OBUs, web crawlers, weather APIs), and constructs a three-dimensional standardized classification framework of time resolution, spatial distribution characteristics, and data structure;

[0047] (2) Communication and protocol module

[0048] Communication is a basic element of system construction, enabling various parts of the sensing system to cooperate with each other. Further traffic management relies heavily on various sensors to collect sensing information. Data preprocessing is a key part of the communication and protocol module, including data decoding according to the protocol, data source verification, data quality verification, and effective data structure formation.

[0049] (3) Database module

[0050] The database module promises an important function, namely, to persist various types of data and query data according to standards. Data persistence is an important part of the system, which means storing data in the appropriate location and providing data support for further data queries. The data that needs to be saved includes three types: structured sensing data, configuration data, and static large content files.

[0051] (4) Client interaction module

[0052] The client interaction module is used to complete the interaction between the sensing system and the client, including processing the underlying data according to the business layer requirements, extracting effective information from the raw data and transmitting the information to the business layer. The client interaction module also converts the client operations into actual system tasks, one of which is data mining according to the requirements. According to the purpose of data mining, the data mining means in the sensing system includes real-time data analysis, timing data analysis and historical data analysis. Real-time task is to process real-time data immediately, and the processing target includes vehicle motion state data and illegal snapshot. Timing data analysis is used to analyze the traffic data collected in a certain time period, and the application scenarios include congestion state analysis of specific areas and index calculation. Historical data analysis is to analyze long-term historical data stored in the database, such as traffic demand distribution on weekdays and holidays. The implementation of data mining takes into account the actual purpose of system construction.

[0053] (5) Simulation module

[0054] The construction of the simulation module depends on the sensing system collected by the actual sensor, but also has independent characteristics. The simulation module can operate the simulation process according to the business requirements, including generating simulation data and performing control simulation operations such as traffic demand changes, signal timing changes, vehicle insertion and deletion. The addition of simulation makes the sensing system more flexible, efficient, highly visualized and easy to operate.

[0055] (6) System control module

[0056] The system control module serves as the core decision and control center of the embodiment, and undertakes the tasks of generating, triggering and scheduling the hard shoulder opening control strategy. Based on the fused traffic state feature input, the module calls the algorithm controller to make real-time opening tendency judgment, generates a multi-objective control scheme in combination with the strategy optimization module, and realizes the closed-loop iteration and update of the strategy through the reinforcement learning controller. The system control module simultaneously coordinates and communicates with the simulation module in both directions, realizes the issuance of control instructions, feedback index collection and real-time evaluation of the strategy, and ensures the timeliness, adaptability and stability of the opening strategy.

[0057] (7) Interface display and operation module

[0058] The interface display and operation module is used for visual display of the system running state and the entrance of manual intervention control, and supports access by multiple terminal devices. Its functions include: dynamically displaying the fused traffic state indicators (such as average speed, queue length, etc.), the execution state of the opening control strategy, the feedback effect of the simulation, etc. core running information, and providing user authorized login, strategy parameter setting, system mode switching (such as automatic / semi-automatic / manual intervention) and other interactive functions. The module interface supports graphical real-time large screen display, which is convenient for management personnel to track and intervene in the hard shoulder management and control situation in the expressway operation monitoring center.

[0059] 2. Freeway traffic state modeling and multi-source data fusion

[0060] (1) Data sources and type classification

[0061] The multi-source traffic data collection system involved in this embodiment includes data sources such as coil detectors, video detectors, microwave radars, OBU vehicle-mounted units, web crawlers, API interfaces, and social platforms. The coil detectors collect high-frequency point structured traffic volume, speed, and occupancy data, the video detectors obtain high-frequency / region-level semi-structured queue length, vehicle ID, and trajectory data, the microwave radars provide high-frequency point segment composite structured speed vector and spatial distance data, the OBU vehicle-mounted devices record high-frequency region-level structured GPS trajectory and lane-changing behavior data, the web crawler collects low-frequency region-level unstructured holiday prediction and sudden event public opinion data, the API interface obtains low-frequency region-level structured weather and sudden weather data, and the social platform mines low-frequency region-level unstructured congestion perception hotword data. The system classifies multi-source heterogeneous traffic data in three dimensions of time resolution (high frequency / low frequency), spatial distribution characteristics (point / segment / region), and data structure (structured / semi-structured / unstructured) to establish a unified metadata framework for data fusion processing.

[0062] Table 1 Multi-source data types

[0063] Data source Data type Temporal resolution Spatial distribution characteristics Data structure Loop detector Flow, speed, occupancy High frequency Point Structured Video detector Queue length, vehicle ID, trajectory High frequency Segment / region Semi-structured Microwave radar Speed vector, spatial distance High frequency Point / segment Structured OBU (On-Board Unit) GPS trajectory, lane-changing behavior High frequency Region Structured Web crawler Holiday prediction, public opinion on emergencies Low frequency Region Unstructured API interface Weather data, sudden weather (fog / rain) Low frequency Region Structured Social platform Congestion awareness hot words Low frequency Region Unstructured

[0064] (2) Core modeling method

[0065] The following main state variables are extracted in this embodiment:

[0066] Average speed (per lane):

[0067]

[0068] Wherein: is the average speed of the nth lane segment at time t (unit: km / h); is the total number of detected vehicles on the nth lane segment at time t; is the instantaneous speed of the jth vehicle on the lane segment at time t.

[0069] Queue length (per direction): calculated by image processing or historical traffic backlog model:

[0070]

[0071] Wherein: is the queue length of the nth lane segment at time t (unit: m); Length of the detection section corresponding to the nth lane segment (unit: m) ; Traffic density of the nth lane segment at time t (unit: veh / km / lane) ; cr Critical density, which refers to the critical value of traffic state transition from free flow to congestion state; ρ jam Congestion density, which refers to the maximum density when the road is completely saturated.

[0072] Compression index:

[0073]

[0074] Wherein: Compression index of the nth lane segment, indicating the degree of traffic congestion; Traffic density of the nth lane segment; Average speed of the nth lane segment.

[0075] Large truck occupancy rate:

[0076]

[0077] Wherein: Large truck occupancy rate in the mth area; N tru Number of large trucks in the mth area; N total Total number of vehicles (including small cars, trucks, etc.) in the mth area.

[0078] Traffic flow density:

[0079]

[0080] Wherein: Traffic density of the nth lane segment; N total Total number of vehicles (including small cars, trucks, etc.) in the nth lane segment; Length of the detection section corresponding to the nth lane segment (unit: m).

[0081] (3) Weighted fusion model

[0082] The observation of the same variable by multiple sensors is recorded as:

[0083]

[0084] Wherein: Fusion result of the ith state variable; Observation value of the ith state variable by the kth type of sensor; K is the total number of sensor types participating in the observation of the variable; The fusion output calculation method is: Wherein, w k(t) is the fusion weight of the kth sensor at time t, satisfying: ∑w k (t) = 1, w k (t) ∈ [0, 1].

[0085] (4) Dynamic confidence mechanism

[0086] Each type of sensor is given a dynamic confidence, calculated as:

[0087]

[0088] Where: λ k (t) is the dynamic confidence value of the kth sensor at time t; is the kth sensor's observation of variable x i The variance of the observation value within a certain time window; σ 2 The empirical standardized variance, used to adjust the sensitivity of confidence to fluctuations, which can be set as the historical average variance.

[0089] The weight is related to the normalization of confidence:

[0090]

[0091] (5) Output interface and format standardization

[0092] The final fused state feature output is:

[0093]

[0094] Where: l n is the nth monitoring lane segment number, used for speed, density, compression, etc. variable index; s m is the mth region number, used for statistical quantities such as large truck proportion; z is the global number, used for weather, accident, etc. macro influence factors; is the set of lane segments monitored at the current time; is the set of all regions involved. This state vector will be uniformly structured and packaged into JSON format, transmitted to the simulation and control platform through TCP / UDP / WebSocket communication interface.

[0095] 3、Actual road network sensor system and simulation integration

[0096] (1) Construction of simulation environment based on SUMO

[0097] Simulation is a method of simulating actual road networks and dynamic traffic flow. Compared with traditional mathematical traffic analysis methods, simulation has high integration in vehicle following models, lane changing models and signal control models. Simulation can consider various factors of road network and intuitively display dynamic traffic state.

[0098] Considering the demand for traffic flow micro-perception in the sensing system, the embodiment selects the SUMO simulation tool for road network simulation. SUMO has a variety of important functions: controlling simulation speed, dynamically allocating traffic flow, and outputting various parameters.

[0099] (2) Synchronization of actual traffic state and simulation

[0100] The integrated SUMO simulation sensing system can intuitively display the micro-target motion state and make up for the missing part of the road network. The key to integration is to synchronize the parameters of the actual road network and the simulation. The synchronization stage includes static road network structure synchronization, processing time synchronization, control strategy synchronization, sensing information synchronization, and induction measures synchronization. The specific data flow and content of each stage are shown in Figure 4 .

[0101] The static road network synchronization stage is the main part of building an integrated sensing system. The main function of this stage is to complete the simulation modeling of the static road network by collecting static road network data. There are many map sources to describe the map structure, and SUMO simulation uses xml format map data sources. According to actual needs, three kinds of map sources are provided: Open Street Map (OSM), map editing GUI, and high-precision point data.

[0102] The embodiment uses xml format map data in SUMO simulation to simulate the road network. The static information of the road network map includes road segment location, intersection area location, road channelization, and connection relationship between different elements.

[0103] Processing time synchronization is the dynamic basis for the integration of SUMO simulation and actual road network sensing system. In the process of mapping the actual sensing state to the simulation, it is necessary to keep the consistency of the actual time and the simulation time, so this embodiment proposes a method based on multi-threading and dynamic step adjustment to optimize the simulation processing time. The characteristics of SUMO simulation are that the simulation is divided into discrete processes in simulation steps, and the simulation is advanced by controlling the simulation step. However, when the tasks executed in each step are different, different steps need different execution times, resulting in a large deviation between the actual simulation time and the actual time. The multi-threading and dynamic simulation step control method takes the simulation step execution content and simulation step execution time as the control object, and realizes the general correspondence between the simulation time and the actual time.

[0104] The control strategy synchronization stage refers to mapping the actual control strategy to the simulation, which extracts the effective information of the control strategy in the actual road network and executes a similar control strategy in the simulation. Synchronization makes the driving environment in the simulation similar to the actual road network. The highway network control strategy mainly includes vehicle speed limit, road traffic sign setting, and vehicle type limit.

[0105] Sensors information synchronization is the further progress of driving environment simulation. Traffic operation state in road network has obvious spatio-temporal distribution difference. The main function of sensors information synchronization step is to map the traffic operation state distribution of actual road network to simulation system. According to the characteristics of sensors, traffic flow information is converted to vehicle motion information in simulation. Sensors are regarded as the generation point of traffic demand, and the position of downstream sensors is regarded as the destination of traffic demand, and OD pairs are created in simulation.

[0106] Induction measures synchronization is the further process of control strategy synchronization, and the main purpose of this stage is to verify the effect of control strategy by changing the current control strategy (such as hard shoulder opening and closing, vehicle type restriction and traffic demand change) after the simulation is completed. By implementing induction measures in simulation early, comprehensive data can be obtained to support traffic management.

[0107] 4. Neuro-FLC determines the hard shoulder opening condition

[0108] In the highway traffic induction system, how to determine whether the hard shoulder should be opened in real time to alleviate congestion is a core decision-making problem. Traditional methods rely on static thresholds (such as vehicle speed below a certain value or large truck ratio exceeding threshold) for judgment, which lack adaptability and are difficult to adapt to complex dynamic scenarios. In addition, the default step control mechanism of simulation platforms such as SUMO cannot guarantee high-precision synchronization of simulation time and real time, affecting the timeliness of control strategies.

[0109] (1) Input variables

[0110] The Neuro-FLC module takes the state characteristics of multi-source data fusion as input, including average speed, queue length, compression index, large truck occupancy rate, traffic flow density, weather and accident.

[0111] (2) Fuzzy logic reasoning

[0112] Fuzzy controller rule setting example:

[0113] IF Δv is High AND ΔQ is High THEN Open tendency = Strong

[0114] IF Δv is Medium AND ΔQ is Increasing THEN Open tendency = Moderate

[0115] Fuzzy output score is denoted as μ open ∈ [0, 1]

[0116] (3) Neural network self-learning

[0117] Neuro-FLC constructs a three-layer feedforward neural network (input layer-hidden layer-output layer) to train and update the parameters of fuzzy membership function. The training data comes from the historical simulation cases in the context of "open strategy is better than non-open". The comprehensive score, road condition and risk index are integrated to output the dynamic threshold score, and the final open logic is:

[0118] Openifμ open >θ(t)∧W≥3.0m∧R risk <ε

[0119] Where: θ(t) is the dynamic score threshold; W is the width of the hard shoulder; R risk The traffic risk value can be predicted by any existing traffic risk prediction model, and ε is the risk threshold. In some other embodiments, the traffic risk value can also be ignored.

[0120] 5. Multi-objective optimization strategy generation module (NSGA-II)

[0121] This embodiment proposes an open hard shoulder management strategy, which uses NSGA-II (non-dominated sorting genetic algorithm) as the core optimization engine to solve multiple conflicting objectives collaboratively, and outputs a set of Pareto optimal control strategies. The strategy set will serve as the basis for the action space of the reinforcement learning controller, providing stable strategy candidates for subsequent closed-loop self-learning of the system.

[0122] This embodiment sets three key performance indicators as optimization objectives

[0123] (1) Maximize average travel speed:

[0124]

[0125] Where: V avg is the average travel speed of all vehicles in the observation area (unit: km / h); is the number of all road segments in the observation area; is the average speed of the nth observation road segment.

[0126] (2) Minimize congestion delay cost:

[0127]

[0128] Where: C delay is the congestion delay cost, representing the time loss of vehicles in different time periods due to non-free flow state; T is the total number of time steps for optimization evaluation (e.g. in minutes); is the average travel time of the mth region at j time steps; is the baseline travel time of the region under free flow conditions.

[0129] (3) Lane-changing disturbance intensity minimization:

[0130]

[0131] Where: F change is the average lane-changing disturbance intensity, representing the frequency of unit vehicle lane-changing behavior within a certain time; T is the total number of time steps corresponding to the observation duration; the total number of vehicles in the lane segment at time t; is whether lane-changing behavior occurs in the lane segment at time t.

[0132] The final optimization objective function is defined as: max{V avg -C delay -F change}

[0133] For the convenience of genetic algorithm processing, the control strategy parameters are encoded as decision variable vectors in this embodiment:

[0134] X = [L, C v , V limit , t open , t close ]

[0135] Where: L is the length of the open hard shoulder (unit: meters); C v is the type of vehicle allowed to pass (enumerated values: 0-small passenger car, 1-truck, 2-mixed) ; V limit is the corresponding section speed limit value (unit: km / h) ; t open is the opening start time (seconds) ; t close is the closing time (seconds).

[0136] The above parameters are represented by mixed integer real coding, which adapts to the NSGA-II chromosome structure.

[0137] (4) Constraint condition definition

[0138] The width of the hard shoulder has an important impact on driving behavior. The wider the width of the hard shoulder, the faster the speed, and vice versa. For details, see Table 2.

[0139] Table 2 Influence of hard shoulder width on driving speed

[0140] Hard shoulder width Passenger car speed Truck speed 3.0m 100 / 3.5m 100 / 3.75m 120 80 4m 120 80 4.25m 120 80

[0141] This embodiment further analyzes the types of vehicles supported by hard shoulders of different widths, including not meeting vehicle traffic, only small passenger car traffic, and mixed passenger and truck traffic, and sets vehicle speed limits according to different hard shoulder widths.

[0142] Table 3 Hard shoulder speed limit and vehicle restriction

[0143] Design speed Hard shoulder width Passing vehicle Speed limit (literature) Speed limit (simulation) 80 km / h ＜3.0 Prohibited / / 100-120 km / h 3.0 Small passenger car 80 km / h 80 km / h 120 km / h 3.5 Small passenger car 110 km / h 110 km / h 120 km / h 3.75 Mixed passenger and truck 120 km / h 120 km / h 120 km / h 4.0 Mixed passenger and truck 120 km / h 120 km / h 120 km / h 4.25 Mixed passenger and truck 120 km / h 120 km / h

[0144] To ensure the engineering feasibility of the generated strategy, the following constraints are set in this embodiment:

[0145] The hard shoulder can only be opened when its width is greater than or equal to 3.0m; the speed limit parameter should meet V limit ∈[80,120]; the opening and closing timing must meet t open ≤t close ; the open vehicle type should match the width condition (e.g., only allowing small cars).

[0146] (5) Brief description of NSGA-II algorithm flow

[0147] This embodiment aims to solve the problem of dynamic opening control of hard shoulder on expressway. The NSGA-II non-dominated sorting genetic algorithm is introduced into the control strategy generation module, combined with traffic state perception and simulation evaluation mechanism, to realize multi-objective optimization control. The specific steps of the algorithm are as follows:

[0148] a. Individual definition and population initialization: according to the control requirements of hard shoulder, each genetic individual is coded as a five-dimensional vector, representing a set of candidate control parameter combinations: X i =[L,C v ,V limit ,t open ,t close ].

[0149] b. Objective function evaluation: call the SUMO simulation platform to simulate the traffic operation results of each candidate scheme, calculate three optimization indicators: {V avg ,-C delay ,-F change}, which is based on the initial traffic state input provided by the perception system, to ensure that the control strategy is adapted to the current road conditions.

[0150] c. Non-dominated sorting and crowding distance calculation: sort the population based on the Pareto dominance relationship to form several front levels, and maintain the diversity of the population through the crowding distance index.

[0151] d. Evolution operation and new population generation: select excellent individuals by using the elite reservation strategy, generate new individuals through tournament selection, simulated binary crossover (SBX) and polynomial mutation, and continuously optimize the quality of solutions.

[0152] e. Convergence judgment and solution output: When the maximum iteration number is reached or the convergence threshold is met, output the non-dominated front set in the current population as the final Pareto optimal strategy set, covering typical control target combinations such as "high traffic efficiency, low lane-changing disturbance, and low congestion delay".

[0153] f. In this embodiment, the NSGA-II genetic algorithm is used to generate a set of Pareto optimal control strategies in each control period according to the real-time traffic state evaluation, including parameter combinations such as opening length, vehicle type restriction, speed limit value, and opening and closing time. This strategy set is not directly used as the controller output, but as an initialization candidate set for the action space of the reinforcement learning controller.

[0154] 6. Closed-loop feedback and reinforcement learning strategy iteration mechanism

[0155] To further improve the intelligent and adaptive ability of the hard shoulder opening control system of the expressway, this embodiment proposes a closed-loop strategy iteration mechanism based on reinforcement learning on the basis of multi-source data feature fusion and multi-objective strategy optimization, which dynamically adjusts the induction strategy through continuous interaction and has the ability to improve long-term performance.

[0156] Specifically, the system takes the traffic state feature vector obtained by current fusion perception as the input of the reinforcement learning controller, including the average speed of each lane, queue length, traffic density, compression index, truck occupancy rate, and the opening tendency score output by the fuzzy controller.

[0157] The reinforcement learning controller (such as DDPG) uses the optimal solution set output by the NSGA-II genetic algorithm to construct its selectable action set A, i.e.: Where: A t is the action candidate space at time t; is the kth Pareto optimal strategy output by NSGA-II; s t is the current environment state (provided by the perception fusion module).

[0158] On this basis, the reinforcement learning controller outputs an action vector based on the deterministic policy gradient method (DDPG), representing the current recommended hard shoulder opening strategy combination: a t = {L, C v , V limit , t open , t close}.

[0159] The action instruction acts on the simulation system or actual induction system through the TRACI interface, and the system collects the current period feedback indicators, including the average vehicle speed V avg , lane-changing disturbance intensity F change , and the estimated accident risk factor Raccident The immediate reward function is thus constructed: r t = ω1V avg - ω2F chang - ω3R accident ; wherein ω1, ω2, ω3 are target weighting coefficients, supporting flexible adjustment according to traffic targets.

[0160] The controller stores the four-tuple of current state, action, reward, and next state (s t , a t , r t , s t+1 ) into the experience pool, and updates the parameters through the policy network (Actor) and the value network (Critic). Both are constructed using two-layer fully connected neural networks, with the Actor network inputting traffic state and outputting continuous control action, and the Critic network inputting the combination of state and action and outputting the corresponding Q value estimate. The value network is based on the minimum mean square loss of target Q value, and the policy network maximizes the Q value output through policy gradient. To improve learning stability, the system introduces a soft update target network and combines the time difference method for reinforcement training.

[0161] The reinforcement learning controller proposed in this embodiment has high compatibility with the SUMO simulation system, and can complete policy evaluation and update once every simulation control period (30-120 seconds), while supporting online learning and offline batch training modes. The system can also use the NSGA-II optimal strategy as the initial action space of DDPG to improve convergence efficiency and policy interpretability.

[0162] The following will illustrate the scheme and effect by combining an actual scene simulation.

[0163] This embodiment selects a section of the Beijing-Shijiazhuang Expressway as an example. There are 32 pairs of monitoring cameras, 32 sets of broadcast, 20 sets of microwave vehicle detectors, and other hardware and software detectors in the section. The parameters detected by the coil detector include traffic flow, queue length, and time interval. The video detector on the road is a high-pixel camera that obtains specific traffic parameters through a video frame processing algorithm. Target information such as the speed of targets in a specific area, the unique identifier of the target, and the time interval can be obtained. Real-time sensing data can be provided for SUMO simulation.

[0164] 1. System composition and architecture

[0165] Multi-source data transmission and acquisition module: including hardware sensors (coil and radar detectors, video detectors) and software sensors, used to collect traffic flow, vehicle speed, lane occupancy, etc.

[0166] Communication and Protocol Module: Active data push is achieved using long connections (such as WEBSOCKET), combined with HTTP query mode for data interaction in low real-time scenarios; UDP communication is used for data with high real-time requirements (coil data, radar data, etc.), and TCP or MQTT protocols are used for scenarios with high data integrity requirements.

[0167] Database module: Used to persist various types of sensor data, configuration data, and large content files, and to provide support for subsequent data mining.

[0168] Client-side interaction module: Transforms raw sensor data into effective information that meets business needs, and achieves intelligent judgment of traffic conditions through data mining (real-time, timed, and historical analysis).

[0169] Simulation Module: Built on the SUMO simulation platform, this module simulates traffic conditions in real-world road networks. It uses XML-formatted map data for static road network modeling and employs multi-threading and dynamic step-size control to achieve basic synchronization between simulation and real-world time. The car-following model uses a modified Krauss model, and the lane-changing model is LC2013.

[0170] 2. Multi-source sensor data fusion and feature extraction

[0171] (1) Standardization of heterogeneous data

[0172] Hardware sensors: The coil detector uploads structured traffic, vehicle speed, and occupancy data at a frequency of 1Hz and transmits it to the edge computing node in real time via UDP protocol; the video detector transmits semi-structured video frame data via RTSP stream and extracts vehicle ID, trajectory, and queue length, outputting JSON format data packets.

[0173] Mobile terminal: The OBU device uploads GPS trajectory and lane change behavior data at a frequency of 0.5Hz through the 5GC-V2X communication module, and uses Protobuf encoding and compression for transmission.

[0174] Internet data: Web crawlers collect traffic congestion hot words from platforms such as Weibo and Gaode Map every hour, and extract semantic features using the BERT-NLP model; the meteorological API obtains regional weather data every 15 minutes and stores it in a structured MySQL database.

[0175] (2) Dynamic weight fusion calculation

[0176] Based on the aforementioned weight allocation model, a dynamic weight allocation matrix is ​​constructed. When the microwave radar detects sudden rainfall, its confidence weight α is... radar The weight of the video detector was increased to 0.4, while the weight of the video detector was reduced to 0.2 to address image recognition errors caused by decreased visibility.

[0177] The fusion output state vector is pushed to the SUMO simulation platform through the WebSocket interface, and an example of the data format is as follows:

[0178] {"timestamp":"2024-03-20T14:30:00Z","speed_avg":65.2,"queue_length":320,"density":28.5,"weather":"rain","incident":false,"shoulder_score":0.78}。

[0179] 3. Integration and synchronization process of sensor data and simulation platform

[0180] The system includes the following stages in the implementation of the data synchronization process:

[0181] (1) Static road network synchronization stage: Collect the static data of the actual road network in the target area (main road segments, interlaced areas, etc.) through OpenStreetMap, and generate XML format map data that meets the requirements of SUMO, and modify the map elements using the map editing GUI.

[0182] (2) Time synchronization stage: Dynamic time synchronization is the basis for subsequent operations after static road network synchronization. To ensure that the simulation processing time is consistent with the actual time, use the multi-threading operation mode in JAVA to control the simulation time. The main thread in multi-threading operation is used to implement simulation step advancement and time synchronization, and any other simulation business requirements are operated through secondary threads. The simulation step interval of the main thread is controlled by thread sleep time operation according to the simulation speed, so that the processing time between simulation and actual time is basically consistent. The multi-threading operation process is shown in Figure 6 . The multi-threading operation mode in the simulation system can help the sensor system to control the simulation progress. After simulation task division and task multi-threading operation, the simulation time change in the sensor system is shown in Figure 7 . From this, it can be seen that the multi-threading operation in the simulation can reduce the simulation time loss caused by user operation, and make the simulation time consistent with the actual time.

[0183] (3) Control Strategy Synchronization Phase: Speed ​​limit control and traffic strategy control are performed during control strategy synchronization. Maximum speed limits are controlled in the same way as the actual road network by restricting the speed of specific road segments and vehicle types. The TRACI interface provided by SUMO is used to map control strategies such as speed limits, road signs, and vehicle type restrictions from the actual road network to the simulation environment, ensuring consistency between the simulation environment and the real road network. Within the area, the speed limit is 120 km / h for passenger cars and 80 km / h for freight cars. During hard shoulder control synchronization, UDP technology is used to access actual variable message signs and other data to publish control strategies.

[0184] (4) During the sensor information synchronization phase, the sensor data of the weaving area and road segment are analyzed and converted into corresponding motion state vehicles. These simulated vehicles are then placed into the simulation for synchronized movement. Sensing areas are treated as nodes to separate the target movement areas in the simulation. Upstream and downstream sensors are considered as the generation and destination points of traffic demand, and OD points are created in the simulation to visually display the road network traffic state and divide the spatial simulation units as follows: Figure 8 As shown.

[0185] The simulated vehicle travel uses sensing areas as features for separating spatial travel segments. Segment camera detectors collect vehicle information for each segment, providing the simulated vehicle with speed and unique identification information. Therefore, the segment at the upstream detector is set as the travel demand generation point, the downstream exit detector as the travel demand destination, and the segment between the generation point and the destination serves as the simulation unit for the simulated vehicle. Figure 8 As shown, the simulated vehicle completes one trip of the simulation unit in simulated road segment 1.

[0186] (5) Guidance Measure Synchronization Phase: After completing the above synchronization phase, the process of constructing the simulation module from the actual road network is complete. Guidance measure synchronization maps the actual guidance measures into the simulation. Improving the quality of travel services and traffic management structure are important goals in constructing the sensing system. The integration of the simulation with the actual road network sensing system can more comprehensively and intuitively display the actual road conditions. Guidance measures for temporarily opening hard shoulders were implemented in the integrated sensing system, such as... Figure 9 As shown in the diagram, real-time and historical traffic data collected by various sensors are analyzed to determine the opening thresholds for main roads, merging, and merging scenarios. After the simulation is completed, a dynamic hard shoulder control strategy is implemented based on real-time data feedback to verify the effectiveness of the control strategy.

[0187] 4. Intelligent decision-making and control strategy generation

[0188] (1) Neuro-FLC dynamic decision-making

[0189] The input vehicle speed fluctuation rate and compression index are standardized by Z-score to eliminate dimensional differences. 32 expert rules are built in, such as:

[0190] IF vehicle speed fluctuation rate IS High (μ = 0.7) AND queue length change rate IS Rising (μ = 0.6)

[0191] THEN open tendency = Strong (output score + 0.5)

[0192] Based on historical data, a training set is constructed (input: traffic state vector, output: optimal opening decision), and the Adam optimizer is used to iteratively update the membership function parameters.

[0193] (2) NSGA-II multi-objective optimization

[0194] The control strategy parameters use hybrid coding, chromosome = [350, 0, 80, 3600, 7200] represents opening 350 meters of hard shoulder, allowing only small cars to pass, speed limit 80 km / h, opening period is from 3600th second to 7200th second. And impose a penalty function on individuals that violate the hard shoulder width-vehicle type matching:

[0195] if width < 3.75m and vehicle_type!= 0:

[0196] fitness* = 0.5 # penalty function to weaken fitness

[0197] Pareto frontier screening: output 15 non-dominated solution sets after 50 generations, select the comprehensive optimal strategy by TOPSIS method. In the early stage of system deployment or when the reinforcement learning controller has not completed the strategy network training, the system uses the TOPSIS method to select the strategy with the largest closeness index from the multiple Pareto solutions output by NSGA-II, according to the weight priority set by the management side (such as traffic efficiency priority or safety priority), and selects the strategy with the largest closeness index as the current control instruction.

[0198] The selection process is as follows:

[0199] a, construct decision matrix: construct a standardized decision matrix according to the three objective indexes {V avg ,-C delay ,-F change} of all Pareto non-dominated solutions, where the negative sign indicates that the cost item needs to be minimized.

[0200] b, normalization: range normalization is performed on each objective index to form a dimensionless evaluation matrix.

[0201] c. Set target weights: Set the weights of each target according to the preferences of the road management department, for example, ω = [0.4, 0.3, 0.3], corresponding to traffic efficiency, delay cost, and lane change disturbance, respectively.

[0202] d. Construct ideal and negative ideal solutions: Define the theoretical optimal solution (i.e., the optimal value of each target indicator) and the worst solution. Calculate the Euclidean distance between each solution and the ideal solution and the Euclidean distance between each solution and the negative ideal solution Calculate the relative closeness: where: T i The larger the value, the closer the solution is to the optimal solution.

[0203] e. Select the optimal strategy: Select the individual with the largest closeness T i as the comprehensive optimal control strategy for the current period, and use it as the reinforcement starting point for the DDPG controller.

[0204] (3) Reinforcement learning closed-loop optimization

[0205] As the system runs for a period of time, the reinforcement learning controller (using the DDPG algorithm) constructs state-action-reward triplets based on traffic perception states, NSGA-II output strategy sets, and feedback indicators, and trains the strategy network (Actor) and value network (Critic) in continuous interaction. In this phase, the controller no longer relies on TOPSIS for static selection, but instead selects actions adaptively from the Pareto solution set.

[0206] The input vector of the state space includes 6 features: [speed_avg, queue_length, density, truck_ratio, shoulder_score, weather_level]. The action space is based on the discrete action set constructed from the Pareto solution set output by NSGA-II, action = [open length, vehicle restriction, speed limit value, open time, close time] ∈ {200m ≤ length ≤ 500m, 0 / 1 / 2, 60 / 80 / 100km / h, 1800-3600s, 3601-6000s}, and the reward function is a weighted combination of the three indicators:

[0207] reward = 0.6 × Δspeed - 0.3 × lane_change_penalty

[0208] -0.1 × risk_factor

[0209] The experience pool uses Prioritized Experience Replay with a storage capacity of 10^5 transition samples. The Actor-Critic network uses a double-target network structure, and the parameters are synchronized every 100 steps to prevent overfitting.

[0210] 5. Dynamic hard shoulder control execution

[0211] (1) Command issuance

[0212] The optimal strategy is pushed to the Road Side Unit (RSU) through the MQTT protocol to drive the Changeable Message Signs (CMS) and lane control signs: {Topic: / highway / shoulder_control

[0213] Payload: {"action": "open", "length": 350, "speed_limit": 80, "valid_time": 3600}}.

[0214] (2) Real-time control algorithm

[0215] For dynamic hard shoulder opening control, the system designs the following control process:

[0216] Step 1: Extend in SUMO road network file <lane>Tag attribute, set hard shoulder only allow certain vehicle type pass, and define dynamic speed limit parameters;

[0217] <lane id="shoulder_1" width="3.0" allow="passenger" / >

[0218] <vClassSpeed dev="0.2" vClass="passenger" speed="33" / >

[0219] <vClassSpeed dev="0.2" vClass="truck" speed="22" / >

[0220] Step two: through the main thread loop to obtain real-time data of sensor, according to the control strategy, once the condition is met, the control instruction is issued by using TRACI interface, and the simulation parameters are modified in real time;

[0221] Step three: at the same time, dynamic simulation step control algorithm is adopted to keep the simulation time and actual time synchronous, so as to realize real-time dynamic control.

[0222] The specific algorithm pseudo code is as follows:

[0223] Table 4 dynamic simulation Python pseudo code

[0224]

[0225] Experimental analysis and system advantages

[0226] In order to verify the universal applicability of the embodiment, three typical scenes of opening hard shoulder of highway network are determined, which are main line hard shoulder opening, outflow ramp hard shoulder opening and inflow ramp hard shoulder opening.

[0227] Main line hard shoulder opening: mainly used for expanding free flow scene, reducing the mutual interference influence of vehicle driving in vehicle flow, and trying to let the vehicle drive in the state close to the speed limit. Mainly solve the problem of reducing the overall driving efficiency caused by large truck multi-lane occupation.

[0228] The experimental results of this embodiment in the two-way 4-lane and 6-lane highway scene show that the intelligent control system using Neuro-FLC dynamic decision, NSGA-II multi-objective optimization and DDPG reinforcement learning significantly improves the road traffic efficiency. In the two-way 4-lane scene, the system dynamically generates the optimal opening strategy (small car speed limit 100km / h) by real-time evaluation of the speed fluctuation rate and the queuing change rate, which makes the average speed increase from 76.6km / h to 83.6km / h (increased by 9.13%), while the lane changing disturbance is reduced by 31.2% and the accident risk index remains below the safety threshold of 0.15. For the two-way 6-lane scene, the system uses the mixed traffic strategy (truck speed limit 80km / h), which increases the average speed from 78.7km / h to 84.8km / h (increased by 7.75%) and reduces the lane changing disturbance by 28.5%. It effectively relieves the road congestion and improves the traffic efficiency.

[0229] Table 5 Main line hard shoulder opening results

[0230]

[0231] The export ramp hard shoulder opening: mainly used for the scene that the export vehicle queue affects the normal driving of the main road, a certain distance upstream of the export is prompted that only the ramp-down vehicles can enter the opened hard shoulder to avoid the straight vehicles entering.

[0232] In the two-way 4-lane export scene, the system dynamically generates the optimal opening strategy (small car speed limit 80km / h) by real-time evaluation of the ramp demand ratio and the main line traffic density, which increases the average speed in the import area from 58.3km / h to 65.7km / h (increased by 12.7%), while the import conflict is reduced by 37.6% and the accident risk index remains below the safety threshold of 0.18. For the two-way 6-lane scene, the system uses the mixed traffic strategy by time period (small car in peak period), which increases the average speed in the import area from 61.2km / h to 67.5km / h (increased by 10.3%) and reduces the import disturbance by 32.8%. It effectively relieves the traffic pressure of the main line and ensures the traffic efficiency and road section operation stability of the vehicles.

[0233] Table 6 Import hard shoulder opening results

[0234]

[0235]

[0236] The import ramp hard shoulder opening: mainly used for the scene that the import vehicle interweaving affects the normal driving of the main road, the hard shoulder is prohibited upstream of the import to avoid excessive conflict with the ramp traffic. The hard shoulder is opened downstream of the import as a speed-up and import lane, and a certain length is set.

[0237] In the two-way 4-lane merging scene, the system generates the optimal opening strategy (passing priority for small cars) by real-time evaluation of the mainline traffic flow (2450 pcu / h) and ramp merging demand (800 pcu / h), so that the average speed of the hard shoulder section is increased to 73.4 km / h, the mainline capacity reaches 3050 pcu / h (increased by 24.4%), while the merging conflict is reduced by 39.2%, and the accident risk index is controlled below 0.17. For the two-way 6-lane scene, the system uses the dynamic lane allocation strategy according to the mainline flow of 3750 pcu / h and the merging flow of 800 pcu / h, so that the average speed of the section is maintained at 73.6 km / h, the capacity is increased to 4460 pcu / h (increased by 18.8%), and the merging disturbance is reduced by 35.6%. In this process, the right-turn vehicles waiting to merge can quickly merge into the main road with the help of the hard shoulder, effectively relieving congestion and improving traffic efficiency.

[0238] Table 7 opening results of merging hard shoulder

[0239]

[0240]

[0241] The embodiment of the application further discloses a computer system, including a memory, a processor and a computer program stored on the memory and executable on the processor, when the computer program is executed by the processor, the steps of the expressway hard shoulder dynamic opening method based on the NFGD model and the simulation platform are realized.

[0242] The embodiment of the application further discloses a computer program product, including a computer program, when the computer program is executed by the processor, the steps of the expressway hard shoulder dynamic opening method based on the NFGD model and the simulation platform are realized.

[0243] The above embodiments are only preferred embodiments of the application, and those skilled in the art can make various modifications and changes to the above embodiments without departing from the spirit and protection scope of the application, which should be regarded as the protection scope of the application.< / lane>

Claims

1. A method for dynamically opening hard shoulders of highways based on NFGD models and simulation platforms, characterized in that, Includes the following steps: Multi-source real-time sensor data were collected and weighted using a dynamic confidence mechanism to obtain fused traffic state features. Dynamic control decisions for the hard shoulder were made based on an NFGD model comprising a fuzzy-neural hybrid controller, a multi-objective genetic algorithm controller, and a reinforcement learning controller. This included: inputting the fused state vector into the fuzzy-neural hybrid controller for real-time opening tendency assessment, outputting an opening tendency score, and determining whether to trigger strategy optimization; when strategy optimization was triggered, a multi-objective genetic algorithm was used to collaboratively solve multiple conflicting objectives, outputting a set of Pareto optimal control strategies. Control parameters included opening length, vehicle type restrictions, speed limits, and opening / closing times. The section describes a reinforcement learning controller constructed using perceived states, optimal control policy sets, and historical feedback. It outputs action vectors based on a deterministic policy gradient method, representing the currently recommended combination of hard shoulder opening policies. The optimal control policy set serves as the basis for the action space of the reinforcement learning controller. Action commands are applied to the simulation system or the actual induction system. The system collects current periodic feedback indicators and iteratively optimizes the policy network and value network to form a closed-loop adaptive control policy. In the section on using a multi-objective genetic algorithm to collaboratively solve multiple conflicting objectives, the NSGA-II non-dominated sorting genetic algorithm is used to generate the multi-objective control policy, with the optimization objective function being max{V avg ,-C delay ,-F change }, V avg C represents the average speed of all vehicles within the observation area. delay For the cost of congestion and delay, F change To average lane change disturbance intensity; the control parameters are encoded into a decision variable vector X = [L, C] v V limit ,t open ,t close ], L is the length of the open hard shoulder section, C v For permitted vehicle types, V limit For the corresponding section speed limit, t open t is the start time of opening. close For the closing time, the control parameters are represented using a mixed integer real number encoding method; the constraints include that the road can only be opened when the hard shoulder width is greater than or equal to a preset threshold, V limit Within the preset range, the opening and closing timing must satisfy t. open ≤t close The vehicle type and width conditions are matched. It interacts with the simulation platform in real time to enable the issuance of control commands, the collection of feedback, and online strategy evaluation.

2. The method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform according to claim 1, characterized in that, The fused traffic status features include average vehicle speed, queue length, compression, truck occupancy rate, traffic density, and weather and accident risk information. Observations of the same variable by multiple sensors Through standardized weights w k (t) Weighted fusion; weights are calculated using a dynamic confidence mechanism. Standardized weights It satisfies the normalization constraint; where K represents the number of sensors. Let Vi represent the observation value of variable i from the k-th sensor, and Var represent the variance of the observation sequence within a certain time window. σi 2 It is expressed as a normalization factor or reference variance constant, used to adjust the tolerance for fluctuation amplitude.

3. The method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform according to claim 1, characterized in that, In the fuzzy-neural hybrid controller, the fuzzy inference layer incorporates expert rules and uses Gaussian membership functions to calculate the fuzzy values ​​of the input variables; the neural network optimization layer trains the fuzzy rule parameters online through a three-layer feedforward network, with training data derived from a set of benefit comparisons of open strategies in historical simulations; when the output openness tendency score is greater than the score threshold, or when the output openness tendency score is greater than the score threshold and the shoulder width and / or traffic risk meet the set conditions, an open command is triggered to optimize the control strategy.

4. The method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform according to claim 1, characterized in that, A multi-objective genetic algorithm is used to first generate the Pareto optimal policy set, and then the TOPSIS method is used to select the comprehensive optimal solution as the reinforcement starting point of the reinforcement learning controller.

5. The method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform according to claim 1, characterized in that, The state space of the reinforcement learning controller includes the average speed of each lane, queue length, traffic density, compression index, truck occupancy rate, and the opening tendency score output by the fuzzy-neural hybrid controller. The action space is constructed based on the Pareto solution set output by the genetic algorithm to form a discrete action set. The actions include the length of the open hard shoulder section, the type of vehicles allowed to pass, the corresponding speed limit value of the section, the opening start time and closing time. The reward function considers the average vehicle speed, the lane changing disturbance intensity, and the accident risk factors estimated by simulation.

6. The method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform according to claim 1, characterized in that, Different expert rules are set up and different models are trained for different scenarios, such as opening the hard shoulder of the main road, opening the hard shoulder of the merging ramp, and opening the hard shoulder of the merging ramp.

7. The method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform according to claim 1, characterized in that, The simulation platform dynamically modifies simulation parameters by expanding lane attribute labels and speed limit configurations, and uses a multi-threaded and dynamic step size adjustment mechanism to ensure that simulation time is synchronized with actual time. It controls the travel path of simulated vehicles by setting OD points.

8. A dynamic opening system for highway hard shoulders based on an NFGD model and simulation platform, used to implement the dynamic opening method for highway hard shoulders based on an NFGD model and simulation platform according to any one of claims 1-7, characterized in that, include: The data transmission and acquisition module is used to collect and weighted fuse multi-source real-time sensor data using a dynamic confidence mechanism to obtain the fused traffic state characteristics. The system control module is used for dynamic control decisions of hard shoulders based on an NFGD model comprising a fuzzy-neural hybrid controller, a multi-objective genetic algorithm controller, and a reinforcement learning controller. This includes: inputting a fused state vector into the fuzzy-neural hybrid controller for real-time opening tendency judgment, outputting an opening tendency score, and determining whether to trigger strategy optimization; when strategy optimization is triggered, using a multi-objective genetic algorithm to collaboratively solve multiple conflicting objectives, outputting a set of Pareto optimal control strategies, with control parameters including opening length, vehicle type restrictions, speed limits, and opening / closing periods; constructing a reinforcement learning controller using the perceived state, the optimal control strategy set, and historical feedback, outputting action vectors based on a deterministic policy gradient method, representing the currently recommended combination of hard shoulder opening strategies, where the optimal control strategy set serves as the basis for the action space of the reinforcement learning controller; and applying action commands to the simulation system or the actual guidance system, with the system collecting current periodic feedback indicators, iteratively optimizing the policy network and value network to form a closed-loop adaptive control strategy. The simulation module is used to interact with the simulation platform in real time, enabling the issuance of control commands, the collection of feedback, and online strategy evaluation.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for dynamic opening of highway hard shoulders based on NFGD model and simulation platform as described in any one of claims 1-7.

Images