Construction and evaluation method of autonomous transportation system scene based on prompting engineering

Abstract

The application discloses a construction and evaluation method of an autonomous traffic system scene based on prompt engineering, relates to the technical field of intelligent traffic, and comprises the following steps: defining the scene dimension of the autonomous traffic system; guiding a large language model to construct the operation scene of the autonomous traffic system according to the scene dimension by using prompt engineering; constructing evaluation indexes according to the constraint rules of the scene dimension; and evaluating the operation scene according to the evaluation indexes. The operation scene constructed by the application can provide basic input for specific technical researches such as self-organization operation and optimization. As the carrier of specific problems, the operation scene can focus on research content and be used for demonstrating the effectiveness of a theoretical framework and technical application.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation technology, and in particular to the construction and evaluation methods of autonomous transportation system scenarios based on prompting engineering. Background Technology

[0002] With the development of transportation demand and emerging technologies, Intelligent Transportation Systems (ITS) are constantly evolving into Autonomous Transportation Systems (ATS). ATS, centered on autonomous perception, learning, decision-making, and response, is characterized by reducing human intervention in passenger and freight transportation to achieve self-organizing operation. Scholars have conducted extensive research on the theoretical architecture of ATS. Related technology 1 constructs a basic theoretical framework for ATS, encompassing five key elements: components, demand, services, functions, and technologies, and systematically elucidates the logical relationships among these elements that support the system's self-organizing operation. Based on this, related technology 2 constructs a knowledge graph of ATS elements and explores the coupling relationships and key nodes between these elements, revealing the self-organizing operation mechanism of ATS in specific scenarios. Related technology 3, addressing the needs of building a personalized service system for ATS, proposes a "trust-privacy-fairness" research framework to optimize the ATS architecture.

[0003] Self-organized operation is a key capability and characteristic of ATS (Automatic Traffic System) in road traffic environments. It refers to the system's ability to balance traffic demand and resource supply through traffic flow shaping and road spatiotemporal resource allocation to achieve the safe and efficient operation of the overall transportation system. Its technological implementation is a crucial step in fully leveraging the overall system effectiveness and advancing the system's autonomy to improve the level of ATS autonomy. In recent years, related fields have deployed a series of projects and achieved results in areas such as holographic perception, interoperability, digital twins of transport equipment, and computing technology. However, in-depth exploration of self-organized operation technologies, such as the behavioral characteristics of all elements, multi-agent collaborative organization, multi-system coupling and linkage, and the matching of mixed traffic flow with road resources, remains limited. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method and related equipment for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, so as to construct and evaluate autonomous transportation system scenarios.

[0005] One aspect of this application provides a method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, the method comprising the following steps:

[0006] Define the scenario dimensions of autonomous transportation systems;

[0007] Using prompting engineering to guide a large language model to construct the operational scenario of the autonomous transportation system based on the aforementioned scenario dimension;

[0008] Evaluation metrics are constructed based on the constraint rules of the aforementioned scenario dimensions;

[0009] The operational scenario is evaluated based on the evaluation metrics.

[0010] In some embodiments, defining the scenario dimension of an autonomous transportation system includes the following steps:

[0011] Road type, scenario scale, operating conditions, service level, CAV penetration rate, control methods, and key objectives are defined as the scenario dimensions of the autonomous transportation system.

[0012] In some embodiments, the step of using prompting engineering to guide a large language model to construct the operational scenario of the autonomous transportation system based on the scenario dimension includes the following steps:

[0013] First and second prompt words are generated and constructed based on role settings, example specifications, thought processes, and search enhancements.

[0014] The scene dimension combination, the scene dimension, and the first prompt word are input into the large language model, and the scene description text is generated using the large language model.

[0015] The scene description text, the second prompt word, the scene dimension, and the elements of the autonomous transportation system are input into the large language model, and the structured operation scene is generated using the large language model.

[0016] In some embodiments, defining the retrieval enhancement generation includes the following steps: converting structured feature attributes into fine-grained knowledge blocks as an external knowledge base, instructing the large language model to retrieve and match the most relevant features based on the current context semantics during the generation instruction input stage, and injecting the most relevant features as enhanced context into prompt words for output;

[0017] Defining the role setting includes the following steps: assigning an expert identity to the large language model in the prompt instruction to activate relevant parameters, and guiding the large language model to output content that conforms to the expert identity;

[0018] Defining the example specification includes the following steps: embedding the expected answer as a template in the instruction, guiding the large language model to learn the output logic and specification for generating the first prompt word and the second prompt word without adjusting the parameters;

[0019] The definition of the thought chain includes the following steps: guiding the large language model to examine its own answers from a set perspective, and repeatedly thinking and reasoning to arrive at the final answer.

[0020] In some embodiments, the elements of the autonomous transportation system are obtained through the following steps:

[0021] The components of the defined elements include user entities, vehicles, infrastructure, traffic environment, and smart devices;

[0022] The components of the elements are defined by perceptibility and organizeability; wherein, perceptibility is used to measure the extent to which the state information of the elements is acquired by the traffic system, and perceptibility includes imperceptibility, indirect perceptibility and direct perceptibility; the organizeability is used to measure the extent to which the elements are controlled by the traffic system, and organizeability includes unorganizable, indirectly organized and directly organized.

[0023] In some embodiments, constructing evaluation metrics based on the constraint rules of the scenario dimension includes the following steps:

[0024] Based on the constraint rules of the aforementioned scenario dimension, scenario pass rate, scenario richness, self-organization level, and expert comprehensive evaluation are constructed as the evaluation indicators.

[0025] Among them, the pass rate of the scenario Represented as:

[0026] ;

[0027] in, The total number of the aforementioned running scenarios. The number of the aforementioned operating scenarios that have passed the test completely;

[0028] The richness of the scene Represented as:

[0029] ;

[0030] in, To set the combination of dimensions, , The total number of dimension combinations, This is a collection of structured scenes generated through scene creation. , The number of structured scenes generated from a single scene combination; The composite feature vector is composed of four sub-features, which are meteorological features. Environmental characteristics Entity features Statistical characteristics ; for dimensionality;

[0031] The self-organizing level Represented as:

[0032] ;

[0033] in, This refers to the number of elements in the described operating scenario with different levels of perceptibility and organization. In response The weight, This refers to the total number of elements in the aforementioned operating scenario;

[0034] The expert comprehensive assessment Represented as:

[0035] ;

[0036] in, For interaction density, Depending on the complexity of the interaction, As to the degree of dependence on people, and They are respectively , The weight.

[0037] In some embodiments, evaluating the operating scenario based on the evaluation metrics includes the following steps:

[0038] Based on the evaluation indicators, an evaluation experiment was designed from the basic model, dimensional combination, and generation scale to evaluate the operating scenario. At the same time, a sensitivity analysis was performed on the temperature parameters of the basic model, and an ablation experiment was carried out on the prompting engineering module. The basic model includes different large language models.

[0039] Another aspect of this application embodiment provides a device for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, the device comprising:

[0040] The scenario dimension definition unit is used to define the scenario dimensions of the autonomous transportation system;

[0041] The scenario construction unit is used to guide the large language model to construct the scenario of the autonomous transportation system according to the scenario dimension using prompting engineering.

[0042] An evaluation index construction unit is used to construct evaluation indexes based on the constraint rules of the scenario dimension.

[0043] The operation scenario evaluation unit is used to evaluate the operation scenario based on the evaluation indicators.

[0044] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0045] The memory is used to store programs;

[0046] The processor executes the program to implement any of the methods described above.

[0047] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0048] This application includes at least the following beneficial effects:

[0049] This application's solution includes defining the scenario dimensions of an autonomous transportation system; using prompting engineering to guide a large language model to construct operational scenarios for the autonomous transportation system based on these scenario dimensions; constructing evaluation indicators based on the constraint rules of the scenario dimensions; and evaluating the operational scenarios based on these evaluation indicators. The operational scenarios constructed in this application can provide basic input for specific technical research such as self-organizing operation and optimization. As a carrier of specific problems, these operational scenarios can focus on research content and be used to demonstrate the effectiveness of the theoretical framework and technical applications. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating the method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, as provided in the embodiments of this application;

[0052] Figure 2 An example flowchart of the method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering provided in the embodiments of this application;

[0053] Figure 3 This is a schematic diagram of the scene generation process provided in the embodiments of this application;

[0054] Figure 4 A schematic diagram of the ATS component framework provided in an embodiment of this application;

[0055] Figure 5 A schematic diagram illustrating the search enhancement generation provided in the embodiments of this application;

[0056] Figure 6 Example diagrams of the prompting project provided in the embodiments of this application;

[0057] Figure 7 A schematic diagram illustrating the contents of a sample example provided in this application embodiment;

[0058] Figure 8An example diagram illustrating the perceptibility and organization of scene elements provided in the embodiments of this application;

[0059] Figure 9 A schematic diagram illustrating the comprehensive scenario evaluation provided in the embodiments of this application;

[0060] Figure 10 A schematic diagram illustrating the experimental results of model parameter changes provided in the embodiments of this application;

[0061] Figure 11 This is a structural block diagram of the device for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, provided in an embodiment of this application. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0064] Related technology 1 designs five typical scenarios—autonomous driving at intersections, highway platooning, mobility-as-a-service, electric buses, and multimodal transport—around the goals of safety, convenience, efficiency, greenness, and economy, for the verification and application of the macro-theoretical architecture. Related technology 4 analyzes user needs in these scenarios based on a short-text two-layer matching model, using a MaaS scenario as an example for demonstration. Related technology 5 proposes an ATS architecture integrity assessment method based on an information-triggered collaboration mechanism, using the scenario of autonomous vehicles passing through intersections as an example for verification. Related technology 6 constructs a hierarchical, evolvable architecture model using road intersections as an empirical scenario, analyzing the intergenerational evolution path of ATS. Many other studies focus on the application of self-organizing traffic scenarios, such as urban road vehicle-road cooperative optimization and autonomous highway platooning. However, the scenarios currently used in theoretical framework or optimization algorithm research rely on empirical settings and have not been systematically demonstrated. Therefore, sorting out and constructing self-organizing ATS operation scenarios can provide scenario input and basis for specific research.

[0065] In traffic research, scenario construction often employs methods such as hierarchical partitioning, data-driven approaches, and parameter combination. Hierarchical partitioning involves dividing scenarios into functional, logical, and specific scenarios based on their level of abstraction, refining scenario parameters layer by layer. Data-driven approaches utilize large amounts of real-world data and machine learning methods such as clustering to construct scenario systems. Parameter combination primarily targets the key performance parameters of the scenario's main components, defining scenario boundaries while combining different parameters to form specific scenarios. The self-organizing operation problem addresses autonomous transportation systems with varying levels of autonomy at multiple levels, currently lacking data support and parameter anchoring. While top-level design methods based on hierarchical partitioning can provide a reference, the numerous and complex relationships within self-organizing scenarios make translating the top-level design into concrete scenarios a crucial challenge.

[0066] The rapid development of Large Language Models (LLMs) has provided new ideas for solving the scene construction problem. LLMs are large-scale parametric models trained on vast amounts of corpora, possessing rich knowledge reserves. They can transform abstract scenes into concrete scene descriptions or standardized text expressions through text generation. Related technology 7 uses methods such as role setting, sample examples, thought chains, and consistency checks to guide LLMs in converting scene description texts into specific format files required for simulation. Related technology 8 uses multimodal LLMs to transform images, videos, and accident reports into unified structured scene descriptions, employs thought chain reasoning to extract key risk elements, and utilizes strategies such as context learning and domain knowledge embedding to guide the model in generating code files that meet simulation requirements. Related technology 9 utilizes LLMs to convert natural language instructions into specific scripts, enabling the automatic generation, customization, and analysis of traffic scenes in the SUMO simulator.

[0067] In summary, the theoretical framework of ATS has been initially established, and LLM is widely used in traffic scenario construction. However, it still has certain limitations in constructing ATS self-organizing operation scenarios. ATS-related research focuses on theoretical architecture construction or breakthroughs in specific functional technologies, lacking a top-level design and evaluation system to support the self-organizing evolution analysis of ATS. Current LLM-based scenario generation research focuses on transforming given scenario descriptions into standardized simulation files, but the construction of ATS scenarios should consider the generation mechanism from abstract theory to concrete scenarios. General-purpose LLM lacks domain knowledge constraints during the generation process, making it difficult to reflect the self-organizing operation characteristics of ATS.

[0068] Against this backdrop, this application proposes a scenario construction and evaluation framework based on LLM (Limited Ledger Model) hinting engineering to address the self-organizing operation requirements of ATS (Automatic Training System). This application analyzes scenario dimensions based on the self-organizing operation mechanism, constructs an ATS scenario element knowledge base, embeds domain knowledge using retrieval-enhanced generation technology, and combines hinting engineering-driven LLM to transform abstract dimensions into concrete scenario descriptions and standardized expressions, establishing multi-dimensional index evaluation outputs. Finally, through multiple sets of comparative experiments, ablation experiments, and sensitivity analyses, the effectiveness and robustness of the proposed method are demonstrated. The core solution of this application includes: comprehensively utilizing LLM hinting engineering to construct concrete scenarios from abstract ATS theory, providing standardized scenario inputs for research on self-organizing operation mechanisms, system simulation, and optimization; and the designed scenario dimensions and evaluation indicators provide a basis for research on self-organizing operation scenarios.

[0069] Reference Figure 1 This application provides a method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, specifically including the following steps S100~S130:

[0070] S100: Defines the scenario dimensions of autonomous transportation systems;

[0071] S110: Use prompting engineering to guide the large language model to construct the operation scenario of the autonomous transportation system based on the scenario dimension;

[0072] S120: Construct evaluation metrics based on the constraint rules of the aforementioned scenario dimension;

[0073] S130: Evaluate the operating scenario based on the evaluation indicators.

[0074] Optionally, defining the scenario dimension of an autonomous transportation system includes the following steps:

[0075] Road type, scenario scale, operating conditions, service level, CAV penetration rate, control methods, and key objectives are defined as the scenario dimensions of the autonomous transportation system.

[0076] Optionally, the step of using prompting engineering to guide the large language model to construct the operational scenario of the autonomous transportation system based on the scenario dimension includes the following steps:

[0077] First and second prompt words are generated and constructed based on role settings, example specifications, thought processes, and search enhancements.

[0078] The scene dimension combination, the scene dimension, and the first prompt word are input into the large language model, and the scene description text is generated using the large language model.

[0079] The scene description text, the second prompt word, the scene dimension, and the elements of the autonomous transportation system are input into the large language model, and the structured operation scene is generated using the large language model.

[0080] Optionally, the definition of the retrieval enhancement generation includes the following steps: converting structured feature attributes into fine-grained knowledge blocks as an external knowledge base, instructing the large language model to retrieve and match the most relevant features based on the current context semantics during the generation instruction input stage, and injecting the most relevant features as enhanced context into the prompt words for output;

[0081] Defining the role setting includes the following steps: assigning an expert identity to the large language model in the prompt instruction to activate relevant parameters, and guiding the large language model to output content that conforms to the expert identity;

[0082] Defining the example specification includes the following steps: embedding the expected answer as a template in the instruction, guiding the large language model to learn the output logic and specification for generating the first prompt word and the second prompt word without adjusting the parameters;

[0083] The definition of the thought chain includes the following steps: guiding the large language model to examine its own answers from a set perspective, and repeatedly thinking and reasoning to arrive at the final answer.

[0084] Optionally, the elements of the autonomous transportation system are obtained through the following steps:

[0085] The components of the defined elements include user entities, vehicles, infrastructure, traffic environment, and smart devices;

[0086] The components of the elements are defined by perceptibility and organizeability; wherein, perceptibility is used to measure the extent to which the state information of the elements is acquired by the traffic system, and perceptibility includes imperceptibility, indirect perceptibility and direct perceptibility; the organizeability is used to measure the extent to which the elements are controlled by the traffic system, and organizeability includes unorganizable, indirectly organized and directly organized.

[0087] Optionally, constructing the evaluation index based on the constraint rules of the scenario dimension includes the following steps:

[0088] Based on the constraint rules of the aforementioned scenario dimension, scenario pass rate, scenario richness, self-organization level, and expert comprehensive evaluation are constructed as the evaluation indicators.

[0089] Among them, the pass rate of the scenario Represented as:

[0090] ;

[0091] in, The total number of the aforementioned running scenarios. The number of the aforementioned operating scenarios that have passed the test completely;

[0092] The richness of the scene Represented as:

[0093] ;

[0094] in, To set the combination of dimensions, , The total number of dimension combinations, This is a collection of structured scenes generated through scene creation. , The number of structured scenes generated from a single scene combination; The composite feature vector is composed of four sub-features, which are meteorological features. Environmental characteristics Entity features Statistical characteristics ; for dimensionality;

[0095] The self-organizing level Represented as:

[0096] ;

[0097] in, This refers to the number of elements in the described operating scenario with different levels of perceptibility and organization. In response The weight, This refers to the total number of elements in the aforementioned operating scenario;

[0098] The expert comprehensive assessment Represented as:

[0099] ;

[0100] in, For interaction density, Depending on the complexity of the interaction, As to the degree of dependence on people, and They are respectively , The weight.

[0101] Optionally, evaluating the operating scenario based on the evaluation metrics includes the following steps:

[0102] Based on the evaluation indicators, an evaluation experiment was designed from the basic model, dimensional combination, and generation scale to evaluate the operating scenario. At the same time, a sensitivity analysis was performed on the temperature parameters of the basic model, and an ablation experiment was carried out on the prompting engineering module. The basic model includes different large language models.

[0103] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0104] Current transportation systems are continuously evolving towards Autonomous Transportation Systems (ATS), and preliminary explorations have been made in their theoretical framework and technology. Self-organizing operation scenarios are crucial for clarifying research boundaries and focusing on real-world problems in ATS research, but currently, top-level design and specific construction methods are lacking. This embodiment proposes a scenario construction and evaluation framework based on LLM (Limited Learning Model) prompting engineering. It designs scenario dimensions based on self-organizing operation characteristics, constructs an ATS scenario element knowledge base, guides LLM to construct structured scenarios through prompting engineering, and conducts evaluation based on multi-dimensional indicators built from self-organizing relevance. Experimental results show that this framework can generate ATS self-organizing operation scenarios in most LLMs, but the pass rate and richness are affected by model capabilities. This embodiment uses ablation experiments to illustrate the key roles of retrieval enhancement generation and the thought chain module in this embodiment, as well as the guiding role of the role setting and example sample modules in scenario generation. This embodiment can provide a scenario library and theoretical support for the analysis and simulation optimization of ATS self-organizing operation mechanisms.

[0105] 1. Research framework and methodology.

[0106] 1.1 Overall framework.

[0107] The technical framework of this embodiment can be divided into three stages, such as Figure 2 As shown. The first stage is scenario-level design, which analyzes multi-dimensional traffic elements from top to bottom based on the ATS self-organizing operation mechanism and traffic business logic to form a complete scenario combination space. The second stage is the scenario generation stage, which uses an element knowledge base as a foundation and leverages prompting engineering to guide LLM to transform from abstract dimensions to concrete semantic descriptions, and outputs structured operational scenarios. The third stage is the scenario evaluation stage, which establishes a multi-index evaluation matrix from the perspectives of logical consistency, generation diversity, and system autonomy level, and conducts eight sets of basic experiments and ablation experiments to clarify the effectiveness of the method and the self-organizing relevance of the scenarios.

[0108] 1.2 Scene Dimension Design.

[0109] Based on the meaning of self-organized operation and the basic characteristics of the scenario, this embodiment designs seven key scenario dimensions from four aspects: spatial attributes, operating environment, technical characteristics and organizational goals, as shown in Table 1.

[0110] Table 1 Scene Dimensions and Values

[0111]

[0112] (1) Spatial scope: Road type anchors the physical background of the scene to distinguish different traffic supply and demand characteristics; scale defines the research scope from the point-line-surface perspective and divides it into three scales: node, road segment and road network.

[0113] (2) Operational scope: Operational conditions are used to classify whether the operating environment is normalized, divided into normal state and emergency state with extremely high requirements for self-organization capability; Service level dimension is divided into free flow, stable flow and saturated flow according to the degree of matching between traffic flow and traffic capacity, which is used to characterize the operational quality of traffic flow in the scene.

[0114] (3) Technical Scope: CAV penetration rate refers to the proportion of connected and automated vehicles (CAVs) in a scenario, categorized into no CAV, mixed traffic, and full CAV states. Vehicles are key carriers for the self-organized operation of traffic. The penetration rate of CAVs has a profound impact on macro-level traffic flow organization and micro-level entity interaction behavior, and at a high level of autonomy, it can serve as a means of regulating traffic flow. Control methods refer to the means or methods by which a traffic system achieves self-organized operation, and are a key perspective for classifying self-organized scenarios. Specifically, the values ​​are divided into four states: no control, human control, roadside facility control, and vehicle-road-cloud collaborative control, covering common scenarios such as no signal control, traffic police command, and intersection signal control, as well as the still developing vehicle-road-cloud collaborative control.

[0115] (4) Scope of objectives: Key objectives are used to clarify the focus of self-organizing operation in the scenario, including safety, convenience, efficiency, greenness and economy.

[0116] 1.3 Scene construction method.

[0117] Dimensions, while dividing scenarios at the top-level design level, still only yield abstract expressions of these scenarios, which are insufficient to directly support specific needs such as twinning, deduction, and optimization. They need to be further mapped to concrete, standardized scenarios. This section guides the LLM to understand the scenario's meaning by designing multiple modules within the project, and uses ATS domain knowledge to generate standardized scenario expressions in stages. The technical process is as follows: Figure 3 As shown.

[0118] In the first stage of scene generation, dimensions and their values ​​are defined as background content explanations, and ATS elements serve as knowledge supplements. The transformation from dimension combination to descriptive text is mainly guided by role setting and mind chain modules. The second stage transforms the scene description text into a structured expression, comprehensively using modules such as role setting, example specification, mind chain, and retrieval enhancement to guide LLM output. The specific module content will be explained in detail in 1.3.2.

[0119] 1.3.1 ATS Element Analysis.

[0120] In the theoretical elements of ATS (Automatic Traffic System), components are the carriers of functions and services, and also relate to needs and technologies. Therefore, to a certain extent, the operational state of components can be considered a concentrated manifestation of the system's self-organizing characteristics. Regarding the self-organizing operational scenario construction problem focused on in this embodiment, it is necessary to further analyze the relevant elements in the road system based on ATS components. The ATS component framework includes categories such as user entities, vehicles, infrastructure, traffic environment, and intelligent devices, such as... Figure 4 As shown.

[0121] To characterize the self-organization level of each element in a scenario, this embodiment defines two core attributes—perceptibility and organizeability—based on the ATS component framework. Perceptibility measures the extent to which the traffic system acquires the state information of an element, categorized as imperceptible, indirectly perceptible, and directly perceptible. Imperceptible means the system cannot acquire any state information of the element; indirectly perceptible means the system can acquire its state information through third-party media, such as roadside cameras recognizing pedestrian movement; directly perceptible means the system can directly acquire the element's state information, such as the status feedback from traffic control facilities. Organability measures the extent to which an element is controlled by the traffic system, categorized as unorganizable, indirectly organizeable, and directly organizeable. Unorganizable means the element cannot be physically or logically controlled by the system in real time, such as road structure or weather conditions; indirectly organizeable means the system can guide the object by coordinating related elements, such as using fences to guide pedestrian flow; directly organizeable means the system has the ability to directly make decisions and control the element's behavioral state, such as signal timing optimization.

[0122] 1.3.2 Tips for the project.

[0123] Related research indicates that carefully crafted prompts and well-designed task templates can enhance the logical reasoning and content generation capabilities of LLM in specific tasks. With the aim of generating high-quality ATS self-organizing scenarios, this embodiment employs a combination of strategies, including retrieval-enhanced generation, role setting, example specification, and thought chaining, to design prompts.

[0124] (1) Enhanced search generation.

[0125] Retrieval-Augmented Generation (RAG) is an effective paradigm for improving the generation quality of LLMs by providing them with a searchable external knowledge base. It alleviates the LLM illusion problem while enhancing the accuracy of their domain knowledge. This embodiment utilizes RAG to embed ATS element knowledge to compensate for the lack of ATS-specific knowledge in general LLMs. First, structured ATS element attributes are transformed into fine-grained knowledge blocks as an external knowledge base. During the generation instruction input stage, the LLM is allowed to retrieve and match the most relevant elements based on the current context semantics, injecting them as enhanced context into the prompt words for output. The specific implementation process is as follows: Figure 5 As shown.

[0126] (2) Character setting.

[0127] The Role Setting (RS) module assigns an expert role to the model in the prompts to activate relevant model parameters and guide the model to output content consistent with that role, thereby improving the accuracy of the results. In this embodiment, the model's role is adjusted according to the task stage to improve both the professionalism of scene generation and the standardization of data format. In the scene description generation stage, the LLM is set as a traffic simulation scene designer, enabling it to perform professional reasoning from a traffic simulation perspective based on dimensional combinations and output logically consistent scene description text, such as... Figure 6 Prompt 1 is shown. During the data structuring phase, the model's identity is switched to a JSON synthesis expert, reinforcing the LLM's consideration of consistency between preceding text and RAG retrieval elements when generating structured text, and requiring the output to conform to a given example specification, such as... Figure 6 The Prompt 2 shown.

[0128] (3) Example specifications.

[0129] The Few-Shot (FS) strategy embeds the expected response as a template into the instruction, guiding the LLM to learn the output logic and specifications for a specific task without adjusting parameters, thereby improving the quality and formatting of the generated content. This embodiment constructs a paradigm from manually verified standard ATS scenario data and embeds it into Prompt 2. This includes multi-dimensional information such as weather, road network topology, infrastructure attributes, and traffic participant behavior, assisting the LLM in learning the mapping relationship between abstract scenario descriptions and JSON objects. Specific examples are shown below. Figure 7 As shown.

[0130] (4) Mind chain.

[0131] Chain of Thought (CoT) simulates the human thought process, guiding the model to examine its own answers from a predetermined perspective, repeatedly considering and reasoning to arrive at a final, accurate answer. This can enhance the interpretability and accuracy of LLM generation to some extent. In this embodiment, corresponding chain of thought strategies are embedded in both the scene description generation and structured expression stages, thereby guiding the model to conduct differentiated reasoning and step-by-step decoupling of the scene. The scene description generation stage forces the model to identify high-frequency scenes and actively adjust the generation logic by reviewing existing scenes, formulating differentiated strategies, and checking the compliance of scene composition, ensuring that the richness of the output can be improved while complying with the specifications. The scene structured expression stage simulates the step-by-step analysis logic of experts, providing guidance through specific steps such as global search, entity naming, element mapping, attribute filling, numerical generation, and integrity verification, mitigating information omissions and parameter illusions in the model output process.

[0132] 1.4 Construction of scenario evaluation indicators.

[0133] To further verify the reliability of the LLM output results, this embodiment constructs a multi-dimensional evaluation system from aspects such as generation quality, output richness, and self-organization correlation. The specific indicators are shown in Table 2.

[0134] Table 2 Scenario Evaluation Index System

[0135]

[0136] Indicator 1 (C1) primarily uses pre-coded verification rules to judge scene quality, as shown in Table 3. It's important to note that only scenes passing the C1 metric can be included in the calculation of subsequent indicators. Indicator 2 (C2) evaluates whether the structured scene representation output by the scene generation framework is sufficiently rich when inputting a single-dimensional combination. Indicator 3 (C3) calculates the degree of self-organization of elements within the scene to assess its self-organization level. Indicator 4 (C4) introduces LLM experts to comprehensively score the scene results from multiple perspectives.

[0137] Table 3 Specific Rules for Calculating Scene Pass Rate Indicators

[0138]

[0139] This embodiment combines specific dimensions. The following example illustrates the calculation of the indicator. , This represents the total number of dimension combinations. A set of structured representations is generated through scene generation. ,in , The number of structured specific scenes generated from a combination of individual scenes.

[0140] (1) Scene pass rate.

[0141] The calculation of scene pass rate mainly focuses on the structural specifications, content quality, condition constraints, and dimension matching of the final output scene. The specific rules are designed based on the importance of information and the observation experience of the experimental process, as shown in Table 3.

[0142] Based on the above rules, the number of scenarios that completely pass the test is: The scene pass rate can then be expressed as:

[0143]

[0144] in, The total amount of structured representation of the generated scene.

[0145] (2) Scene richness.

[0146] This embodiment constructs a richness index based on a composite feature space to examine whether there are a large number of similar scenes in the model output. First, any scene is structurally represented. Transformed through semantic model Numerical eigenvectors of dimension The expression is as follows:

[0147]

[0148] in, The composite feature vector is composed of four sub-features, which are meteorological features. Environmental characteristics Entity features Statistical characteristics .

[0149] Subsequently, a weighted method was used to calculate the difference distances for different types:

[0150]

[0151] in, Cosine distance For Jaccard distance, The normalized Euclidean distance. These are the weight coefficients for each feature dimension.

[0152] cosine distance Defined as:

[0153]

[0154]

[0155] Jaccard Distance Defined as:

[0156]

[0157] Furthermore, the scene richness index It can be defined as:

[0158]

[0159] in, To generate a combination of dimensions A scenario.

[0160] (3) Scene self-organization level.

[0161] A scene's self-organizing capability depends on the degree of autonomy of its constituent elements. Based on the ATS element attribute table, the self-organization level of elements within a scene can be statistically analyzed. Figure 8 In the scenario shown, elements such as CAVs and traffic lights are perceptible and controllable elements, while elements such as motorcycles, weather, and road topology are perceptible but uncontrollable elements. Vehicles obscured by obstacles are unperceptible and uncontrollable elements under low levels of autonomy technology.

[0162] Therefore, the self-organization level of a scene can be characterized by the distribution of the perceptibility and organizeability of elements in the output scene:

[0163]

[0164] in, The number of elements with different levels of perception and organization in the scene. For the corresponding weights, This represents the total number of elements in the scene.

[0165] (4) Scenario comprehensive evaluation score.

[0166] This embodiment introduces LLM to simulate human experts, requiring the model to conduct a comprehensive evaluation from specific dimensions, such as... Figure 9 As shown. Interaction density Used to measure the scale of participants and the complexity of interactions within an area of ​​mutual influence in a scenario. This assesses the game-theoretic nature of traffic behavior and the degree of dependence on human intervention. The self-organization level of a scenario is measured by following the principle that "the less the need for people, the higher the level of self-organization".

[0167] All indicators are scored on a scale of 1 to 5, and the final comprehensive score is calculated using a weighted average method.

[0168]

[0169] in, and In the experiment, a value of 0.5 was used to balance spatial density and interaction. Considering the potential bias caused by the LLM illusion problem, the evaluation scores were manually verified during the experiment.

[0170] 2. Experimental Design and Results.

[0171] The effectiveness verification of this embodiment framework is mainly carried out from three aspects: first, verifying whether the structured scene meets the simulation research requirements in terms of content and format; second, evaluating whether the richness of the scene set generated by the combination of a single dimension is sufficient to support subsequent simulation and optimization research; and third, examining whether the scene is composed of ATS elements and reflects the level of self-organization.

[0172] Based on the evaluation metrics established above, we designed experiments from the perspectives of basic model, dimensional combination, and generation scale. We also conducted sensitivity analysis on model temperature parameters and carried out ablation experiments on key engineering modules. Therefore, this study selected five LLMs as the basic models for scene generation for comparative analysis: DeepSeek-Reasoner (DS), Doubao-Seed-1.6-thinking (DB), Qwen3-Max (Qwen), GPT-5.2 (GPT), and Gemini-2.5-Pro ​​(Gemini). The expert scoring process uniformly used DeepSeek, maintaining consistent hardware and hyperparameter configurations.

[0173] Due to the large space for combining scenario dimensions, this embodiment selects two sets of significantly different dimension combinations as inputs for the experiment, as specifically defined in Table 4. Here, S1 represents the roadside facility control scenario in an urban mixed traffic environment, while S2 represents the vehicle-road-cloud cooperative self-organizing control scenario in a high-level autonomous driving environment.

[0174] Table 4. Experimental Scenario Dimension Combinations

[0175]

[0176] Considering that the generation scale may affect the calculation of scene richness index, this embodiment sets up 3 sets of experiments to analyze the degree of influence of the generation scale, with the number of scene outputs being 10, 50, and 100, respectively. Taking into account the base model, input scene, and generation scale, this embodiment conducts a total of 8 sets of experiments. The specific configuration and evaluation results of the experiments are shown in Table 5. The experimental configurations are named in the format of "Scene ID-Model Abbreviation-Number of Generations". For example, S1-DS-10 represents the experimental group that uses scene S1 as input and generates 10 scene samples based on the DeepSeek model.

[0177] Table 5 Experimental Evaluation Results

[0178]

[0179] 3. Analysis and Discussion.

[0180] 3.1 Comparative experimental analysis.

[0181] Table 5 shows experiments 1-5 comparing the performance of the proposed method in different LLMs. The DS, DB, GPT, and Gemini models showed high scene generation pass rates, while the QW model had a significantly lower C1 score due to poor performance in generating standardization. Regarding the C2 scene richness index, only the DB model had a relatively low result of 0.562. Since the richness index calculations are based on scenes validated by the C1 index, and QW generates a smaller number of usable samples, even with a C2 index of 0.633, it cannot be concluded that QW performed well in terms of richness. Considering the number of scenes included in the calculation, the DS model performed slightly better than GPT and Gemini in the C2 richness index. Regarding the C3 scene self-organization level index, QW and DB had the highest and lowest values, respectively, while the other models had values ​​around 0.8, confirming that the proposed framework can effectively drive the LLM to invoke self-organizing elements. Regarding the comprehensive evaluation score C4, apart from the QW model, which may have a relatively high score due to the smaller number of scenario samples included in the calculation, the other four models performed similarly.

[0182] Based on the differences in the index results of various models under the same configuration, it can be concluded that: (1) The pass rate index is a prerequisite. The scenarios generated by models such as DS, DB, GPT, and Gemini are relatively more in line with the requirements. Among them, DS, GPT, and Gemini perform relatively well in terms of scenario richness index. The three models perform similarly in terms of C3 and C4 index; (2) There are certain differences in the capabilities of the models. Reasonable selection can improve the effectiveness of the generated scenarios.

[0183] To verify the robustness of the framework in this embodiment under different scenario input conditions, experiments were conducted using two sets of dimension combinations as inputs based on the DS model, namely Experiment 1 and Experiment 6 shown in Table 5. Although the C1 and C2 indices decreased due to the influence of the randomness of the LLM output, the C3 and C4 indices increased significantly. This situation is presumably due to the difference in the self-organization level of the scenario combinations themselves. Scenario S2 is a fully automated driving cooperative control environment, and its corresponding specific scenario should logically have a higher level of self-organization than scenario S1.

[0184] Experiments 1, 7, and 8 used the same model and input scenarios, but with varying numbers of output scenarios to explore the impact of the number of output scenarios on the model's performance. Table 5 shows that as the generation scale increases, the four metrics decrease to varying degrees, with decreases of 17%, 14.8%, 4.8%, and 7.2%, respectively. Considering the results of other experiments, a C1 metric performance above 80% is considered normal, and the results of Experiments 7 and 8 differ by only 1%, further indicating that the model's performance on the C1 metric is close to 80%. The decrease in the C2 metric is mainly due to the use of a pairwise distance metric mechanism in the metric calculation; the increase in the number of parameters involved in the calculation itself affects the metric results. The C3 and C4 metrics remained relatively stable in this experiment, demonstrating the robustness of the framework at the level of self-organized correlation.

[0185] 3.2 Parameter sensitivity analysis.

[0186] Temperature is a key variable controlling the randomness of LLM generation. In this embodiment, DS is used as the base model to perform parameter sensitivity analysis on two prompts in the scene generation process. The results are as follows: Figure 10 As shown.

[0187] pass Figure 10 It can be observed that, regarding the C1 pass rate indicator, among the 36 groups of experiments conducted, 30 groups had a pass rate of 90% or higher, including 12 parameter groups with a pass rate of 100%, which indicates the role of the prompting engineering strategy in standardizing the quality of the generated product.

[0188] In the 12 experiments where C1 achieved a 100% pass rate, the maximum C2 richness score was 0.670 and the minimum was 0.592, corresponding to temperature parameter combinations (0, 0.3) and (0.5, 0.3), respectively; the maximum C3 self-organization level score was 0.811 and the minimum was 0.750, corresponding to temperature parameter combinations (1, 1) and (0.5, 0.3), respectively; and the maximum C4 expert comprehensive evaluation score was 0.414 and the minimum was 0.338, corresponding to temperature parameter combinations (0, 1) and (0.5, 0.5), respectively. It can be observed that temperature parameter changes do not have a significant regular impact on the experimental results; therefore, a relatively balanced temperature parameter can be selected. When the Prompt temperature parameter is set to (1.0, 1.0), the system maintains a relatively balanced richness, self-organization level, and comprehensive score while ensuring a 100% pass rate.

[0189] 3.3 Ablation experiment.

[0190] To evaluate the effectiveness of each module in the method, ablation experiments were conducted using S1-DS-10 as the baseline, and the results are shown in Table 6.

[0191] Table 6 Ablation Experiment Results

[0192]

[0193] In Table 6, w / o represents no corresponding module.

[0194] In Experiment 2, the scene pass rate (C1) and scene richness (C2) were slightly lower than the baseline, while the self-organization level (C3) was 15.7% lower than the baseline, resulting in a comprehensive evaluation score that differed from the baseline by 0.011. This indicates that RS can appropriately improve scene richness and drive LLM to construct scenes using self-organizing elements. In Experiments 3 and 5, the scene pass rate (C1) was 0, indicating that the absence of the RAG and CoT modules would prevent LLM from generating effective output. This highlights that the extensive ATS domain knowledge provided by RAG and the guidance rules embedded in CoT are key components of this implementation framework, making these two modules crucial. In Experiment 4, the scene pass rate (C1) decreased by 30%, and the self-organization level (C3) decreased by 4.5%, demonstrating that the example specifications provided by FS can play a guiding role and effectively improve the standardization of generated scenes.

[0195] 4. Conclusion.

[0196] This embodiment addresses the lack of scenario support in ATS self-organizing operation research by proposing a scenario construction and evaluation framework based on LLM suggestion engineering. Scenario dimensions are defined according to the characteristics of self-organizing operation. A scenario generation method is designed based on LLM-integrated retrieval enhancement generation, role setting, sample examples, and thought chains. The dimensions are combined to transform into specific structured scenarios, and multi-dimensional indicators are constructed to evaluate the scenario generation method and the self-organization level of the scenarios. The effectiveness of the method is demonstrated through multiple sets of comparative experiments, ablation experiments, and sensitivity analysis. The results can provide a rich scenario library and theoretical support for research on self-organizing operation mechanisms and simulation evolution.

[0197] In the future, the self-organizing operation process of scenarios can be enriched by specifying the values ​​of scenario dimensions and strengthening the LLM guidance specifications. Simulation software such as Carla and SUMO can be used to transform the structured expression scenarios into simulation experimental scenarios. The operational efficiency of different self-organizing operation strategies in diverse scenarios can be compared and analyzed to further improve the scenario typicality evaluation model and explore the collaborative optimization research of ATS traffic organization strategies.

[0198] Reference Figure 11 This application provides an apparatus for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, including:

[0199] The scenario dimension definition unit is used to define the scenario dimensions of the autonomous transportation system;

[0200] The scenario construction unit is used to guide the large language model to construct the scenario of the autonomous transportation system according to the scenario dimension using prompting engineering.

[0201] An evaluation index construction unit is used to construct evaluation indexes based on the constraint rules of the scenario dimension.

[0202] The operation scenario evaluation unit is used to evaluate the operation scenario based on the evaluation indicators.

[0203] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0204] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented in this application. Alternative embodiments are contemplated, in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0205] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the device disclosed in this application, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0206] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0207] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0208] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0209] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0210] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0211] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0212] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, characterized in that, The method includes the following steps: Define the scenario dimensions of autonomous transportation systems; Using prompting engineering to guide a large language model to construct the operational scenario of the autonomous transportation system based on the aforementioned scenario dimension; Evaluation metrics are constructed based on the constraint rules of the aforementioned scenario dimensions; The operational scenario is evaluated based on the evaluation metrics.

2. The method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering according to claim 1, characterized in that, The definition of the scenario dimension of the autonomous transportation system includes the following steps: Road type, scenario scale, operating conditions, service level, CAV penetration rate, control methods, and key objectives are defined as the scenario dimensions of the autonomous transportation system.

3. The method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering according to claim 1, characterized in that, The method of using prompting engineering to guide the large language model to construct the operational scenario of the autonomous transportation system based on the scenario dimension includes the following steps: First and second prompt words are generated and constructed based on role settings, example specifications, thought processes, and search enhancements. The scene dimension combination, the scene dimension, and the first prompt word are input into the large language model, and the scene description text is generated using the large language model. The scene description text, the second prompt word, the scene dimension, and the elements of the autonomous transportation system are input into the large language model, and the structured operation scene is generated using the large language model.

4. The method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering according to claim 3, characterized in that, The definition of the retrieval enhancement generation includes the following steps: converting structured feature attributes into fine-grained knowledge blocks as an external knowledge base; in the generation instruction input stage, instructing the large language model to retrieve and match the most relevant features based on the current context semantics; and injecting the most relevant features as enhanced context into the prompt words for output. Defining the role setting includes the following steps: assigning an expert identity to the large language model in the prompt instruction to activate relevant parameters, and guiding the large language model to output content that conforms to the expert identity; Defining the example specification includes the following steps: embedding the expected answer as a template in the instruction, guiding the large language model to learn the output logic and specification for generating the first prompt word and the second prompt word without adjusting the parameters; The definition of the thought chain includes the following steps: guiding the large language model to examine its own answers from a set perspective, and repeatedly thinking and reasoning to arrive at the final answer.

5. The method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering according to claim 3, characterized in that, The elements of the autonomous transportation system are obtained through the following steps: The components of the defined elements include user entities, vehicles, infrastructure, traffic environment, and smart devices; The components of the elements are defined by perceptibility and organizeability; wherein, perceptibility is used to measure the extent to which the state information of the elements is acquired by the traffic system, and perceptibility includes imperceptibility, indirect perceptibility and direct perceptibility; the organizeability is used to measure the extent to which the elements are controlled by the traffic system, and organizeability includes unorganizable, indirectly organized and directly organized.

6. The method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering according to claim 1, characterized in that, The process of constructing evaluation metrics based on the constraint rules of the scenario dimension includes the following steps: Based on the constraint rules of the aforementioned scenario dimension, scenario pass rate, scenario richness, self-organization level, and expert comprehensive evaluation are constructed as the evaluation indicators. Among them, the scene pass rate Represented as: ； in, The total number of the aforementioned running scenarios. The number of the aforementioned operating scenarios that have passed the test completely; The richness of the scene Represented as: ； in, To set the combination of dimensions, , The total number of dimension combinations, This is a collection of structured scenes generated through scene generation. , The number of structured scenes generated from a single scene combination; The composite feature vector is composed of four sub-features, which are meteorological features. Environmental characteristics Entity features Statistical characteristics ; for dimensionality; The self-organization level Represented as: ； in, This refers to the number of elements in the described operating scenario with different levels of perceptibility and organization. In response The weight, This refers to the total number of elements in the aforementioned operating scenario; The expert comprehensive assessment Represented as: ； in, For interaction density, Depending on the complexity of the interaction, As to the degree of dependence on people, and They are respectively , The weight.

7. The method for constructing and evaluating autonomous transportation system scenarios based on prompting engineering according to any one of claims 1 to 6, characterized in that, The evaluation of the operating scenario based on the evaluation indicators includes the following steps: Based on the evaluation indicators, an evaluation experiment was designed from the basic model, dimensional combination, and generation scale to evaluate the operating scenario. At the same time, a sensitivity analysis was performed on the temperature parameters of the basic model, and an ablation experiment was carried out on the prompting engineering module. The basic model includes different large language models.

8. A device for constructing and evaluating autonomous transportation system scenarios based on prompting engineering, characterized in that, The device includes: The scenario dimension definition unit is used to define the scenario dimensions of the autonomous transportation system; The scenario construction unit is used to guide the large language model to construct the scenario of the autonomous transportation system according to the scenario dimension using prompting engineering. An evaluation index construction unit is used to construct evaluation indexes based on the constraint rules of the scenario dimension. The operation scenario evaluation unit is used to evaluate the operation scenario based on the evaluation indicators.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 7.

Images