Intersection phase structure optimization method based on large language model

Abstract

The invention belongs to the technical field of urban traffic planning and intelligent traffic systems, particularly relates to an intersection phase structure optimization method based on a large language model, and aims to improve the intelligent level and operation efficiency of traffic signal control. According to the method, semantic mapping cues of traffic flow and a phase structure are constructed, and a large language model is guided to generate a phase structure scheme adapted to an actual traffic state. Compared with a traditional scheme depending on artificial experience and a fixed structure, the method can automatically generate diversified and data-driven phase structure combinations, and has higher adaptability and generalization ability. In the aspect of technical implementation, the method fuses prompt engineering and guides a large language model to generate an initial phase scheme, and performs evaluation and feedback by using a value network, so that optimization of a phase structure is realized, and the overall operation efficiency of a traffic system is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of urban traffic planning and intelligent transportation systems, relates to large language model technology and phase structure combination method, and specifically relates to a phase structure optimization method for an intersection based on a large language model. BACKGROUND

[0002] The phase structure of an intersection is a global design framework for allocating traffic rights in a signal cycle in a traffic signal control system, and the core thereof is composed of a phase combination scheme and a phase switching scheme. The phase combination scheme defines a traffic flow set that can share traffic rights in a specific time window, and eliminates the motion conflict within the intersection by discretizing the traffic timing of the conflicting traffic flow, so as to maximize the intersection traffic efficiency, and therefore the phase combination scheme is the first key step in signal timing optimization. However, the traditional phase optimization method highly depends on historical data and has problems such as low efficiency and poor adaptability. With the continuous development of artificial intelligence technology, some studies attempt to introduce methods such as reinforcement learning to optimize phase parameters, such as the document “Single Intersection Traffic Signal Control Method Based on Q-Learning” which proposes to dynamically adjust the phase length through the Q-Learning algorithm, but such methods still face many challenges, such as being limited by the constraints of fixed phase structure, being difficult to meet the demand for dynamic adjustment of phase structure under complex traffic flow scenarios, having a large state space dimension resulting in low learning efficiency and poor model generalization ability, and the “black box” structure lacking explainability, which limits its deployment feasibility in real traffic systems. In recent years, large language models have shown strong semantic understanding and logical reasoning capabilities in the field of natural language processing, but their application in the field of traffic signal control is still in the exploratory stage. SUMMARY

[0003] To solve the problems of traditional phase structure optimization such as strong dependence on historical data and poor adaptability of fixed phase structure, the application provides a phase structure optimization method based on a large language model, which realizes the dynamic generation of intersection phase structure by fusing traffic data and the semantic reasoning capability of a large language model. The application realizes automatic generation of phase structure schemes through a semantic reasoning and feedback optimization mechanism, effectively improving the intersection traffic efficiency and adaptability.

[0004] The technical scheme of the application is as follows:

[0005] A phase structure optimization method for an intersection based on a large language model, which includes four core steps: data preparation, simulation platform construction, value network training, and large language model optimization (including interaction, cycle calculation, and iterative optimization). By constructing the semantic mapping relationship between traffic flow characteristics and phase structure, the large language model can select the phase structure.

[0006] The specific steps are as follows:

[0007] Step 1: Preparation of traffic flow data and road network information data

[0008] The process involves collecting traffic flow and road network information for the area to be optimized within a specific time period. This information is used for traffic simulation model construction and phase optimization scheme design, and includes two parts:

[0009] Step 1.1, Traffic Flow Data Collection:

[0010] Data such as through traffic flow and turning traffic flow at the intersection to be optimized within a specific time period are collected using detection equipment such as geomagnetic sensors, radar, and video detectors.

[0011] Step 1.2: Obtaining Road Network Information Data:

[0012] Obtain the road network geometry parameters (such as road network topology, number of lanes, lane function division, channelization design, etc.), traffic control parameters (such as existing phase structure, cycle duration), and traffic management system (such as traffic control information such as no left turn, no right turn, etc.) of the target area.

[0013] Step 2: Construction of the simulation platform

[0014] Based on the data obtained in step 1, a traffic simulation model is constructed to simulate the traffic operation effects under different phase structures, and the following functions are achieved through simulation software:

[0015] Step 2.1: Dynamically load real-time traffic flow data;

[0016] Step 2.2: Simulate traffic flow operation under different phase structure schemes;

[0017] Step 2.3: Output multi-objective performance evaluation indicators, including delay, queue length, and throughput capacity.

[0018] Step 3: Value Network Training

[0019] Based on the constructed traffic simulation model, the value network is trained, specifically including the following sub-steps:

[0020] Step 3.1, Data Generation:

[0021] In traffic simulation models, the traffic flow at intersections is randomly perturbed to construct diverse traffic scenarios; let the original traffic flow at a certain intersection be... The new flow after the disturbance is :

[0022]

[0023] in, The set flow rate adjustment range parameter, It is a random function.

[0024] Step 3.2, Phase Scheme Simulation:

[0025] For each generated traffic scenario, all possible phase structure schemes are enumerated, the Webster method is used to calculate the base green light ratio for each scheme, and the multi-objective evaluation results for each scheme are recorded; let the set of phase structure schemes be . For any phase structure scheme Its traffic flow Comprehensive evaluation indicators for:

[0026]

[0027] in, Represents the topological feature vector of an intersection, containing static attributes such as geometric layout and lane configuration; function Represents traffic and phase scheme Average waiting time for vehicles, function Represents traffic and phase scheme Average vehicle travel time; Represents the weighting coefficient, satisfying and , representing the weights of waiting time and passage time in the comprehensive evaluation, respectively. Their weight values ​​depend on the optimization objective and can be initially set to 0.5; This represents the normalization coefficient, used to eliminate the influence of dimensions, and is taken from the current traffic scenario. The maximum value of the corresponding indicator among all the alternative options, i.e. .

[0028] Step 3.3, Data Augmentation:

[0029] Repeat steps 3.1 and 3.2 to enhance the diversity and coverage of the data. The final dataset is then constructed. Depend on The sample consists of 5 tuples, each of which is a quintuple:

[0030]

[0031] in, For the first Traffic flow data generated by random perturbation. For the number of perturbations, This represents the total number of phase structure schemes.

[0032] Step 3.4, Multi-intersection data expansion:

[0033] Assume there is a total There are several intersections, for each intersection Repeat steps 3.1 to 3.3 to obtain the complete dataset:

[0034]

[0035] in, Represents a multi-intersection dataset. Indicates the first A dataset of intersections;

[0036] Step 3.5, Value Network Training:

[0037] Based on the constructed multi-intersection dataset Establish a nonlinear residual neural network model As a value network, it is used to predict the comprehensive evaluation index of phase schemes. Network model. It consists of an embedding layer and residual blocks, and accepts three types of input features: intersection topology features. Traffic flow information and phase scheme The traffic light parameters are used; among them, topological features and traffic light parameters are used as classification features and are soft-encoded through an embedding layer, and then combined with numerical traffic information to form a feature vector. Network mapping can be represented as ,in For comprehensive evaluation indicators The predicted value, These are the network parameters. The training objective is to minimize the mean squared error loss function. :

[0038]

[0039] in, The total number of training samples, The actual evaluation index is obtained from the simulation.

[0040] Step 4: Interactive Iteration of the Large Language Model

[0041] A large language model is used to intelligently generate and interactively optimize the phase structure scheme. This process combines actual traffic flow, road network information, and professional traffic knowledge to obtain a traffic signal control strategy. The obtained traffic signal control strategy is evaluated using the network model trained in step 3. Specifically, it includes the following sub-steps:

[0042] Step 4.1, Prompt word generation and input construction:

[0043] Traffic flow information based on the intersection to be optimized Transportation network structure And supplementary traffic background information Constructing prompt words This is used to guide the generation of the initial phase structure scheme for a large language model. The prompt word generation function is:

[0044]

[0045] in, This is a prompt word generation function used to transform the input into a text format acceptable to large language models. It can also contain heuristic language optimization strategies to guide the output control suggestions of large language models.

[0046] prompt words Interact with large language models (such as GPT-like models) to obtain their generated initial phase structure schemes. Each of them A signal configuration that indicates a phase, including the lanes and steering directions covered.

[0047] Step 4.2: Calculate the green light ratio based on the phase structure constructed from the large language model:

[0048] To convert the text-based schemes generated by the large language model into implementable control parameters, it is necessary to calculate the specific signal cycles and green light durations for each phase using traffic engineering methods, based on the phase structure scheme provided by the large language model. By combining traffic flow information with Webster's optimal cycle length calculation method, the optimal cycle length is calculated. Compared to the green light of each import channel. If there is a common Each phase, Then the period The calculation formula is:

[0049]

[0050] in, This represents the total lost time within the period. For the first The green light ratio of each phase satisfies:

[0051]

[0052] in, For the first Green light duration for each phase;

[0053] If the first The intersection entrance lanes with phase action have If there are 1, then the set of calculated green light durations for that phase entrance lane is 1. :

[0054]

[0055] Step 4.3, Solution Evaluation and Value Network-Driven Iterative Optimization:

[0056] The performance of phase structure schemes generated by a large language model is evaluated using a value network, and prompt words are updated based on the evaluation results. This guides the large language model to iteratively generate better-performing schemes. The process specifically includes:

[0057] Step 4.3.1, Solution Evaluation:

[0058] The phase structure scheme and green ratio calculated in step 4.2, as well as the current intersection topology and traffic flow, are used as inputs and fed into the value network model trained in step 3 to obtain the performance evaluation results of the scheme under the current traffic scenario.

[0059] Let the current solution be The comprehensive evaluation index of the value network output is:

[0060]

[0061] in, For current traffic flow, For the current phase structure, For a trained neural network model, These are network parameters.

[0062] Step 4.3.2, Optimality Judgment:

[0063] Compare the current solution score with the historical best result. The solution is compared; if the current solution scores better than the known optimal solution, the optimal solution is updated. :

[0064]

[0065] Step 4.3.3, Feedback Guidance and Prompt Update:

[0066] If the current solution does not achieve the preset optimal evaluation, or if there is still significant room for optimization, then feedback information containing the structure and performance score of the current solution is constructed. Update the prompt words of the large language model. :

[0067]

[0068] Among them, the function By combining traffic data, background information, and performance feedback, new optimization prompts are constructed to guide the large language model in generating the next round of solutions. .

[0069] Step 4.3.4, Iterative optimization process:

[0070] Steps 4.3.1-4.3.3 involve continuous iteration. In each round, the large language model generates new candidate solutions, and the value network independently scores their performance. This process continues until the solution score reaches the set optimal evaluation target or the maximum number of iterations is reached. The process terminates at a certain point. Finally, the output is the phase structure scheme evaluated as optimal by the value network. As the final traffic signal control configuration scheme.

[0071] The beneficial effects of this invention are:

[0072] (1) Dynamic phase structure generation: By utilizing the semantic reasoning capabilities of large language models, the model automatically identifies the matching relationship between traffic flow characteristics and phase structure, breaking through the limitations of traditional fixed phase schemes. For example, when the left-turn flow exceeds the threshold, the model can autonomously generate a dedicated left-turn phase and adjust the phase order.

[0073] (2) Enhanced real-time adaptability: Based on real-time traffic data, the phase scheme is optimized, and the response cycle is shortened from hours to minutes in the traditional method, which significantly improves the ability to respond to sudden traffic events.

[0074] (3) Reduce reliance on human intervention: Transform the phase structure design from the design of traffic engineers to the automated design of large language models, thereby reducing the deviation of the scheme caused by human factors. Attached Figure Description

[0075] Figure 1 This is a basic flowchart of the method of the present invention.

[0076] Figure 2 This is a road network structure diagram (6 intersections) of a certain area in an embodiment of the present invention.

[0077] Figure 3 This is a road network structure diagram (intersection 31) of a certain area in an embodiment of the present invention. Detailed Implementation

[0078] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0079] The basic process of the method of this invention is as follows: Figure 1 As shown, the following example uses SUMO software to select a portion of the road network in a certain area (31 intersections and 6 simplified intersections of the road network) as a calculation example.

[0080] Step 1: Preparation of traffic flow data and road network data

[0081] Step 1.1: Traffic Flow Data Preparation

[0082] Traffic flow and density at intersections in the area between 7:00 and 8:00 are obtained through traffic sensors and video surveillance.

[0083] Step 1.2: Road Network Data Preparation

[0084] Data required to obtain the road network topology and related traffic information of the target area: number of lanes, lane capacity, and speed limits of the road network in the target area.

[0085] Step 2: Setting up the simulation platform

[0086] Based on the collected traffic flow data and road network data, a traffic simulation model is built. An example area road network is shown below. Figure 2 As shown.

[0087] Step 3: Value Network Training

[0088] Based on the simulation model built in step 2, multiple intersections to be optimized were selected, and disturbance flow samples and corresponding combinations of phase structures were constructed. For each intersection, a random function was used to disturb the straight-through and turning flows, forming diverse traffic scenarios. For each disturbance traffic scenario, phase structure schemes were enumerated and evaluation indicators were calculated according to the Webster method. The above process was repeated to form multiple sets of training data, covering multiple intersections and flow intervals. A three-layer fully connected neural network was constructed, using phase structure and flow characteristics as inputs, to predict traffic performance indicators. The model used the activation function ReLU (Rectified LinearUnit), the optimizer Adam (Adaptive Moment Estimation), and set the learning rate to 0.001. The loss function was the MSE (Mean Squared Error) function, and the training data and test data were divided in an 8:2 ratio.

[0089] This value network is obtained as a scorer for subsequent evaluation and feedback iteration of large language model schemes.

[0090] Step 4: Interaction of the Large Language Model

[0091] Step 4.1: Dynamic prompt word generation

[0092] Input traffic flow information, road network information, and supplementary traffic background information of the intersection to be optimized, and construct prompt words that can clearly and accurately convey the problem context and solution objectives to the large language model (DeepSeek-V3 is used in this example).

[0093] Step 4.2: Interacting with the large language model

[0094] The prompts automatically generated based on traffic flow characteristics are input into the large language model, and the model infers the initial phase structure scheme.

[0095] Step 4.3: Calculate the period duration and preliminary phase duration.

[0096] Based on the large language model, a phase structure scheme and traffic flow data are given, and the signal cycle and green light duration for each phase are calculated using the Webster method.

[0097] Step 4.4: Write the calculated results into the configuration file.

[0098] Write the phase selection given by the large model and the calculated results into the configuration file to create a signal light configuration file for a single intersection.

[0099] Step 4.5: Iterative Optimization

[0100] Step 4.5.1: Integrate the phase scheme given by the large model with the calculated green ratio information into a complete scheme;

[0101] Step 4.5.2: Use the model trained in step 3 to judge the merits of the large language model scheme;

[0102] Step 4.5.3: If the current solution does not reach the preset performance threshold or is not better than the historical best value, then construct feedback information to prompt for an update.

[0103] Step 5: Run the simulation and perform result analysis

[0104] In the streamlined road network of 6 intersections (such as...) Figure 2 The test was conducted in the system shown in the figure. After checking the road network structure and related phase configuration files, SUMO simulation was performed and the output results were analyzed. The weighted average waiting time of vehicles was selected as the comparison index. To ensure the scientificity and accuracy of the results, the control effects of the phase scheme before and after optimization were compared in multiple sets of experiments. The data are shown in Table 1 (the original scheme refers to the manually configured phase scheme in the current actual project).

[0105] Table 1: Comparison of Phase Scheme Results for Six Intersections

[0106]

[0107] As shown in the table, the phase structure of this method can achieve a certain degree of optimization at most intersections. Looking at the overall 6-intersection road network, the phase scheme generated by this method based on the large language model achieves a 17% optimization in weighted average waiting time compared to the original phase scheme, verifying the optimization capability and practical applicability of this method in small-scale road networks.

[0108] In addition, in the complete network of 31 intersections (such as...)Figure 3 The test was conducted in the system shown in the figure. After checking the road network structure and related phase configuration files, SUMO simulation was performed and the output results were analyzed. The weighted average waiting time of vehicles was selected as the comparison index. To ensure the scientific validity and accuracy of the results, multiple sets of experiments were conducted to compare the results of the phase schemes. The data are shown in Table 2.

[0109] Table 2: Comparison of Phase Scheme Results for 31 Intersections

[0110]

[0111] Experimental results show that in a large-scale urban road network covering 31 intersections, the phase structure scheme generated by this method based on a large language model reduces the weighted waiting time by an average of about 20% compared with the original scheme, fully demonstrating its optimization potential and generalization ability in large-scale traffic environments.

[0112] This invention achieves automatic generation and performance evaluation of signal control schemes by constructing a semantic mapping between traffic data and phase structure. By combining a value network as a scorer and guiding iterative optimization based on the score results, a data-driven, structurally dynamic, and performance-oriented intelligent signal control scheme configuration process is realized.

Claims

1. A method for optimizing intersection phase structure based on a large language model, characterized in that, The steps are as follows: Step 1: Preparation of traffic flow data and road network information data Collect the traffic flow and road network information of the area to be optimized in a specific time period for traffic simulation model construction and phase optimization scheme design, which includes: Step 1.1, traffic flow data collection: Collect the straight traffic flow and turning traffic flow data of the intersection to be optimized in a specific time period through detection equipment including geomagnetic sensors, radars, and video detectors; Step 1.2, acquisition of road network information data: Obtain the road network geometric parameters, traffic control parameters, and traffic management system of the target area; Step 2: Construction of simulation platform Based on the data obtained in step 1, construct a traffic simulation model; Step 3: Value network training Based on the constructed traffic simulation model, train the value network, which includes the following sub-steps: Step 3.1, data generation: Random disturbance is made to the intersection flow in the traffic simulation model to construct diversified traffic scenarios. Let the original flow of a certain intersection be and the new flow after disturbance be : wherein, is a set flow adjustment magnitude parameter, is a random function; Step 3.2, phase scheme simulation: For each generated traffic scene, all possible phase structure schemes are enumerated, the base green ratio of each scheme is calculated by using Webster method, and the multi-objective evaluation results of each scheme are recorded; let the phase structure scheme set be For any phase structure scheme , the comprehensive evaluation index under the traffic flow is: wherein, represents intersection topology feature vector, including geometric layout, lane configuration; function represents the average waiting time of vehicles under traffic and phase scheme ; function represents traffic and phase scheme ; vehicle average travel time; represents weight coefficient, satisfying and , respectively representing the weight of waiting time and travel time in comprehensive evaluation; represents normalization coefficient, taking the maximum value of the corresponding index in all alternative schemes under the current traffic scene , that is ; Step 3.3, data augmentation: Steps 3.1 and 3.2 are repeated to enhance the diversity and coverage of the data; the final constructed dataset consists of samples, each sample being a five-tuple: wherein, is the traffic flow data generated by the is the number of perturbations, is the total number of phase structure schemes; Step 3.4, multi-intersection data expansion: Let there be crossings, for each crossing repeat steps 3.1 to 3.3, resulting in a complete data set: wherein, represents a multi-intersection data set, represents a data set of the th intersection; Step 3.5, value network training: Step 4: Interactive iteration of large language model Use the large language model to intelligently generate and interactively optimize the phase structure scheme. This process combines actual traffic flow, road network information, and professional traffic knowledge to obtain a traffic signal control strategy. The obtained traffic signal control strategy is evaluated through the network model trained in step 3.

2. The intersection phase structure optimization method based on a large language model according to claim 1, characterized in that, In step 1.2, the road network geometric parameters include road network topology, lane number, lane function division, and channelization design. The traffic control parameters include existing phase structure and cycle length. The traffic management system includes left turn prohibition and right turn prohibition.

3. The intersection phase structure optimization method based on a large language model according to claim 1, characterized in that, The traffic simulation model in step 2 is used to simulate the traffic running effect under different phase structures and achieve the following functions through simulation software: Step 2.1, dynamically load real-time traffic flow data; Step 2.2, simulate the traffic flow running state under different phase structure schemes; Step 2.3, output multi-objective performance evaluation indicators including delay, queue length, and capacity.

4. The intersection phase structure optimization method based on a large language model according to claim 1, characterized in that, Step 3.5 is as follows: Based on the constructed multi-intersection dataset , a nonlinear residual neural network model is established As a value network, the prediction of comprehensive evaluation index for phase scheme; network model It is composed of embedding layer and residual block, and accepts three types of input features: intersection topology features , traffic flow information and signal lamp parameters of phase scheme ; among them, the topology features and signal lamp parameters are soft coded as classification features through the embedding layer, and then combined with the numerical traffic flow information as a feature vector ; the network mapping is represented as , wherein is the predicted value of the comprehensive evaluation index , and is the network parameter; the training target is to minimize the mean square error loss function : wherein, is the total number of training samples, is the real evaluation index obtained by simulation.

5. The intersection phase structure optimization method based on a large language model according to claim 1, characterized in that, Step 4 is as follows: Step 4.1, prompt word generation and input construction: Based on the traffic information of the intersection to be optimized , traffic network structure , and supplementary traffic background information , construct prompt words for guiding large language models to generate initial phase structure schemes; the prompt word generation function is: wherein, is a prompt generation function for converting the input into a text format acceptable to the large language model, including heuristic language optimization strategies for guiding the large language model to output control recommendations; The prompt word Interact with a large language model to obtain its generated initial phase structure scheme Wherein each Represents the signal configuration of a phase, including the covered lanes and turns; Step 4.2, green ratio calculation of phase structure based on large language model: In order to convert the text scheme generated by the large language model into the control parameters that can be implemented, the specific signal cycle and the green time of each phase are calculated according to the traffic engineering method, and the phase structure scheme given by the large language model is combined with the traffic flow information and the Webster optimal cycle time calculation method to calculate the optimal cycle time and the green split ratio of each approach. If there are common phases, , the cycle calculation formula is:​​​ wherein, is the total loss time in a cycle, is the green-to-signal ratio of the phase, satisfying: wherein, is the length of the green light for the first phase; If the intersection approach of the first phase has approaches, then the calculated green time set of the approaches of the first phase is . : Step 4.3, scheme evaluation and iterative optimization driven by value network: Evaluate the performance of the phase structure scheme generated by the large language model through the value network, and update the prompt words based on the evaluation results to guide the large language model to iteratively generate a better performance scheme. This process includes: Step 4.3.1, scheme evaluation: Input the phase structure scheme, green ratio, and current intersection topology characteristics and traffic flow calculated in step 4.2 into the value network model trained in step 3 to obtain the performance evaluation results of the scheme in the current traffic scenario; Let the current solution be , the value network outputs its comprehensive evaluation index as: wherein, is the current traffic flow, is the current phase structure, is the trained neural network model, is the network parameter; Step 4.3.2, optimality judgment: compare the current solution score to the historical best result update the best solution if the current solution score is better than the known best solution : Step 4.3.3, feedback guidance and prompt word update: If the current solution does not reach the preset optimal evaluation or still has obvious optimization space, feedback information containing the structure and performance score of the current solution is constructed , update the prompt word of the large language model : wherein the function Combine traffic data, background information and performance feedback to construct new optimization hints to guide the large language model to generate the next round of solutions ; Step 4.3.4, iterative optimization process: Step 4.3.1 - Step 4.3.3 are iteratively performed, each round generating new candidate solutions by the large language model and independently scoring the performance by the value network; the whole process continues until the solution score meets the set optimal evaluation target or reaches the maximum number of iterations is terminated; finally, the phase structure solution evaluated as optimal by the value network is output as the final traffic signal control configuration solution.

Images