Large-model-driven urban railway newly-opened station flow prediction method
By constructing a multi-dimensional reference system and a prediction method driven by a large language model, the problem of data cold start for passenger flow prediction at newly opened stations was solved, achieving high-precision and interpretable passenger flow prediction, adapting to complex scenarios, and improving the scientific and refined level of urban rail transit operation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN URBAN TRANSPORT PLANNING CENT CO LTD
- Filing Date
- 2026-04-14
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies lack historical data for passenger flow forecasting at newly opened urban rail stations, resulting in insufficient forecast accuracy. Traditional models have poor interpretability and lack flexibility in reasoning under complex scenarios, making it difficult to meet the requirements of refined operation of rail transit.
We construct a multi-dimensional reference system of 'grid space - site type - macro indicators', combine a rule engine and a large language model, drive prediction through multi-dimensional data, generate structured prompt words and perform large model inference, output quantitative prediction values and natural language inference chains, and optimize the prediction method through multi-model cross-validation and expert review.
It achieves accurate passenger flow forecasting even without historical data, improves the interpretability and operational reliability of forecast results, reduces the average absolute percentage error by 28%, adapts to complex and atypical scenarios, provides high-precision passenger flow data support, and enhances the scientific and refined level of operational organization.
Smart Images
Figure CN122022896A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of urban rail transit operation and management technology, specifically to a large-model-driven method for predicting traffic flow at newly opened urban rail transit stations. Background Technology
[0002] The opening of new urban rail transit lines and stations has become a normal part of urban development. Passenger flow forecasts for newly opened stations are an important basis for rail transit capacity allocation, train operation organization, and other operational work. However, the core problem is the lack of historical passenger flow data. Traditional forecasting methods are difficult to directly adapt, resulting in insufficient forecast accuracy and restricting the scientific and refined level of operation organization.
[0003] Currently, existing technologies attempt to address this problem primarily through three methods: First, the virtual passenger flow allocation method based on route similarity calculation, which predicts passenger flow by retrieving similar routes or station migration data. However, it heavily relies on retrieving similar objects and has limited generalization ability in special areas such as new urban areas, easily leading to prediction distortion. Second, the multi-source data prediction method based on BP neural networks, which utilizes the nonlinear mapping capability of neural networks to mine data associations. However, the model is a "black box" structure, the reasoning process lacks interpretability, and the results are difficult to understand and utilize. Third, the attribute association prediction method based on historical data clustering, which constructs a fixed statistical relationship between station attributes and passenger flow. However, it is highly dependent on historical data, has rigid reasoning logic, and is prone to prediction failure when there are no similar stations for reference.
[0004] Existing technologies have not broken through the core technology bottlenecks and generally suffer from problems such as cold start of data, poor model interpretability, and insufficient flexibility of reasoning in complex scenarios. They also lack an effective closed-loop optimization mechanism and cannot adapt to the prediction needs of atypical scenarios such as new urban areas and special functional areas. They are difficult to meet the actual requirements of refined operation of rail transit. There is an urgent need for a more adaptable and accurate method for predicting passenger flow at newly opened stations. Summary of the Invention
[0005] To address the technical problems of low prediction accuracy, poor interpretability of traditional models, and insufficient inference flexibility in complex urban scenarios caused by the lack of historical passenger flow data for newly opened urban rail transit stations, this invention provides a large model-driven method for predicting passenger flow at newly opened urban rail transit stations, comprising:
[0006] S1. Collect and preprocess multi-dimensional data related to station passenger flow;
[0007] S2. For the target newly opened site, construct a multi-dimensional reference system of "grid space - site type - macro indicators". Within the spatial geographic range of the target newly opened site, search for existing sites with the same type label as the target newly opened site as similar references. If a similar reference is found, extract the average daily entry and exit volume of at least one characteristic day among normal working days, normal rest days, the day before a holiday, and the holiday period in the past 30 days. If no similar reference is found, extract the static urban environmental data of the spatial geographic range of the target newly opened site as macro indicator references.
[0008] S3. Based on the rule engine, call the prompt word template. According to the operation status of the target newly opened site and the search results of S2, the rule engine automatically determines the scenario category to which the target newly opened site belongs, and calls the preset prompt word template corresponding to the scenario category. Fill the relevant information processed by S1 and S2 according to the organizational structure of the prompt word template to generate structured prompt words.
[0009] S4. Large model-driven reasoning and prediction generation: Input the prompt words from S3 into the large language model and guide the large language model to analyze according to the preset reasoning thought chain template, and output the prediction results including quantitative prediction values and natural language reasoning chains.
[0010] S5. Verify the prediction results described in S4, and optimize and fine-tune the prediction method based on the verification feedback.
[0011] Furthermore, in S1, the multi-dimensional data specifically includes:
[0012] Spatial basic data, including urban geographic information grids divided into preset sizes, latitude and longitude coordinates of newly opened stations and all existing stations;
[0013] Site characteristic data includes basic attributes such as type labels for existing sites based on historical passenger flow patterns, the line to which the site belongs, transfer attributes, number of entrances and exits, as well as historical inbound and outbound traffic data for existing sites;
[0014] Urban environmental cross-domain data includes static and dynamic data; the static data includes the number of various points of interest mapped to the raster, population density, and job density; the dynamic data includes at least calendar information, weather information, and event information.
[0015] The knowledge vector database is constructed by vectorizing expert experience knowledge and research report text data.
[0016] Furthermore, in S2, the spatial geographic range is a 2km*2km urban geographic information grid; the station type label is obtained by clustering analysis of the historical passenger flow patterns of existing stations, including at least one of the following: commuter type, commercial and entertainment type, transportation hub type, residential type, and mixed type.
[0017] Furthermore, in S3, the scenario categories include: a new site is already in operation with a small amount of sample data, a new site is not yet open but has similar references, and a new site is not yet open but has no similar references;
[0018] When a new site is not yet open and there are similar references, the corresponding prompt word template includes site density and passenger flow rules: if the number of similar references is >3, it is marked as diversion effect; if the number of similar references is ≤3, it is marked as induced passenger flow effect.
[0019] Furthermore, in S4, the reasoning chain template is a causal analysis template after setting the traffic prediction expert role in the large language model. The analysis steps include sequentially judging the station attributes, reference benchmark values, passenger flow trends, holidays, special weather, large-scale events, operational and policy adjustments, newly built POIs, the impact of station special characteristics, and the superposition or reduction of multiple factors.
[0020] Furthermore, in S4, the quantified predicted value is output in CSV text format with a header, and the natural language inference chain is output as an integrated whole with the quantified predicted value after being distinguished by preset tags.
[0021] Furthermore, in S5, the verification involves cross-validating the prediction results of the large language model with the results of the traditional regression model, and triggering manual review when the difference is significant; the optimization and tuning include supplementing the review feedback into the knowledge vector base and optimizing the template, and inputting real passenger flow data into the large language model in a few-sample learning manner after the new site is in operation.
[0022] The beneficial effects of this invention are:
[0023] This solution effectively addresses the challenge of cold-starting passenger flow forecasting data for newly opened stations by constructing a multi-dimensional reference system of "grid space - station type - macro indicators," enabling forecasting without relying on the target station's own historical data. By leveraging a rule engine to match customized prompt word templates and combining them with a pre-set reasoning chain to drive large-scale model inference, it not only achieves accurate adaptation of the large model to urban rail passenger flow forecasting scenarios but also provides forecast results with a complete natural language reasoning chain. This completely changes the "black box" output mode of traditional models, improving the interpretability of forecast results and the trust of operators. Simultaneously, through multi-model... The closed-loop optimization mechanism of cross-validation, expert review and feedback, and online learning with few samples continuously iterates to improve prediction accuracy. Experimental verification shows that the average absolute percentage error is reduced by 28% compared with traditional regression models. Moreover, the solution can be adapted to different stages of new station operation and various scenarios such as whether there are similar stations nearby. It still has good robustness and adaptability in complex and atypical scenarios such as new urban areas and areas with special functional positioning. It provides high-precision and highly practical passenger flow data support for capacity allocation, train operation organization, passenger transport organization and emergency response of newly opened urban rail transit stations, and improves the scientific and refined level of operation organization. Attached Figure Description
[0024] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0025] The technical solution of the present invention will be further described below with reference to embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. In the following embodiments, process equipment or devices not specifically specified are all conventional equipment or devices in the art. Unless specifically specified, the technical means used in the embodiments of the present invention are all conventional means well known to those skilled in the art.
[0026] Example 1, combined with Figure 1 This embodiment describes a large-model-driven method for predicting traffic flow at newly opened urban rail transit stations, including:
[0027] S1. Collect and preprocess multi-dimensional data related to station passenger flow;
[0028] S2. For the target newly opened site, construct a multi-dimensional reference system of "grid space - site type - macro indicators". Within the spatial geographic range of the target newly opened site, search for existing sites with the same type label as the target newly opened site as similar references. If a similar reference is found, extract the average daily entry and exit volume of at least one characteristic day among normal working days, normal rest days, the day before a holiday, and the holiday period in the past 30 days. If no similar reference is found, extract the static urban environmental data of the spatial geographic range of the target newly opened site as macro indicator references.
[0029] S3. Based on the rule engine, call the prompt word template. According to the operation status of the target newly opened site and the search results of S2, the rule engine automatically determines the scenario category to which the target newly opened site belongs, and calls the preset prompt word template corresponding to the scenario category. Fill the relevant information processed by S1 and S2 according to the organizational structure of the prompt word template to generate structured prompt words.
[0030] S4. Large model-driven reasoning and prediction generation: Input the prompt words from S3 into the large language model and guide the large language model to analyze according to the preset reasoning thought chain template, and output the prediction results including quantitative prediction values and natural language reasoning chains.
[0031] S5. Verify the prediction results described in S4, and optimize and fine-tune the prediction method based on the verification feedback.
[0032] Specifically, this invention effectively solves the problem of passenger flow prediction for newly opened stations under conditions of few or zero samples by constructing a novel prediction framework that integrates multi-source heterogeneous data with the reasoning capabilities of large language models.
[0033] As described in S3, this invention achieves robustness and adaptability in predictions under complex and atypical scenarios through a multi-dimensional reference system and the generalization reasoning capabilities of a large model. Predictive reasoning templates are provided for scenarios such as the initial / pre-opening phase of a new site, and the presence / absence of similar sites in the vicinity.
[0034] As described in S4, the core output of this invention is not only a predicted value, but also includes a natural language inference chain generated by a large model, which changes the "black box" output mode of small AI models. The inference chain clearly shows how the prediction is derived step by step.
[0035] As described in S5, domain experts can visually examine the reasoning chain. If logical flaws are found, this feedback can be used to optimize subsequent prediction prompts, forming a closed-loop optimization mechanism of "human-machine collaboration."
[0036] In S1, the multi-dimensional data specifically includes:
[0037] Spatial basic data, including urban geographic information grids divided into preset sizes, latitude and longitude coordinates of newly opened stations and all existing stations;
[0038] Site characteristic data includes basic attributes such as type labels for existing sites based on historical passenger flow patterns, the line to which the site belongs, transfer attributes, number of entrances and exits, as well as historical inbound and outbound traffic data for existing sites;
[0039] Urban environmental cross-domain data includes static and dynamic data; the static data includes the number of various points of interest mapped to the raster, population density, and job density; the dynamic data includes at least calendar information, weather information, and event information.
[0040] The knowledge vector database is constructed by vectorizing expert experience knowledge and research report text data.
[0041] Specifically, the data collected in this solution includes static data, dynamic data, and knowledge vector base data. The static data covers the number of various Points of Interest (POIs) mapped to rasters, such as residential buildings and office buildings, as well as macro-statistical indicators such as population density, job density, nighttime light intensity, and number of high-tech enterprises in each raster, obtained based on statistical yearbooks. The dynamic data includes calendar, weather, short-term event, and long-term event data. Calendar data records information such as date, day of the week, and holidays. Weather data is associated with date, geographic scope, and weather type to the corresponding raster. Short-term and long-term events record relevant time, geographic location, and event description, all of which are associated with the raster. Short-term events include concerts and temporary traffic control, while long-term events include shopping mall openings and policy releases. The knowledge vector base data is formed by vectorizing research report text data such as expert experience knowledge and data association analysis results.
[0042] In S2, the spatial geographic range is a 2km*2km urban geographic information grid; the station type label is obtained by clustering analysis of the historical passenger flow patterns of existing stations, including at least one of the following: commuter type, commercial and entertainment type, transportation hub type, residential type, and mixed type.
[0043] In S3, the scenario categories include: a new site is already in operation with a small amount of sample data, a new site is not yet open but has similar references, and a new site is not yet open but has no similar references.
[0044] When a new site is not yet open and there are similar references, the corresponding prompt word template includes site density and passenger flow rules: if the number of similar references is >3, it is marked as diversion effect; if the number of similar references is ≤3, it is marked as induced passenger flow effect.
[0045] Specifically, (1) when a new site has been operating and has a small amount of sample data (when the site has already started operating), the following is an example of the prompt template:
[0046] ## Basic Attributes
[0047] [Site Feature Data (Example: The subway station belongs to Line 11, Line 5, and Line 1; the station type is commuter station)]
[0048] ## Reference Data
[0049] [Site Inbound and Outbound Traffic Data (Example: Featured Day, Inbound Volume, Outbound Volume)]
[0050] ## Knowledge Tips
[0051] [Knowledge base text retrieval (Example: If weekday traffic is greater than weekend traffic, it indicates that the entity mainly handles commuting. Conversely, if weekday traffic is greater than weekend traffic, it indicates that the entity mainly handles leisure travel. Leisure travel will further increase on holidays compared to regular weekends.)]
[0052] Question: Predict traffic on the following dates
[0053] [Cross-domain Urban Environmental Data - Dynamic Data (Example: November 1, 2025, a typical weekday, Monday, sunny)]
[0054] (2) When the new site has not yet opened up similar references (i.e., the similar reference retrieval in S2 is not empty), the following is an example of the prompt template:
[0055] ## Basic Attributes
[0056] [Site Feature Data (Example: The subway station belongs to Line 11, Line 5, and Line 1; the station type is commuter station)]
[0057] This is a newly opened site, and there are [X1] similar sites within a 2-square-kilometer radius.
[0058] When the number of stations is greater than 3: Due to the high station density, the marginal effect of adding new stations on attracting new passenger flow is limited, and the total number of trips will not change significantly. The new stations bring a diversion effect.
[0059] When the number of stations is ≤3: Due to the low station density, adding new stations will induce new passenger flow, and the total number of trips will first increase and then be diverted.
[0060] ## Reference Data
[0061] [Site Inbound and Outbound Traffic Data (Example: Featured Day, Inbound Volume, Outbound Volume)]
[0062] ## Knowledge Tips
[0063] [Knowledge base text retrieval (Example: If weekday traffic is greater than weekend traffic, it indicates that the entity mainly handles commuting. Conversely, if weekday traffic is greater than weekend traffic, it indicates that the entity mainly handles leisure travel. Leisure travel will further increase on holidays compared to regular weekends.)]
[0064] Question: Predict traffic on the following dates
[0065] [Cross-domain Urban Environmental Data - Dynamic Data (Example: November 1, 2025, a typical weekday, Monday, sunny)]
[0066] (3) When a new site has not been opened and there are no similar references (i.e., no similar references are found in S2), the following is an example of the prompt template:
[0067] ## Basic Attributes
[0068] [Site Feature Data (Example: The subway station belongs to Line 11, Line 5, and Line 1; the station type is commuter station)]
[0069] This is a newly launched site, and there are no similar sites within a 2-square-kilometer radius. You can refer to the following information to estimate the site's traffic.
[0070] 1. Macroeconomic indicators within a 2-square-kilometer radius:
[0071] [Urban Environmental Data (Example: The grid is located in District B of City A, with a total population of [X2] people, [X3] jobs, nighttime light intensity of [X4], number of high-tech enterprises of [X5]...)]
[0072] 2. Data from other types of sites within the scope of the study
[0073] Site Feature Data (Example: Category C sites [X6]...)
[0074] ## Knowledge Tips
[0075] [Knowledge base text retrieval example: Population density determines the total travel volume; economic level affects travel mode selection, determining the total travel volume of entity types; finally, traffic allocation is performed based on the number of entities of the same type.]
[0076] Question: Predict traffic on the following dates
[0077] [Cross-domain Urban Environmental Data - Dynamic Data (Example: November 1, 2025, a typical weekday, Monday, sunny)]
[0078] In this document, the data in brackets 【】 represents dynamically updated input data, and the data in parentheses () represents example data.
[0079] In S4, the reasoning chain template is a causal analysis template after setting the traffic prediction expert role in the large language model. The analysis steps include sequentially judging station attributes, reference benchmark values, passenger flow trends, holidays, special weather, large-scale events, operational and policy adjustments, newly built POIs, the impact of station special characteristics, and the superposition or reduction of multiple factors.
[0080] Specifically, S4 utilizes Large Language Modeling (LLM) for knowledge reasoning and quantitative prediction. Roles and tasks are defined for the large model, reasoning steps are designed, and the large model is guided to fully analyze the aforementioned prompts. The reasoning thought chain template is as follows:
[0081] You are a traffic forecasting expert. Based on the case scenarios described by users, historical data, and experiential knowledge, please conduct a thorough causal analysis and deduce a logically sound prediction result.
[0082] ## Reasoning Steps
[0083] First, analyze the basic attributes to determine if it is a newly added traffic entity. If so, analyze the traffic flow characteristics of the new station.
[0084] Then, from historical sample data, reference benchmark values for the same type of date are selected for each prediction date according to "the day before the holiday, the first day of the holiday, the middle of the holiday, the last day of the holiday, a normal working day, and a normal rest day".
[0085] Next, analyze whether there are any anomalies in the baseline values; the closer the reference date is to the prediction date, the higher the reliability of the data. When anomalies are found, consider whether they are caused by related events and eliminate the impact of related events on traffic.
[0086] Then, analyze whether the annual flow is increasing year by year, and determine whether the predicted value is higher or lower than the benchmark value.
[0087] Next, consider whether holidays are involved. If so, analyze the traffic flow trend from day 1 to day n of similar holidays based on historical data to reflect the differences in daily forecasts and determine which day the maximum traffic occurs.
[0088] Next, consider whether special weather conditions are involved. If so, analyze the relationship between the duration and intensity of rainfall and the decrease in flow rate to determine whether the predicted value is higher or lower than the baseline value.
[0089] Next, consider whether large-scale events are involved. If so, analyze the relationship between the event location, scale, and duration and traffic growth, and determine the growth coefficient of the predicted value.
[0090] Next, consider whether there are any adjustments to operating hours or capacity. If so, analyze whether operating hours are extended or shortened, or whether capacity is increased or decreased, and determine whether the forecast value is higher or lower than the baseline value.
[0091] Next, consider whether emergency measures or policy adjustments are involved. If so, analyze what changes will occur in travel characteristics after the policy is released, during the policy implementation period, and after the policy ends.
[0092] Next, consider whether it involves the opening of POIs in shopping malls, scenic spots, etc. If so, analyze the relationship between the newly created POIs and traffic growth, and determine the growth coefficient of the predicted value.
[0093] Next, consider whether the site type and location are special. If so, analyze the sensitivity of this type of site to the above events and adjust the prediction value upward, downward, or remain unchanged.
[0094] Then, when the above multiple factors occur simultaneously, analyze the superposition or reduction of their effects, and test the rationality of the predicted values.
[0095] ## Output Format
[0096] Put your thought process into <think>< / think> In the tag, place your final results in CSV text format with headers. <result>< / result> Marked.
[0097] After receiving prompts, the large model simulates expert thinking, understanding and connecting all the structured and unstructured information input. It performs analogical reasoning based on similar references, quantifying and comparing differences; and it performs causal reasoning based on knowledge prompts, inferring the positive or negative impact of each factor on passenger flow. Finally, it outputs a specific predicted value, along with a complete, natural language-described reasoning chain (i.e., the thought process).
[0098] Furthermore, in S4, the quantified predicted value is output in CSV text format with a header, and the natural language inference chain is output as an integrated whole with the quantified predicted value after being distinguished by preset tags.
[0099] Furthermore, in S5, the verification involves cross-validating the prediction results of the large language model with the results of the traditional regression model, and triggering manual review when the difference is significant; the optimization and tuning include supplementing the review feedback into the knowledge vector base and optimizing the template, and inputting real passenger flow data into the large language model in a few-sample learning manner after the new site is in operation.
[0100] Specifically, to improve the reliability and adaptability of predictions, S5 establishes a human-computer interaction verification mechanism:
[0101] 1. Multi-model cross-validation: The prediction results of the large model are compared with the prediction results based on traditional methods (such as regression models). If the difference is significant, manual review is triggered.
[0102] 2. Expert Intervention and Prompt Optimization: Domain experts review the "reasoning chain" provided by the large model. If logical errors or oversights are found, this feedback is added to the knowledge prompts as new knowledge.
[0103] 3. Online learning mechanism: As the new site begins operation, a small amount of real passenger flow data from the first few days is used as a new constraint and input into the large model in a "few-sample learning" manner to improve prediction accuracy.
Claims
1. A large-model-driven method for predicting traffic flow at newly opened urban rail transit stations, characterized in that, include: S1. Collect and preprocess multi-dimensional data related to station passenger flow; S2. For the target newly opened site, construct a multi-dimensional reference system of "grid space - site type - macro indicators". Within the spatial geographic range of the target newly opened site, search for existing sites with the same type label as the target newly opened site as similar references. If a similar reference is found, extract the average daily entry and exit volume of at least one characteristic day among normal working days, normal rest days, the day before a holiday, and the holiday period in the past 30 days. If no similar reference is found, extract the static urban environmental data of the spatial geographic range of the target newly opened site as macro indicator references. S3. Based on the rule engine, call the prompt word template. According to the operation status of the target newly opened site and the search results of S2, the rule engine automatically determines the scenario category to which the target newly opened site belongs, and calls the preset prompt word template corresponding to the scenario category. The relevant information processed by S1 and S2 is filled in according to the organizational structure of the prompt word template to generate structured prompt words; S4. Large model-driven reasoning and prediction generation: Input the prompt words from S3 into the large language model and guide the large language model to analyze according to the preset reasoning thought chain template, and output the prediction results including quantitative prediction values and natural language reasoning chains. S5. Verify the prediction results described in S4, and optimize and fine-tune the prediction method based on the verification feedback.
2. The method for predicting traffic flow at newly opened urban rail transit stations driven by a large model, as described in claim 1, is characterized in that... In S1, the multi-dimensional data specifically includes: Spatial basic data, including urban geographic information grids divided into preset sizes, latitude and longitude coordinates of newly opened stations and all existing stations; Site characteristic data includes basic attributes such as type labels for existing sites based on historical passenger flow patterns, the line to which the site belongs, transfer attributes, number of entrances and exits, as well as historical inbound and outbound traffic data for existing sites; Urban environmental cross-domain data includes static and dynamic data; the static data includes the number of various points of interest mapped to the raster, population density, and job density; the dynamic data includes at least calendar information, weather information, and event information. The knowledge vector database is constructed by vectorizing expert experience knowledge and research report text data.
3. The method for predicting traffic flow at newly opened urban rail transit stations driven by a large model, as described in claim 2, is characterized in that... In S2, the spatial geographic range is a 2km*2km urban geographic information grid; the station type label is obtained by clustering analysis of the historical passenger flow patterns of existing stations, including at least one of the following: commuter type, commercial and entertainment type, transportation hub type, residential type, and mixed type.
4. The method for predicting traffic flow at newly opened urban rail transit stations driven by a large model, as described in claim 3, is characterized in that... In S3, the scenario categories include: a new site is already in operation with a small amount of sample data, a new site is not yet open but has similar references, and a new site is not yet open but has no similar references. When a new site is not yet open and there are similar references, the corresponding prompt word template includes site density and passenger flow rules: if the number of similar references is >3, it is marked as diversion effect; if the number of similar references is ≤3, it is marked as induced passenger flow effect.
5. The method for predicting traffic flow at newly opened urban rail transit stations driven by a large model, as described in claim 4, is characterized in that... In S4, the reasoning chain template is a causal analysis template after setting the traffic prediction expert role in the large language model. The analysis steps include sequentially judging station attributes, reference benchmark values, passenger flow trends, holidays, special weather, large-scale events, operational and policy adjustments, newly built POIs, the impact of station special characteristics, and the superposition or reduction of multiple factors.
6. The method for predicting traffic flow at newly opened urban rail transit stations driven by a large model, as described in claim 5, is characterized in that... In S4, the quantified prediction value is output in CSV text format with a header, and the natural language inference chain is output as an integrated whole with the quantified prediction value after being distinguished by preset tags.
7. The method for predicting traffic flow at newly opened urban rail transit stations driven by a large model, as described in claim 6, is characterized in that... In S5, the verification involves cross-validating the prediction results of the large language model with the results of the traditional regression model, and triggering manual review when the difference is significant; the optimization and tuning include supplementing the review feedback into the knowledge vector base and optimizing the template, and inputting real passenger flow data into the large language model in a few-sample learning manner after the new site is in operation.