Global traffic signal control method based on regional context enhanced large language model

Abstract

The invention discloses a global traffic signal control method based on a regional context enhanced large language model, and the method comprises the steps: dividing a target region into K traffic signal lamp control regions according to a traffic network structure of the target region; global real-time traffic information of the target area is collected and converted into global traffic text information; and inputting the global traffic text information into the large language model after LoRA fine tuning, and outputting a global traffic signal lamp control strategy of the target area, namely a traffic signal lamp phase of each intersection. In the fine tuning process of the large language model, an optimal phase decision is generated by using a reinforcement learning agent, a decision result is converted into text data, a region-level optimal phase decision data set is constructed according to an adjacent region division strategy, fine tuning is performed on the model by using the data set, region loss is calculated, and the large language model is optimized through the loss. According to the invention, optimization of global traffic signal lamp control can be realized, the overall traffic delay is reduced, and the traffic efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent traffic signal control, and in particular to a global traffic signal control method based on a regional context enhanced large language model. Background Art

[0002] With the acceleration of urbanization and the surge in the number of motor vehicles, traffic congestion has become a major issue affecting travel efficiency and quality of life. Traffic signal control manages road traffic flow through traffic lights. It uses traffic lights at intersections and other key areas to instruct vehicles to stop, move, turn, and change lanes. Intelligent traffic signal control effectively alleviates congestion, improves road traffic efficiency, and reduces fuel consumption and exhaust emissions by adjusting traffic lights in real time, thereby improving the urban environment and residents' quality of life. Existing traffic signal control technologies can be mainly divided into three categories: rule-based traffic signal control methods, reinforcement learning-based traffic signal control methods, and large language model-based traffic signal control methods.

[0003] Rule-based traffic signal control methods primarily include fixed-time control, sensor-based control, adaptive control, and optimization-based control. These methods rely on preset rules and heuristic algorithms to adjust the signal states at multiple intersections. For example, fixed-time control, based on historical traffic flow statistics, predefines the signal cycle, phase duration, and phase sequence, making it the most widely deployed control method in practice. However, rule-based control methods have numerous limitations. They lack real-time adaptability, making them difficult to adapt to complex and changing traffic flows and effectively address sudden congestion and abnormal events. Furthermore, they are often based on local optimization, making it difficult to achieve efficient coordination across multiple intersections, limiting the optimization capabilities of the overall traffic network. To overcome these limitations, multi-agent reinforcement learning (MARL) has been widely used in the field of traffic signal control. These methods model signal control as a Markov decision process (MDP) and learn optimal control strategies through the interaction of multiple agents with the environment. However, existing MARL approaches typically employ a distributed architecture, where a single agent controls only one intersection. Intersection collaboration is constrained by the topology of the road network, making it difficult for the system to develop a global understanding of the overall traffic environment. Furthermore, the difficulty in interpreting the decision-making process of reinforcement learning models limits their applicability to practical traffic management.

[0004] With the widespread application of large language models (LLMs) in specialized verticals, a growing number of researchers are exploring their potential in traffic signal control. LLMs are a neural network-based natural language processing technology. Trained on massive amounts of text data, LLMs can learn and predict patterns and regularities in natural language text. They can not only generate natural language text but also deeply understand its meaning, enabling them to handle various natural language tasks such as text summarization, question-answering, and translation. The fundamental idea behind LLMs is to treat natural language text as a sequence of data. By taking this input and applying it through computation and transformation across multiple layers of neurons, LLMs generate corresponding output sequences.

[0005] Currently, the main approaches to applying LLM to intelligent traffic signal control include using LLM to directly control signals at a single intersection, enabling it to achieve more human-like decision-making capabilities, or using LLM to assist reinforcement learning models in decision optimization. However, how to fully utilize the reasoning capabilities of LLM to achieve coordinated optimization of signals across multiple intersections remains an urgent problem.

[0006] In summary, the current multi-intersection traffic signal control has the following problems:

[0007] (1) Lack of global coordination capabilities. Most existing methods adopt a distributed structure, and the decision-making of the intelligent agent is usually based only on the information of adjacent or local intersections. It is difficult to achieve efficient coordination between multiple intersections. Moreover, the communication process is limited by the road network topology and cannot optimize the overall traffic flow.

[0008] (2) Insufficient adaptability and generalization capabilities. Traditional methods rely on fixed rules and are difficult to adapt to sudden traffic conditions in real time. Existing multi-agent reinforcement learning (MARL) methods are usually trained in specific environments, and their performance degrades when generalized to different road network structures, limiting their practicality. Summary of the Invention

[0009] To solve the problems in the prior art, the present invention provides a global traffic signal control method based on a regional context-enhanced large language model, the method comprising:

[0010] S101: Divide the target area into K traffic signal light control areas according to the traffic road network structure of the target area.

[0011] S102: Collecting global real-time traffic information of the target area and converting it into global traffic text information.

[0012] S103: Input the global traffic text information into the fine-tuned large language model, and output the traffic phases of each traffic light in the target area.

[0013] S104: Converting the phases of the traffic lights in the target area to control the global traffic light state of the target area.

[0014] Furthermore, based on the horizontal and vertical main roads of the target area, the target area is divided into K traffic light control areas, each control area includes m adjacent intersections, and an intelligent agent is set in each control area.

[0015] Furthermore, the state space S of the agent is defined as the traffic information data of each intersection obtained by observation, and the traffic light control action space A of the agent is defined as the phase transition of controlling the traffic light.

[0016] Furthermore, each control area R k The traffic light control is expressed as:

[0017] ;

[0018] Where, Represents the kth control region R k The traffic light control action set, π (.) represents the traffic signal control strategy adopted by the agent, R k (t) represents the kth control area at time step t, a i (t) represents the action of a single intersection in the control area at time step t, i∈R means that the i-th intersection belongs to the control area R;

[0019] The global traffic light control strategy of the target area is expressed as:

[0020] ;

[0021] Where K represents the number of traffic light control areas divided in the target area, π k (.) indicates that the agent adopts the traffic signal control strategy.

[0022] Furthermore, the global real-time traffic information includes the average waiting time, vehicle queue length, average vehicle speed, and number of vehicles at each intersection.

[0023] Furthermore, the fine-tuning process of the large language model includes:

[0024] Obtain the traffic network structure of the target area, divide the target area into K traffic light control areas, and set an intelligent agent for each control area;

[0025] In the simulation environment, real-time traffic information data is collected at each intersection in the target area, including average waiting time, vehicle queue length, average vehicle speed, and number of vehicles;

[0026] Based on the collected real-time traffic information data of each intersection, each intelligent agent uses a pre-trained reinforcement learning model to generate the optimal phase decision for each intersection. The optimal phase decision is serialized to obtain the optimal phase decision sequence.

[0027] The real-time traffic information data and optimal phase decision sequence of each intersection are converted into a spatiotemporal semantic embedding representation, and the traffic context information of each intersection in the same control area is integrated to construct a regional optimal phase decision dataset;

[0028] The large language model is fine-tuned using the regional optimal phase decision dataset. The cross-entropy loss between the model output phase decision and the optimal decision in the dataset is calculated, and the large language model is optimized using this cross-entropy loss.

[0029] Beneficial effects of the present invention:

[0030] This invention combines the advantages of reinforcement learning (RL) and large language models (LLMs) to propose an intelligent, interpretable, and adaptable multi-intersection traffic signal control method. Through a road network partitioning module, traffic control areas are divided based on horizontal and vertical main road strategies. Traffic status data such as average waiting time, queue length, and vehicle flow at each intersection are acquired in real time, thereby establishing a global traffic perception capability. Compared to traditional methods that focus only on a single intersection or local area, this invention enables cross-intersection collaborative optimization, improving overall traffic efficiency in complex urban road networks.

[0031] This method leverages a pre-trained reinforcement learning model to generate optimal control policies for individual intersections and constructs a regional-level optimal decision dataset to accommodate large-scale dynamic traffic patterns. It employs spatiotemporal semantic embedding to transform traffic data into a context parsable by the Large Language Model (LLM). Combined with LoRa low-rank fine-tuning, the Large Language Model (LLM) infers global signal control policies based on real-time data and outputs multi-step decision trajectories.

[0032] The application of this invention can effectively improve global traffic optimization capabilities, reduce overall traffic delays, improve traffic efficiency, and enhance the adaptability and interpretability of signal control; it can provide key technical support for traffic management in future smart cities. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flowchart of steps in an embodiment of the present invention;

[0034] Figure 2 Schematic diagram of the decision-making process of traffic lights at multiple intersections in an embodiment of the present invention;

[0035] Figure 3 This is an example of a prompt word for fine-tuning a large prediction model using LoRA in an embodiment of the present invention;

[0036] Figure 4 Schematic diagram of the process of using reinforcement learning to assist data collection and fine-tune the LLM model in an embodiment of the present invention;

[0037] Figure 5 This is a comparison table of test performances obtained by testing the present invention and various existing models on a real road network in Jinan;

[0038] Figure 6 The following is a comparison table of test performances of the present invention and existing models obtained by testing on a real road network in Hangzhou. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] It should be noted that, in this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements.

[0041] The embodiment of the present invention proposes a global traffic signal control method based on regional context enhanced large language model. Figure 1 As shown, the method includes:

[0042] S101: Divide the target area into K traffic signal light control areas according to the traffic road network structure of the target area.

[0043] S102: Collecting global real-time traffic information of the target area and converting it into global traffic text information.

[0044] S103: Input the global traffic text information into the fine-tuned large language model, and output the traffic phases of each traffic light in the target area.

[0045] S104: Converting the phases of the traffic lights in the target area to control the global traffic light state of the target area.

[0046] Specifically, in S101, the target area can be an urban area of ​​any size, such as a township, a city, or a district. In open-source road network databases, such as AutoNavi, Tencent Maps, and Baidu Maps, the public can query road network information, including intersection settings and traffic light settings.

[0047] In existing technologies, the method of using reinforcement learning (RL) and large language models (LLM) to control traffic lights usually involves setting up an intelligent agent at a single intersection, collecting traffic information in real time and converting it into traffic information text, and then using LLM to perform inference and output control information to control the traffic lights at a single intersection.

[0048] Figure 2 Schematic diagram of the decision-making process of traffic lights at multiple intersections in an embodiment of the present invention. Figure 2 In this example, a traffic network for the target area is constructed, divided according to horizontal and vertical main roads. Adjacent intersections are connected into control areas, such as control areas 1-4. Each control area is populated with an agent, which observes the state (i.e., traffic information) and performs actions (i.e., controlling the phase transitions of traffic lights). Each agent converts the observed state (i.e., the traffic status information for that control area) into traffic information text and inputs it into a large language model (LLM), such as GPT or Llama. The LLM can deeply understand the meaning of the text, process various natural language tasks, perform reasoning, and output global traffic light state configuration information. Based on the traffic light state configuration information output by the LLM, the agent controls the traffic light phase transitions. For example, intersection 1-1 selects the east-west left turn phase, and intersection 1-2 selects the east-west through phase.

[0049] Due to the large number of intersections in cities, especially large and medium-sized ones, directly inputting textual data for the entire road network into the LLM is limited by the input length, which restricts the model's reasoning capabilities. To address this issue, the present invention partitions the road network based on a proximity strategy, since geographically close intersections typically exhibit similar traffic flow characteristics. The control range of a single agent can be dynamically adjusted to accommodate areas with varying numbers of intersections.

[0050] In the illustrated embodiment, reference is made to Figure 2As shown in the figure, based on the target area's transverse and longitudinal trunk roads, the target area is divided into K traffic signal control zones. Each control zone includes m adjacent intersections, and an agent is assigned to each control zone. Thus, the target area is globally configured with K agents, each of which is a general-purpose agent. For example, m is the number of intersections, with a value varying depending on the zone (e.g., if one agent controls three intersections, m = 3). K is the number of zones in the road network (for example, if the Jinan road network is divided into four zones, K = 4).

[0051] The state space S of the agent is defined as the traffic information data of each intersection obtained by observation. The state observed at the i-th intersection at time t is expressed as .

[0052] The signal light control action space A of the agent is defined to control the phase transition of the traffic light. The action performed at the i-th intersection at time t is expressed as For example, "North-South Straight (NTST)", "North-South Left Turn (NLSL)", "East-West Left Turn (ELWL)", and "East-West Straight (ETWT)".

[0053] Definition: kth control region R k The traffic light control is expressed as:

[0054] ;

[0055] Where, Represents the kth control region R k The traffic light control action set, π (.) represents the traffic signal control strategy adopted by the agent, R k (t) represents the set of all intersections in the kth control area at time step t, a i (t) represents the action of a single intersection in the control area at time step t, and i∈R means that the i-th intersection i belongs to area R.

[0056] Then, the global traffic light control strategy of the target area is expressed as:

[0057] ;

[0058] Where A k (t) represents the kth control region R at time step t k Traffic light control action, K represents the number of traffic light control areas divided in the target area, π k (.) indicates that the agent adopts the traffic signal control strategy.

[0059] Compared with the existing technology, step S101 (i.e., from the control strategy of each control area to the global control strategy) can better obtain global traffic context information, consider the relationship between intersections, output global traffic signal decisions, and significantly reduce the number of required traffic signal control agents.

[0060] In the illustrated embodiment, the global real-time traffic information includes the average waiting time at each intersection, the length of a vehicle queue, the average speed of vehicles, and the number of vehicles.

[0061] Specifically, in S102, traffic information is collected in real time at each intersection, including average waiting time, queue length, average speed, and number of vehicles. The agent corresponding to each control area observes the current state, including traffic information at each intersection within the control area, such as average waiting time, queue length, average speed, and number of vehicles.

[0062] Reference Figure 2 As shown, in S102 , the global real-time traffic information is converted into global traffic text information (ie, the collected data information is converted into text information).

[0063] Specifically, the state and action pairs observed by each agent at the i-th time step are Convert to text description:

[0064] ;

[0065] Where, (.) indicates converting traffic status data information into traffic text information.

[0066] Reference Figure 2 As shown, the example:

[0067] At intersection 1_1: 6 vehicles are on sections 3 and 4 of the east-west through lanes. Considering that releasing the east-west through lanes will cause congestion at intersection 1_2, we choose to release the east-west left-turning vehicles.

[0068] For intersection 1_2: east-west through lane, 4 vehicles are on section 4, select the east-west through phase.

[0069] Through the regional division strategy - taking the horizontal trunk road or the longitudinal trunk road as a region, constructing the control area Prompt word and answer text:

[0070] ;

[0071] Where, represents the control area including the i-th intersection at time step t, and N represents the number of intersections in the control area.

[0072] The prompt words and answer texts are related to traffic scene description information, traffic light control task description information, traffic priority information and traffic information in the current control area.

[0073] Traffic scene description information can be expressed as follows: "Traffic lights control an intersection with four directions: east, south, west and north." Because traffic scenes such as traffic lights and lanes at different intersections vary in actual applications, it is necessary to describe the intersection based on the actual traffic scene to obtain traffic scene description information that can be understood by the large language model.

[0074] The control task description information of the traffic light can be expressed in the following text: "Which is the next optimal traffic light configuration?", "Which is the next optimal lane to be released?", and "What is the direction of traffic allowed next?", etc. The purpose of these control task description information is to enable the large language model to understand what tasks should be completed, that is, to realize the control of traffic lights, which can be described in the form of questions. The specific tasks to be completed need to be determined according to the actual traffic conditions, which can be to determine the next lane to be released and the next traffic light control action, etc. In actual applications, other expressions or other tasks can also be used, and no specific requirements are made here. The corresponding fine-tuning prompt words can be referred to. Figure 3 .

[0075] Traffic priority information primarily reflects common sense in the traffic field, such as "lanes with more queued vehicles should be given priority" and "vehicles farther from the intersection should be given less attention." In practice, other representations may be used. Actual intersections may have different traffic priority information, which can be added as needed.

[0076] The phase transition of traffic lights is the specific control action of each intelligent agent, such as "North-South Straight (NTST)", "North-South Left Turn (NLSL)", "East-West Left Turn (ELWL)", and "East-West Straight (ETWT)". Other phase settings can also be used in actual applications. The information represented by the above text controls the corresponding color of each traffic light.

[0077] Reference Figure 4 As shown, the present invention treats the decision-making process for multiple intersections within a target area as a sequence generation problem. In S103, the global traffic text information is input into a fine-tuned large language model. By leveraging the contextual understanding and reasoning capabilities of the LLM, the potential spatiotemporal dependencies between intersections are captured to enhance decision-making (i.e., the actions selected by the agent).

[0078] The relationship between the target intersection and other intersections in the area is expressed as follows through the large language model LLM reasoning:

[0079] ;

[0080] Where P(.) represents the probability of the target intersection taking action after considering the traffic conditions of surrounding intersections, ( ) represents the states of all intersections observed by the agent in the control area where the target intersection is located, (.) represents the traffic signal control strategy learned by large language model (LLM) contextual reasoning.

[0081] For the kth control region R k ,LLM generates control actions and relational reasoning trajectories at multiple intersections , which is expressed as:

[0082] ;

[0083] Where, ( ) represents the actions of all intersections in the control area, Represents the inference trajectory generated by the large language model, prompt(.) is the prompt input to the large language model, which includes the traffic status in the area, (.) represents the traffic signal control strategy, Indicates the parameters of LLM.

[0084] For example, the LLM model may be GPT, Llama, etc.

[0085] Finally, LLM outputs the global traffic light status configuration information of the target area.

[0086] For example, refer to Figure 2 As shown, for intersection 1_1: select the east-west left turn phase; for intersection 1_2: select the east-west straight phase, etc.

[0087] Figure 4 Schematic diagram of the process of using reinforcement learning to assist data collection and fine-tune the LLM model in an embodiment of the present invention. Figure 4In this paper, the road network structure is obtained from an open-source road traffic network database. In a simulation environment, an intelligent agent is deployed at each intersection to observe the state and perform actions. Feature extraction is performed on the observed traffic state information, including evaluated waiting time, vehicle queue length, evaluated vehicle speed, and vehicle number. Using reinforcement learning, the agent generates the optimal phase decision (i.e., action) for a single intersection in the simulation environment. The optimal phase decision data for each intersection is serialized, and the traffic state data and phase decision sequence for each intersection are converted into a parsable spatiotemporal semantic embedding. By integrating traffic context information from all intersections within the same control area, a regional optimal phase decision dataset is constructed, providing rich and high-quality fine-tuning data for the LLM learning model, thereby enhancing its understanding and generalization capabilities across diverse traffic scenarios. The regional optimal phase decision dataset is used to fine-tune the LLM model. During the fine-tuning process, the loss function is optimized, enabling the LLM to infer a global signal control policy based on real-time traffic conditions.

[0088] In the illustrated embodiment, reference is made to Figure 4 As shown, the fine-tuning process of the large language model (LLM) includes:

[0089] S201: Obtain the traffic network structure of the target area, divide the target area into K traffic light control areas, and set an intelligent agent for each control area.

[0090] S202: In the simulation environment, real-time traffic information data of each intersection in the target area is collected, including average waiting time, vehicle queue length, average vehicle speed, and number of vehicles.

[0091] S203: Each intelligent agent uses a pre-trained reinforcement learning model based on the collected real-time traffic information data of each intersection to generate the optimal phase decision for each intersection, and serializes the optimal phase decision to obtain the optimal phase decision sequence.

[0092] S204: Convert the real-time traffic information data of each intersection and the optimal phase decision sequence into a spatiotemporal semantic embedding representation, integrate the traffic context information of each intersection in the same control area, and construct a regional optimal phase decision dataset.

[0093] A pre-trained reinforcement learning model is used to generate high-quality traffic signal control data, activating the Large Language Model (LLM)'s understanding and decision-making capabilities in traffic signal control tasks. Combined with the LLM's powerful contextual modeling capabilities, it can explore potential spatiotemporal interactions between intersections. Furthermore, leveraging the LLM's zero-shot reasoning capabilities, it enhances its understanding and generalization capabilities for diverse traffic scenarios, enabling more intelligent and robust traffic signal control strategies.

[0094] S205: Using the regional optimal phase decision dataset, the large language model is fine-tuned using the LoRA low-rank fine-tuning method. The cross-entropy loss between the model output phase decision and the optimal decision in the dataset is calculated, and the large language model is optimized using this loss.

[0095] Furthermore, the calculation formula of the loss function is:

[0096] ;

[0097] Where, represents the cross entropy loss function, represents the parameters of the large language model LLM, K represents the number of traffic light control areas divided in the target area, T represents the number of time steps, (.) represents the probability of LLM generating the corresponding decision text in a given area, Represents the text description of traffic phase decision, A text description of the traffic area status. During fine-tuning, by optimizing the loss function, the large language model (LLM) can infer global signal control strategies based on real-time traffic conditions, achieving coordinated optimization of multiple intersections and improving overall traffic flow.

[0098] Experimental verification:

[0099] Simulation software: Cityflow.

[0100] The real road network data of Hangzhou and Jinan were obtained through the existing platform (Openstreetmap), and simulation tests were conducted on the present invention and the existing model to obtain a test performance comparison table, as shown in the figure below. Figure 5 、 6 As shown in the figure, the bold ones are the best and the underlined ones are the second best.

[0101] like Figure 5 、 6 As shown in FIG, the experimental results show that the present invention significantly reduces the average travel time and the average queue length, reflecting its effectiveness in global traffic signal control.

[0102] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A global traffic signal control method based on a large language model enhanced by regional context, characterized by: include: According to the traffic network structure of the target area, the target area is divided into K traffic signal light control areas; Collect global real-time traffic information of the target area and convert it into global traffic text information; Inputting the global traffic text information into the fine-tuned large language model to output the traffic phases of each traffic light in the target area; The phase of each traffic light in the target area is converted to control the global traffic light state of the target area.

2. The global traffic signal control method based on regional context enhanced large language model according to claim 1 is characterized in that: Based on the horizontal and vertical main roads of the target area, the target area is divided into K traffic light control areas, each control area includes m adjacent intersections, and an intelligent agent is set in each control area.

3. The global traffic signal control method based on regional context enhanced large language model according to claim 2 is characterized in that: The state space S of the intelligent agent is defined as the traffic information data of each intersection obtained by observation, and the signal light control action space A of the intelligent agent is defined as the phase transition of the traffic light.

4. The global traffic signal control method based on regional context enhanced large language model according to claim 1 or 2, characterized in that: Each control area R k The traffic light control is expressed as: ； Where, Represents the kth control region R k The traffic light control action set, π (.) represents the traffic signal control strategy adopted by the agent, R k (t) represents the kth control area at time step t, a i (t) represents the action of a single intersection in the control area at time step t, i∈R means that the i-th intersection belongs to the control area R; The global traffic light control strategy of the target area is expressed as: ； Where A k (t) represents the kth control region R at time step t k Traffic light control action, K represents the number of traffic light control areas divided in the target area, π k (.) indicates that the agent adopts the traffic signal control strategy.

5. The global traffic signal control method based on regional context enhanced large language model according to claim 1 is characterized in that: The global real-time traffic information includes the average waiting time at each intersection, the length of the vehicle queue, the average vehicle speed, and the number of vehicles.

6. The global traffic signal control method based on regional context enhanced large language model according to claim 1 is characterized in that: The fine-tuning process of the large language model includes: Obtain the traffic network structure of the target area, divide the target area into K traffic light control areas, and set an intelligent agent for each control area; In the simulation environment, real-time traffic information data is collected at each intersection in the target area, including average waiting time, vehicle queue length, average vehicle speed, and number of vehicles; Based on the collected real-time traffic information data of each intersection, each intelligent agent uses a pre-trained reinforcement learning model to generate the optimal phase decision for each intersection. The optimal phase decision is serialized to obtain the optimal phase decision sequence. The real-time traffic information data and optimal phase decision sequence of each intersection are converted into a spatiotemporal semantic embedding representation, and the traffic context information of each intersection in the same control area is integrated to construct a regional optimal phase decision dataset; The large language model is fine-tuned using the regional optimal phase decision dataset. The cross-entropy loss between the model output phase decision and the optimal decision in the dataset is calculated, and the large language model is optimized using this cross-entropy loss.

7. The global traffic signal control method based on regional context enhanced large language model according to claim 6 is characterized in that: During the fine-tuning of the large language model, the loss function used is: ； in, represents the cross entropy loss function, represents the parameters of the large language model LLM, K represents the number of traffic light control areas divided in the target area, T represents the number of time steps, (.) represents the probability of LLM generating the corresponding decision text in a given area, Represents the text description of traffic phase decision, Indicates a text description of the traffic area status.

Images