A traffic signal control optimization method and system based on a large model agent

Abstract

The application discloses a traffic signal control optimization method and system based on a large model intelligent agent, and belongs to the technical field of intelligent traffic control. In view of the problems of low intelligent degree, unexplainable AI decision black box, fragmented tool chain, insufficient use of domain knowledge and poor scene adaptability of the existing traffic signal control optimization method, the application proposes a large language model as the reasoning core, combines retrieval enhancement generation technology and programmable tool chain, adopts a seven-stage workflow programming mode (data collection -> performance evaluation -> problem diagnosis -> control mode decision -> strategy generation -> scheme evaluation -> scheme output) to schedule tools to complete the whole process of signal control optimization, and realizes the whole process transparency of the reasoning process through four mechanisms of thought chain recording, tool calling log, knowledge reference traceability and decision basis chain. The application shortens the traditional manual signal control optimization cycle of several days to minutes, realizes the whole process auditability of the optimization scheme, reduces the professional operation threshold, and can adapt to various application scenarios of sensing control, adaptive control and event response control.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation systems, specifically relating to a traffic signal control optimization method and system based on a large language model intelligent agent. Background Technology

[0002] Traffic signal control is a core component of urban traffic management, and the quality of signal timing schemes directly affects road efficiency, traffic congestion levels, and driving safety. With the acceleration of urbanization, traffic signal control optimization technology has evolved from fixed timing to intelligent optimization.

[0003] Traditional traffic signal control systems primarily rely on fixed timing schemes or adaptive control based on simple induction coils (such as SCOOT and SCATS systems), using preset rule engines to adjust signal timing. These schemes suffer from poor flexibility, are prone to causing empty intersections and excessively long queues, and cannot adapt to dynamic changes in traffic flow.

[0004] In recent years, reinforcement learning methods have been widely applied in the field of traffic signal control. For example, Chinese patent CN117809469A (applicant: Hefei University of Technology, publication date: April 2, 2024) discloses a traffic signal timing control method based on deep reinforcement learning, which uses a double-Q learning method to optimize signal timing through experience playback and a target Q-value network. Chinese patent CN112700664A (applicant: Beijing University of Technology, publication date: April 23, 2021) discloses a traffic signal timing optimization method based on deep reinforcement learning, which optimizes timing by dynamically adjusting the signal phase and phase length. However, the above-mentioned reinforcement learning-based methods are all black-box decision-making methods. The decision-making process of the intelligent agent is uninterpretable, and traffic engineers cannot understand or audit the optimization basis, making it difficult to establish trust in practical engineering.

[0005] With the development of large language model technology, large models have begun to be introduced into the field of traffic control. For example, Chinese patent CN120954237A (applicant: Zhejiang Zhongkong Information Industry Co., Ltd., publication date: November 14, 2025) discloses a vehicle-road cooperative signal control method based on a large model agent, which constructs a V2X cooperative architecture of perception agent, decision agent, execution agent and communication agent, and realizes real-time generation of signal phase schemes by combining a knowledge graph library. However, this scheme relies on V2X vehicle-road cooperative hardware infrastructure and is geared towards millisecond-level real-time online control scenarios. Its reasoning process still lacks a complete transparency mechanism, and traffic engineers cannot trace the complete reasoning chain of optimization decisions.

[0006] Furthermore, existing traffic signal control optimization systems generally suffer from a fragmented toolchain: traffic data collection, performance index calculation, problem analysis, and timing scheme generation are scattered across different software systems, lacking a unified intelligent scheduling and coordination mechanism. At the same time, a wealth of professional knowledge accumulated in the field of traffic engineering (national standards, industry specifications, historical optimization cases, traffic flow theory) has not been effectively integrated and utilized.

[0007] In summary, existing technologies still have room for improvement in terms of intelligence level, decision interpretability, toolchain collaboration, knowledge utilization, and scenario adaptability, and require further optimization. Summary of the Invention

[0008] In view of the above-mentioned problems in the prior art, the technical problem to be solved by the present invention is to provide a traffic signal control optimization method and system based on a large model intelligent agent, so as to realize the automation, intelligence and full transparency of signal control optimization.

[0009] To address the aforementioned technical problems, this invention provides a traffic signal control optimization method based on a large-scale intelligent agent model, comprising the following steps: Step S10: Receive a traffic signal control optimization request input by the user, the request including a natural language description or structured parameters.

[0010] The large language model serves as the core of inference, parsing user optimization requests through natural language understanding. Users can describe their optimization needs in natural language (e.g., "Analyze the causes of morning rush hour congestion at intersection XX and optimize signal timing"), or input structured parameters (intersection number, time period, optimization target, etc.).

[0011] Step S20: The large language model understands the request, decomposes the request into a sequence of subtasks, and generates a tool call plan according to a predefined workflow orchestration.

[0012] Based on its reasoning capabilities, the large language model decomposes optimization requests into executable subtasks and plans the tool call sequence according to the professional process of credit control optimization. Workflow orchestration defines a standardized credit control optimization process framework and supports dynamic adjustment of subsequent tool call strategies based on intermediate results.

[0013] Step S30: Retrieve professional knowledge related to the subtask from the traffic control domain knowledge base using retrieval enhancement generation technology, and inject the retrieved knowledge as context into the reasoning process of the large language model.

[0014] The traffic control knowledge base includes the following: traffic engineering standards and specifications (such as GB 14886-2016 "Specifications for the Installation and Setting of Road Traffic Signals"), signal timing manuals and guidelines, historical optimization case database (including comparative data before and after optimization), traffic flow theory literature, and intersection feature knowledge graphs.

[0015] Furthermore, the retrieval enhancement generation technology employs a hybrid retrieval strategy combining semantic retrieval and keyword matching. Semantic retrieval, based on vector similarity calculation, captures the semantic connections between queries and knowledge; keyword matching, based on precise term matching, ensures accurate retrieval of specialized terms and standard numbers. The combined relevance score of the two retrieval strategies is calculated using the following formula:

[0016] in, This indicates the overall relevance score. This represents the similarity score for semantic retrieval. This indicates the relevance score of keyword matching. This represents the semantic retrieval weight coefficient. This represents the keyword matching weight coefficient. and satisfy . The value range is from 0.5 to 0.9. The value range is from 0.1 to 0.5.

[0017] Step S40: The tool is scheduled to execute the following seven stages in sequence according to the workflow orchestration: Step S41: Use the data acquisition tool to obtain traffic flow data and current timing scheme for the target intersection.

[0018] The data acquisition tool obtains real-time or historical traffic flow data from the traffic detector interface, including data such as flow rate, speed, occupancy, and queue length of each approach lane, as well as parameters such as the phase structure, green light duration, and cycle length of the current signal timing scheme.

[0019] Step S42: Call the indicator calculation tool to calculate the signal control performance index of the target intersection.

[0020] Performance metrics include average intersection delay, level of service (LOS) for each approach lane, saturation for each phase, traffic efficiency, green light ratio, and cycle time loss. The metric calculation tool transforms the detector's raw data into a standardized set of performance metrics that can be used for problem diagnosis.

[0021] Step S43: Invoke the problem diagnosis tool to identify the signal control problem of the target intersection based on the performance indicators.

[0022] The problem diagnosis tool compares the calculated performance indicators with preset standard thresholds and, in conjunction with relevant standard clauses retrieved from the knowledge base, identifies the type and severity of traffic control problems. Identifiable problem types include oversaturation, wasted green time, phase conflicts, poor coordination, and insufficient pedestrian safety.

[0023] Step S44: Make a control mode decision based on the type of signal control problem and determine the corresponding control mode.

[0024] Furthermore, the specific process for determining the control mode is as follows: when the signal control problem is a routine problem with unreasonable cycles or an imbalanced green light ratio, the control mode is determined to be the inductive control mode; when the signal control problem is a complex traffic flow problem with multi-period traffic flow fluctuations or significant changes in demand, the control mode is determined to be the adaptive control mode; when the signal control problem is a sudden event caused by intersection overflow, traffic accidents, or large-scale events, the control mode is determined to be the event response control mode; when the signal control problem includes multiple types, the control mode is determined to be the combined control mode.

[0025] Furthermore, the control mode decision is based on the following judgment conditions: when the maximum saturation value of each approach lane at the intersection reaches a preset saturation threshold, an event response control mode is triggered, the saturation threshold being in the range of 0.80 to 0.95; when the queue length exceeds a preset queue length threshold, an event response control mode is triggered, the queue length threshold being in the range of 100 meters to 200 meters; when the coefficient of variation of traffic flow within the statistical period exceeds a preset coefficient of variation threshold, the coefficient of variation threshold being in the range of 0.2 to 0.4; when none of the above conditions are triggered, an inductive control mode is adopted.

[0026] Step S45: Based on the control mode, call the timing calculation tool to generate candidate optimization strategies and corresponding timing parameters.

[0027] Different timing calculation methods are employed for different control modes. Under inductive control mode, the optimal cycle and green ratio are calculated based on the Webster method; under adaptive control mode, dynamic timing parameters are calculated based on multi-period flow characteristics; under event-response control mode, timing schemes incorporating special strategies such as overflow mitigation and emergency priority are generated. At least two candidate optimization strategies are generated for each control mode for subsequent evaluation.

[0028] Step S46: Call the simulation verification tool to perform simulation verification on the candidate optimization strategy, and select the optimal solution based on the verification results.

[0029] The simulation verification tool calls the microscopic traffic simulation engine to simulate each candidate optimization strategy, calculates the key performance indicators (average delay, queue length, traffic efficiency, etc.) after optimization, and compares them with the baseline data before optimization to obtain the improvement rate of each candidate solution.

[0030] Furthermore, if the verification results show that the improvement rate of the key performance indicators of the candidate solution is lower than the preset convergence threshold, the process automatically reverts to step S45 to regenerate the candidate optimization strategy and repeats the simulation verification process until the improvement rate reaches the convergence threshold or the number of iterations reaches the preset maximum number of iterations. The convergence threshold ranges from 3% to 10%, and the maximum number of iterations ranges from 2 to 5.

[0031] Step S47: Use the scheme generation tool to generate a standardized signal timing scheme.

[0032] The generated standardized signal timing scheme includes: timing parameter table (green light duration, yellow light duration, all-red time and cycle length for each phase), phase diagram (signal light status and release direction for each phase), time distance diagram (signal timing relationship of each intersection during multi-intersection coordinated control), and expected effect evaluation (comparison data of key performance indicators before and after optimization).

[0033] Step S50: Record the complete reasoning process from steps S10 to S47, generate a transparent reasoning report including thought chain records, tool call logs, knowledge reference tracing, and decision basis chains, and output it together with the standardized signal timing scheme.

[0034] Furthermore, the transparency of the reasoning process includes four mechanisms: Thought chain recording: Generates natural language descriptions for each reasoning step, recording the reasons and logic behind the reasoning; Tool call log: Records the input parameters, output results, and triggering reason for each tool call; Knowledge citation tracing: Mark the sources of knowledge cited in the reasoning process, including standard clause numbers, historical case numbers, and theoretical literature sources; Decision-making basis chain: A complete causal reasoning chain is formed from problem diagnosis to solution recommendation, recording the basis and reasoning path of each decision node.

[0035] Furthermore, the workflow orchestration supports a dynamic adjustment mechanism, including: conditional branching, selecting different strategies to generate paths in step S45 based on the diagnostic result type of step S43; iterative optimization, determining whether to revert to step S45 based on the verification result of step S46; anomaly handling, triggering supplementary data collection or using historical data to complete the data when missing data is detected; and manual intervention, pausing at the scheme selection node in step S46 and waiting for manual confirmation before resuming execution.

[0036] This invention also provides a traffic signal control optimization system based on a large model intelligent agent, comprising: The interaction and input module is used to receive traffic signal control optimization requests from users; The intelligent agent reasoning module understands the request based on a large language model, decomposes it into sub-tasks, and generates a tool call plan for workflow orchestration. The knowledge retrieval module retrieves relevant professional knowledge from the traffic control knowledge base and injects it into the reasoning process through retrieval-enhanced generation technology. The tool execution module sequentially schedules the data acquisition tool, indicator calculation tool, problem diagnosis tool, timing calculation tool, simulation verification tool, and scheme generation tool according to the workflow arrangement to execute the seven-stage signal control optimization task; The transparent recording module records the complete reasoning process and generates a transparent reasoning report that includes thought chain records, tool call logs, knowledge reference tracing, and decision basis chains.

[0037] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0038] The present invention has the following beneficial effects: (1) Improve optimization efficiency: Through the autonomous reasoning and workflow orchestration of the large language model agent, the traditional signal control optimization process that relies on human experience and takes several days is shortened to minutes of automated completion, which significantly improves the efficiency of signal control optimization.

[0039] (2) Enhance the interpretability of the solution: Through the four-fold transparency mechanism of thought chain recording, tool call log, knowledge reference tracing and decision basis chain, the complete reasoning process of traffic control optimization is made explicit, and traffic engineers can understand and audit each step of optimization decision, which solves the problem of black box decision-making in the existing AI traffic control system.

[0040] (3) Improve optimization quality: By using retrieval enhancement generation technology, professional knowledge in the field of transportation engineering (national standards, historical cases, theoretical literature) is dynamically injected into the reasoning process to ensure that the optimization scheme conforms to engineering specifications and avoids non-compliant schemes generated by pure data-driven methods.

[0041] (4) Lowering the professional threshold: Through the natural language interaction interface, non-signal control professionals can also initiate signal control optimization tasks, expanding the applicable population of signal control optimization tools.

[0042] (5) High adaptability: Through the control mode decision-making mechanism and dynamic arrangement mechanism, it can automatically match the inductive control, adaptive control, event response control or combined control mode according to different traffic scenarios and problem types, and adapt to various application scenarios such as single intersection optimization, trunk line coordination and regional optimization. Attached Figure Description

[0043] Figure 1 This is a flowchart of the traffic signal control optimization method based on a large model intelligent agent according to the present invention; Figure 2 This is a schematic diagram of the architecture of the traffic signal control optimization system based on a large model intelligent agent according to the present invention; Figure 3 This is a detailed flowchart of the control mode decision-making stage of the present invention; Figure 4 This is a schematic diagram of the output structure of the reasoning transparency mechanism of this invention; Figure 5 This is the interaction timing diagram between the intelligent agent inference layer and the tool execution layer of this invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0045] Example 1

[0046] like Figure 1 and Figure 2 As shown, this embodiment provides a traffic signal control optimization method based on a large-model intelligent agent, applied to a traffic signal timing optimization scenario at an intersection during the morning rush hour in a certain city. This method is executed by a traffic signal control optimization system based on a large-model intelligent agent, which includes an interaction and input layer 10, an intelligent agent reasoning layer 20, a knowledge enhancement layer 30, and a tool execution layer 40.

[0047] The interaction and input layer 10 includes a natural language interface module 11 and a parameter parsing module 12, which are used to receive user input.

[0048] The intelligent agent reasoning layer 20 includes a task understanding and decomposition module 21, a workflow orchestration engine 22, and a thought chain recording module 23, with a large language model as the core of reasoning.

[0049] The knowledge enhancement layer 30 includes a knowledge base 31, a hybrid retrieval module 32, and a knowledge injection module 33, which realizes the retrieval enhancement generation function.

[0050] The tool execution layer 40 includes a data acquisition tool 41, an indicator calculation tool 42, a problem diagnosis tool 43, a timing calculation tool 44, a simulation verification tool 45, and a scheme generation tool 46.

[0051] The specific steps of the method are as follows: In step S10, the user inputs an optimization request through the natural language interface module 11 of the interaction and input layer 10: "Analyze the causes of congestion at the intersection of Renmin Road and Zhongshan Road during the morning rush hour (7:00-9:00) and optimize signal timing." The parameter parsing module 12 extracts key parameters from the natural language: intersection number "Renmin Road-Zhongshan Road", time period "7:00-9:00", and optimization goal "congestion relief".

[0052] In step S20, the task understanding and decomposition module 21 of the agent inference layer 20 understands the request based on a large language model, identifies it as a signal control optimization task, and decomposes the task into the following sub-task sequence: data acquisition → performance evaluation → problem diagnosis → control mode decision → strategy generation → solution evaluation → solution output. The workflow orchestration engine 22 invokes the plan according to the standardized seven-stage workflow generation tool.

[0053] In step S30, the hybrid retrieval module 32 of the knowledge enhancement layer 30 retrieves professional knowledge related to the intersection type (urban arterial road-arterial road intersection) and the optimization objective (morning rush hour congestion) from the knowledge base 31. The retrieval adopts the hybrid retrieval strategy of Formula 1, taking... , The search results include: relevant clauses on intersection signal timing in GB 14886-2016, three historical optimization cases of the city, and relevant literature on HCM delay calculation methods. The knowledge injection module 33 injects the retrieved knowledge as context into the reasoning process of the large language model.

[0054] Step S41: Data acquisition tool 41 obtains traffic flow data from the intersection of Renmin Road and Zhongshan Road between 7:00 and 9:00 from the traffic detector interface. Hourly traffic flow (vehicles / hour) for each entrance lane: East entrance: straight 856, left turn 312, right turn 198; West entrance: straight 792, left turn 287, right turn 176; South entrance: straight 623, left turn 245, right turn 153; North entrance: straight 598, left turn 218, right turn 142.

[0055] Current timing scheme: four-phase control, cycle 120 seconds, east-west straight green light 42 seconds, east-west left turn green light 18 seconds, north-south straight green light 30 seconds, north-south left turn green light 15 seconds, yellow light duration is 3 seconds, and all red light duration is 2 seconds.

[0056] Intersection geometry information: A standard four-way intersection with four lanes in and four lanes out. Each entrance lane includes one dedicated left-turn lane, two straight lanes, and one dedicated right-turn lane.

[0057] Step S42, the indicator calculation tool 42 calculates each performance indicator: Average intersection delays: East entrance 58 seconds, West entrance 52 seconds, South entrance 45 seconds, North entrance 42 seconds, weighted average delay 50.3 seconds; Service Levels (LOS) of each approach lane: East approach LOS=E, West approach LOS=E, South approach LOS=D, North approach LOS=D; Saturation values ​​for each phase: East-West straight ahead 0.92, East-West left turn 0.78, North-South straight ahead 0.76, North-South left turn 0.65; Key conflict phase: East-west straight saturation is too high (0.92>0.85), indicating insufficient green time in this phase.

[0058] Step S43: Problem diagnosis tool 43 compares the performance indicators with standard thresholds and, in conjunction with relevant standard clauses retrieved from the knowledge base, diagnoses the following problems: Question 1: The east-west straight-line phase saturation reaches 0.92, exceeding the saturation threshold of 0.85, which is an oversaturated state, resulting in a severe lack of green time; Question 2: The queue length in the east-west straight direction has reached 168 meters, exceeding the queue length threshold of 150 meters, posing a risk of overflow. Question 3: The cycle of 120 seconds is too short and fails to fully utilize the intersection's traffic capacity; Question 4: The saturation of left turns from north to south is only 0.65, which results in a waste of green time. Green time can be allocated to the east-west direction.

[0059] Step S44: Make a control mode decision based on the diagnostic results of step S43 (e.g., Figure 3 (as shown) Judgment condition 1: The saturation of the east-west straight line is 0.92, which reaches the saturation threshold of 0.85 → satisfied; Judgment condition 2: The east-west queue length of 168 meters exceeds the queue length threshold of 150 meters → Satisfied; Since both the saturation and queue length exceed the thresholds, the scenario is identified as a sudden congestion with overflow risk, and the control mode is determined to be the intersection overflow mitigation sub-mode in the event response control mode.

[0060] Step S45, the timing calculation tool 44 generates two candidate optimization strategies based on the event response control mode (intersection overflow mitigation): Candidate Strategy A: Adjust the cycle to 140 seconds, reduce the green time for north-south left turns from 15 seconds to 10 seconds, allocate the saved green time to east-west straight traffic (increase by 5 seconds), and adjust the east-west straight traffic green light to 50 seconds. It is expected that the east-west straight traffic saturation will decrease to 0.78, and the average intersection delay will decrease to 42 seconds.

[0061] Candidate Strategy B: Adjust the cycle to 150 seconds, reduce the green time for left turns from 15 seconds to 10 seconds, and for straight traffic from 30 seconds to 26 seconds. Allocate the saved green time to straight traffic from east to west (increasing it by 9 seconds), and adjust the green time for straight traffic from east to west to 54 seconds. It is expected that the saturation of straight traffic from east to west will decrease to 0.72, and the average intersection delay will decrease to 39 seconds.

[0062] Step S46: Simulation verification tool 45 calls the microscopic traffic simulation engine to perform simulation verification on the two candidate strategies during a 2-hour morning rush hour. The simulation results are as follows: Candidate Strategy A: Average intersection delay of 42.8 seconds (a 15.1% reduction from before optimization), maximum east-west queue length of 105 meters (a 37.5% reduction from before optimization), simulation improvement rate of 15.1%, exceeding the convergence threshold of 5%; Candidate Strategy B: The average intersection delay is 38.5 seconds (a 23.5% reduction from before optimization), the maximum queue length in the east-west direction is 88 meters (a 47.6% reduction from before optimization), the simulation improvement rate is 23.5%, exceeding the convergence threshold of 5%.

[0063] Both candidate strategies met the convergence threshold of 5%, eliminating the need for iterative backoff. Based on simulation results, candidate strategy B showed a higher average delay improvement rate, and was therefore selected as the optimal solution.

[0064] Step S47, the scheme generation tool 46 generates a standardized signal timing scheme, including: Timing parameters: Cycle 150 seconds, east-west straight green light 54 seconds, east-west left turn green light 18 seconds, north-south straight green light 26 seconds, north-south left turn green light 10 seconds, yellow light time 3 seconds, all red time 2 seconds; Phase diagram: Displays the status of traffic lights and the direction of passage for the four phases; Time interval diagram: Shows the time distribution relationship of each phase on the timing time axis; Expected results assessment: Average delay decreased from 50.3 seconds to 38.5 seconds (23.5% improvement), and maximum queue length in the east-west direction decreased from 168 meters to 88 meters (47.6% improvement).

[0065] Step S50: The transparent recording module records the complete reasoning process and generates a transparent reasoning report (e.g., ...). Figure 4 (as shown), including: Mind Chain Record 51: Records the complete mind chain from "East-West straight saturation 0.92 exceeds threshold 0.85, judged as oversaturation" → "Queue length 168 meters exceeds threshold 150 meters, judged as overflow risk" → "Select event response control mode" → "Generate two candidate strategies and conduct simulation comparison" → "Select strategy B with higher improvement rate". Tool call log 52: Records the input parameters, output results, and triggering reasons for 5 tool calls (data acquisition → indicator calculation → problem diagnosis → timing calculation → simulation verification); Knowledge Citation Traceability 53: The sources of the cited knowledge are marked, including GB 14886-2016, Clause 6.3 (Intersection Signal Timing Standard), Historical Case No. Case-2024-015 (Overflow Optimization Case of Similar Intersections in this City), and HCM Chapter 6 (Delay Calculation Method). Decision-making chain 54: forms a complete causal reasoning chain of "data acquisition → performance evaluation (delay of 50.3 seconds, saturation of 0.92) → problem diagnosis (oversaturation + overflow risk) → pattern decision (event response) → strategy generation (two candidate solutions) → simulation verification (strategy B is 23.5% better) → solution output".

[0066] Finally, the standardized signal timing scheme and transparent inference report are output to the user.

[0067] Example 2

[0068] In another embodiment of the invention, it is applied to a scenario of coordinated control optimization on a city's main thoroughfare. The user inputs an optimization request: "Optimize the signal coordination at 5 intersections along XX Avenue to achieve bidirectional green wave."

[0069] After the system executes steps S10 to S43, the problem diagnosis results are as follows: the timing of each intersection is basically reasonable (saturation is less than 0.85), but the phase difference between adjacent intersections is not coordinated, causing vehicles to frequently encounter red lights between intersections, indicating poor coordination. The coefficient of variation of travel time between intersections is 0.35, exceeding the coefficient of variation threshold of 0.3.

[0070] Control mode decision (step S44): Since the traffic flow variation coefficient exceeds the threshold, it is judged to be a complex scenario with significant traffic flow fluctuations, and the control mode is determined to be the adaptive control mode.

[0071] Strategy Generation (Step S45): Timing calculation tool 44, based on adaptive control mode and common cycle, calculates the optimal phase difference for each intersection and generates a two-way green wave coordination scheme. The MAXBAND method is used to calculate the green wave bandwidth, generating two candidate strategies: Strategy A optimizes the green wave bandwidth primarily in the east-west direction, and Strategy B performs balanced optimization in both directions.

[0072] Simulation verification (step S46): The simulation results show that the east-west green wave bandwidth of strategy B is 32 seconds and the north-south bandwidth is 28 seconds. The average number of stops is reduced from 5.2 times to 1.8 times (an improvement of 65.4%), which meets the convergence threshold requirement.

[0073] Scheme output (step S47): Generate coordinated timing schemes for 5 intersections, including common cycle, phase difference and time distance diagram for each intersection.

[0074] Example 3

[0075] In another embodiment of the invention, an iterative optimization rollback scenario is described. During the optimization process of a certain intersection, the simulation verification results of step S46 show that the improvement rate of the candidate strategy is 2.8%, which is lower than the convergence threshold of 5%. The system automatically rolls back to step S45, adjusts the strategy parameters (increases the green time for east-west straight traffic by 3 seconds and reduces the green time for north-south left turns by 2 seconds), regenerates the candidate strategy, and performs simulation verification. The second simulation results show an improvement rate of 6.2%, which meets the convergence threshold requirement, and the selected scheme is output. This iteration is executed twice, which does not exceed the maximum number of iterations of 3.

[0076] Example 4 (System Example) This embodiment provides a traffic signal control optimization system based on a large model intelligent agent (such as...). Figure 2 (as shown), including: The interaction and input layer 10 includes a natural language interface module 11 and a parameter parsing module 12, which are used to receive traffic signal control optimization requests input by users. The intelligent agent reasoning layer 20 includes a task understanding and decomposition module 21, a workflow orchestration engine 22, and a thought chain recording module 23, which understands the request based on a large language model, decomposes the sub-tasks, and generates a tool call plan for workflow orchestration. The knowledge enhancement layer 30 includes a knowledge base 31, a hybrid retrieval module 32, and a knowledge injection module 33. It retrieves professional knowledge related to the sub-task from the traffic control domain knowledge base through retrieval enhancement generation technology and injects it into the reasoning process. The tool execution layer 40 includes a data acquisition tool 41, an indicator calculation tool 42, a problem diagnosis tool 43, a timing calculation tool 44, a simulation verification tool 45, and a scheme generation tool 46, which execute the seven-stage signal control optimization task sequentially according to the workflow arrangement. The transparent recording module, integrated into the thought chain recording module 23, records the complete reasoning process.

[0077] The various modules of the system interact with each other through standardized interfaces. The large language model of the agent reasoning layer 20 coordinates the collaborative work of the knowledge enhancement layer 30 and the tool execution layer 40 through the workflow orchestration engine 22, forming a closed loop of perception-reasoning-execution.

[0078] Example 5 (Computer Example) This embodiment provides an electronic device, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the method described in Embodiment 1. The processor is connected to the memory, a communication interface, and a user interface via a bus. The user interface includes a display and input devices, and the communication interface is used for data communication with traffic detectors and signal controllers.

[0079] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1. The computer-readable storage medium includes at least one of flash memory, a portable hard disk, and an optical disc.

Claims

1. A traffic signal control optimization method based on a large-scale intelligent agent model, characterized in that, Includes the following steps: Step S10: Receive a traffic signal control optimization request input by the user, the request including a natural language description or structured parameters; Step S20: The large language model understands the request, decomposes the request into a sequence of sub-tasks, and generates a tool call plan according to a predefined workflow orchestration. Step S30: Retrieve professional knowledge related to the subtask from the traffic control domain knowledge base using retrieval enhancement generation technology, and inject the retrieved knowledge as context into the reasoning process of the large language model; Step S40: The tool is scheduled to execute the following seven stages in sequence according to the workflow orchestration: Step S41: Use the data acquisition tool to obtain traffic flow data and current timing scheme for the target intersection; Step S42: Call the indicator calculation tool to calculate the signal control performance index of the target intersection; Step S43: Invoke the problem diagnosis tool to identify the signal control problem of the target intersection based on the performance indicators; Step S44: Make a control mode decision based on the type of signal control problem and determine the corresponding control mode; Step S45: Based on the control mode, call the timing calculation tool to generate candidate optimization strategies and corresponding timing parameters; Step S46: Call the simulation verification tool to perform simulation verification on the candidate optimization strategy, and select the optimal solution based on the verification results; Step S47: Use the scheme generation tool to generate a standardized signal timing scheme; Step S50: Record the complete reasoning process from steps S10 to S47, generate a transparent reasoning report including thought chain records, tool call logs, knowledge reference tracing, and decision basis chains, and output it together with the standardized signal timing scheme.

2. The method according to claim 1, characterized in that, In step S30, the retrieval enhancement generation technology adopts a hybrid retrieval strategy of semantic retrieval and keyword matching. The comprehensive relevance score of the hybrid retrieval strategy is calculated according to the following formula: This represents the similarity score for semantic retrieval. This indicates the relevance score of keyword matching. This represents the semantic retrieval weight coefficient. This represents the keyword matching weight coefficient. and satisfy ,in The value range is from 0.5 to 0.

9. The value range is from 0.1 to 0.

5.

3. The method according to claim 1, characterized in that, In step S44, the specific process of control mode decision-making is as follows: When the signal control problem is a common problem such as unreasonable cycle or unbalanced green signal ratio, the control mode is determined to be the inductive control mode. When the signal control problem is a complex traffic flow problem with multi-period traffic flow fluctuations or significant changes in demand, the control mode is determined to be an adaptive control mode. When the signal control problem is an emergency caused by intersection overflow, traffic accident, or large-scale event, the control mode is determined to be event response control mode. When the signal control problem involves multiple types, the control mode is determined to be a combined control mode.

4. The method according to claim 3, characterized in that, The control mode decision is based on the following judgment conditions: When the maximum saturation value of each approach lane at the intersection reaches the preset saturation threshold, the event response control mode is triggered. When the queue length exceeds the preset queue length threshold, the event response control mode is triggered. When the coefficient of variation of traffic flow exceeds the preset coefficient of variation threshold within the statistical period, the adaptive control mode is triggered. When none of the above conditions are triggered, the sensor control mode is used; The saturation threshold ranges from 0.80 to 0.95, the queue length threshold ranges from 100 meters to 200 meters, and the coefficient of variation threshold ranges from 0.2 to 0.

4.

5. The method according to claim 1, characterized in that, The scheduling tools in step S40 include: Data acquisition tools are used to obtain real-time or historical traffic flow data from traffic detector interfaces; Indicator calculation tool for calculating at least one performance indicator among delay, service level, saturation, traffic efficiency, green ratio and cycle loss time; Problem diagnosis tool is used to identify at least one signal control problem among oversaturation, wasted green time, phase conflict, poor coordination, and insufficient pedestrian safety based on the comparison results of performance indicators and preset thresholds; Timing calculation tools are used to calculate signal timing parameters based on the Webster method or the HCM method. Simulation verification tools are used to call the microscopic traffic simulation engine to verify the traffic efficiency improvement effect of candidate solutions; The scheme generation tool is used to generate standardized signal timing schemes that include phase diagrams, timing tables, and time-distance diagrams.

6. The method according to claim 5, characterized in that, In step S46, after the simulation verification tool performs simulation verification on the candidate optimization strategy, if the verification result shows that the improvement rate of the key performance indicators of the candidate solution is lower than the preset convergence threshold, it will automatically revert to step S45 to regenerate the candidate optimization strategy and repeat the simulation verification process until the improvement rate of the key performance indicators reaches the convergence threshold or the number of iterations reaches the preset maximum number of iterations. The convergence threshold ranges from 3% to 10%, and the maximum number of iterations ranges from 2 to 5.

7. The method according to claim 1, characterized in that, In step S50, making the reasoning process transparent includes: Thought chain recording: Generates natural language descriptions for each reasoning step, recording the reasons and logic behind the reasoning; Tool call log: Records the input parameters, output results, and triggering reason for each tool call; Knowledge citation tracing: Mark the knowledge sources cited in the reasoning process, including standard clause numbers, historical case numbers, and theoretical literature sources; Decision-making basis chain: A complete causal reasoning chain is formed from problem diagnosis to solution recommendation, recording the basis and reasoning path of each decision node.

8. The method according to claim 1, characterized in that, The standardized signal timing scheme generated in step S47 includes: The timing parameter table includes the green light duration, yellow light duration, all-red time, and cycle length for each phase; The phase diagram shows the status of the traffic lights and the direction of passage for each phase. The time-distance diagram shows the signal timing relationship of each intersection during multi-intersection coordinated control; The reasoning process report includes a complete record of the thought process, a tool usage log, knowledge reference tracing, and a chain of decision-making basis. The expected results assessment includes a comparison of key performance indicators before and after optimization.

9. The method according to claim 1, characterized in that, The workflow orchestration supports a dynamic adjustment mechanism, including: Conditional branch: Based on the diagnostic result type of step S43, select different strategies in step S45 to generate paths; Iterative optimization: Determine whether to revert to step S45 based on the verification result of step S46; Anomaly handling: When missing data is detected, trigger supplementary data collection or use historical data to complete the data; Manual intervention: The scheme selection node in step S46 is paused and execution will continue after manual confirmation.

10. A traffic signal control optimization system based on a large-scale intelligent agent model, characterized in that, include: The interaction and input module is used to receive traffic signal control optimization requests from users; The intelligent agent reasoning module understands the request based on a large language model, decomposes it into sub-tasks, and generates a tool call plan for workflow orchestration. The knowledge retrieval module retrieves relevant professional knowledge from the traffic control knowledge base and injects it into the reasoning process through retrieval-enhanced generation technology. The tool execution module schedules the data acquisition tool, indicator calculation tool, problem diagnosis tool, timing calculation tool, simulation verification tool, and solution generation tool in sequence according to the workflow arrangement, and performs seven stages: data acquisition, performance evaluation, problem diagnosis, control mode decision, strategy generation, solution evaluation, and solution output. The transparent recording module records the complete reasoning process and generates a transparent reasoning report that includes thought chain records, tool call logs, knowledge reference tracing, and decision basis chains.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 9.

Images