Scenic spot ticketing scheduling method and system

Abstract

The application provides a scenic spot ticket scheduling method and system, wherein the method comprises: acquiring numerical time series data and text data of a scenic spot, and acquiring statistical baseline prediction results and text evidence of future time periods; based on a large language model, according to the numerical time series data, the statistical baseline prediction results and the text evidence, outputting passenger flow prediction values of the future time periods, and fusing the statistical baseline prediction results to obtain passenger flow demand predictions of the future time periods; based on the passenger flow demand predictions, according to dynamic effective carrying capacity and a tourist stay time length distribution in the scenic spot, calculating real-time residual sellable capacity, and according to the real-time residual sellable capacity, taking minimization of load rate variance of the future time periods as an objective, optimizing and solving sellable ticket quotas of the future time periods. The method improves the accuracy of passenger flow demand prediction and the response ability to text signals such as sudden public opinion, and realizes balanced utilization of peak shaving, valley filling and scenic spot carrying capacity.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and smart tourism technology, and in particular to a method and system for scheduling ticketing at scenic spots. Background Technology

[0002] In recent years, with the advancement of smart scenic area construction, scenic area ticketing systems have gradually shifted towards online and time-slot reservation models. However, existing scenic area ticketing systems still have significant shortcomings in terms of visitor flow forecasting and ticket quota management, which restricts the improvement of scenic area operational efficiency and visitor experience.

[0003] In terms of visitor flow forecasting, tourist flow at scenic spots is influenced by multiple factors, including weather changes, holiday arrangements, traffic control, and social media dissemination. Many key signals exist in the form of natural language text, such as weather warnings, holiday adjustment announcements, traffic control notices, and public opinion content. Traditional visitor flow forecasting methods mainly rely on fitting historical numerical time-series data, which can only process structured data and struggles to effectively integrate the aforementioned textual signals. This results in a significant delay in forecasting before sudden surges in visitor flow, failing to provide timely and accurate information for ticketing decisions.

[0004] In terms of ticket quota management, common ticketing systems typically release tickets according to a fixed daily or time-slot quota, treating the scenic area's carrying capacity as a static constant. This means they cannot adjust in real time according to weather, facility availability, or unforeseen events. When a surge in ticket purchases occurs, the system can only implement hard sell-out or circuit breaker measures, lacking smooth flow control and proactive peak-shifting methods. This can easily lead to a situation where tickets are snapped up online, resulting in overcrowding at the park, which not only affects the visitor experience but may also pose safety hazards.

[0005] Therefore, there is an urgent need for a scenic area ticketing scheduling method that can incorporate textual signals into passenger flow prediction and dynamically adjust ticket quotas based on real-time carrying capacity. Summary of the Invention

[0006] This invention provides a scenic area ticketing scheduling method and system to solve the problems in the prior art where the prediction of scenic area visitor flow is difficult to integrate with text signals, resulting in prediction lag, and the disconnect between ticket quotas and real-time carrying capacity, leading to offline congestion. It improves the accuracy of visitor flow prediction, achieves a balance of visitor load in different time periods, and avoids offline congestion in scenic areas.

[0007] On one hand, the present invention provides a scenic area ticketing scheduling method, comprising: acquiring numerical time-series data and text data of the scenic area, and obtaining statistical baseline prediction results for each future time period based on the numerical time-series data, and obtaining text evidence for each future time period based on the text data; based on a large language model, outputting passenger flow prediction values ​​for each future time period based on the numerical time-series data, the statistical baseline prediction results, and the text evidence, and fusing the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain passenger flow demand predictions for each future time period; based on the passenger flow demand predictions for each future time period, calculating the real-time remaining saleable capacity based on the dynamic effective carrying capacity and the distribution of visitor dwell time in the scenic area, and optimizing the saleable ticket quota for each future time period based on the real-time remaining saleable capacity with the objective of minimizing the load rate variance for each future time period.

[0008] Furthermore, the step of outputting passenger flow prediction values ​​for future time periods based on the large language model, according to the numerical time series data, the statistical baseline prediction results, and the textual evidence, includes: constructing structured prompt words, wherein the structured prompt words include the numerical time series data, the statistical baseline prediction results for future time periods, the textual evidence for future time periods, and output constraints; inputting the structured prompt words into the large language model, and outputting passenger flow prediction values ​​for future time periods and corresponding natural language inference chains.

[0009] Furthermore, by fusing the passenger flow forecasts for each future time period with the statistical baseline forecasts, passenger flow demand forecasts for each future time period are obtained. This includes: determining the fusion weights online based on the exponential smoothing relative error and confidence level of the large language model prediction and the exponential smoothing relative error of the statistical baseline prediction; and weighting and fusing the passenger flow forecasts for each future time period with the statistical baseline forecasts based on the fusion weights to obtain the passenger flow demand forecasts for each future time period.

[0010] Furthermore, the dynamic effective carrying capacity is calculated based on the basic carrying capacity, weather attenuation factor, emergency event factor, and facility opening rate factor; wherein, the weather attenuation factor is obtained by parsing weather warning text, the emergency event factor is obtained by parsing control announcement text, and the facility opening rate factor is determined based on the real-time opening status of scenic area facilities.

[0011] Furthermore, the process of obtaining the passenger flow demand forecast for each future time period includes: decomposing the passenger flow demand forecast for each future time period into a predetermined arrival volume and a controllable demand, depending on whether it is subject to quota control; wherein, the predetermined arrival volume includes the portion of sold but unredeemed tickets expected to arrive during that time period, as well as the arrival volume of free tickets, annual passes, and group tickets that do not occupy the available ticket quota; the controllable demand is the demand of individual tourists who have not yet purchased tickets and can be absorbed by the available ticket quota.

[0012] Furthermore, the calculation of the real-time remaining saleable capacity based on the predicted passenger flow demand for each future time period, according to the dynamic effective carrying capacity and the distribution of visitor dwell time within the scenic area, includes: calculating the conditional departure probability of each visitor within the scenic area based on the distribution of visitor dwell time within the scenic area; determining the predicted number of visitors leaving the scenic area based on the conditional departure probability of each visitor within the scenic area; and calculating the real-time remaining saleable capacity based on the dynamic effective carrying capacity, the current number of visitors in the scenic area, the predicted number of visitors leaving the scenic area, and the portion of the sold but unredeemed tickets in the predetermined arrival volume that is expected to arrive during that time period.

[0013] Furthermore, based on the real-time remaining available capacity, and with the objective of minimizing the load factor variance for each future time period, the available ticket quota for each future time period is optimized and solved. This includes: determining the actual admission volume for each future time period as the sum of the predetermined arrival volume and the available ticket quota for the corresponding time period; and solving the available ticket quota for each future time period with the objective of minimizing the load factor variance of the actual admission volume for each future time period relative to the dynamic effective carrying capacity. The sum of the available ticket quotas for each future time period does not exceed the real-time remaining available capacity and does not exceed the adjustable demand for the corresponding time period.

[0014] Furthermore, it also includes: performing heat statistics on the public opinion data in the text data to obtain the public opinion heat index for each time period; calculating the congestion level for each future time period based on the predicted number of people in the park, the dynamic effective carrying capacity, the number of reservations, and the public opinion heat index; and adaptively adjusting the real-time release rate of ticket purchase requests for the corresponding time period based on the congestion level for each future time period.

[0015] Furthermore, the scenic area ticketing management is carried out collaboratively by the operation intelligence agent and the maintenance intelligence agent; wherein, the operation intelligence agent is responsible for business-side control, and the maintenance intelligence agent is responsible for system-side support; the operation intelligence agent and the maintenance intelligence agent achieve linkage by exchanging events through shared state indexes.

[0016] Secondly, the present invention also provides a scenic area ticketing scheduling system, comprising: a multi-source data acquisition and text evidence generation module, used to acquire numerical time-series data and text data of the scenic area, and to acquire statistical baseline prediction results for each future time period based on the numerical time-series data, and to acquire text evidence for each future time period based on the text data; a large model prediction and passenger flow demand fusion module, used to output passenger flow prediction values ​​for each future time period based on a large language model, according to the numerical time-series data, the statistical baseline prediction results and the text evidence, and to fuse the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain passenger flow demand predictions for each future time period; and a dynamic capacity calculation and quota optimization module, used to calculate the real-time remaining saleable capacity based on the passenger flow demand predictions for each future time period, according to the dynamic effective carrying capacity and the distribution of tourist dwell time in the scenic area, and to optimize the saleable ticket quota for each future time period based on the real-time remaining saleable capacity, with the goal of minimizing the load rate variance for each future time period.

[0017] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the scenic spot ticketing scheduling method described above.

[0018] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the scenic area ticketing scheduling method as described above.

[0019] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the scenic area ticketing scheduling method described above.

[0020] The scenic area ticketing scheduling method provided by this invention acquires numerical time-series data and textual data of the scenic area, obtains statistical baseline prediction results for each future time period based on the numerical time-series data, and obtains textual evidence for each future time period based on the textual data; based on a large language model, it outputs passenger flow prediction values ​​for each future time period based on the numerical time-series data, statistical baseline prediction results, and textual evidence, and integrates the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain passenger flow demand predictions for each future time period; based on the passenger flow demand predictions for each future time period, it calculates the real-time remaining saleable capacity based on the dynamic effective carrying capacity and the distribution of visitor dwell time in the scenic area, and optimizes the saleable ticket quota for each future time period based on the real-time remaining saleable capacity with the goal of minimizing the load rate variance for each future time period. This method improves the accuracy of passenger flow demand forecasting and the ability to respond to text signals such as sudden public opinion events by integrating numerical time-series data and textual data and using a large language model for logical reasoning prediction. By combining dynamic effective carrying capacity and the distribution of tourist dwell time in real time to calculate the remaining saleable capacity, the capacity assessment is upgraded from static setting to dynamic prediction. Finally, the ticket quota is optimized with the goal of minimizing the variance of the load rate in each time period, which can smoothly distribute passenger flow demand to all time periods of the day, effectively achieving peak shaving and valley filling and balanced utilization of the scenic area's carrying capacity. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the scenic area ticketing scheduling method provided in this embodiment of the invention.

[0023] Figure 2 This is a flowchart illustrating the construction process of textual evidence provided in an embodiment of the present invention.

[0024] Figure 3 This is a flowchart of passenger flow demand prediction based on a large language model provided in an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram illustrating the optimized process of the available ticket quota provided in an embodiment of the present invention.

[0026] Figure 5 This is a time-series diagram illustrating the multi-turn dialogue profile extraction and package combination generation provided in this embodiment of the invention.

[0027] Figure 6This is a schematic diagram of the workflow of the operational intelligent agent provided in the embodiment of the present invention.

[0028] Figure 7 This is a schematic diagram of the workflow of the operation and maintenance intelligent agent provided in the embodiment of the present invention.

[0029] Figure 8 This is a schematic diagram of the closed-loop feedback and online calibration of the scenic area ticketing scheduling method provided in this embodiment of the invention.

[0030] Figure 9 This is an overall architecture diagram of the scenic area ticketing scheduling method provided in the embodiments of the present invention.

[0031] Figure 10 This is a schematic diagram of the scenic area ticketing scheduling system provided in an embodiment of the present invention.

[0032] Figure 11 This is a schematic diagram of the physical structure of the electronic device provided in the embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0034] Figure 1 A flowchart illustrating the scenic area ticketing scheduling method provided in an embodiment of the present invention is shown. Figure 1 As shown, the method includes: S110, acquiring numerical time-series data and text data of the scenic area, and obtaining statistical baseline prediction results for each future time period based on the numerical time-series data, and obtaining text evidence for each future time period based on the text data; S120, based on a large language model, outputting passenger flow prediction values ​​for each future time period based on the numerical time-series data, the statistical baseline prediction results, and the text evidence, and fusing the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain passenger flow demand predictions for each future time period; S130, based on the passenger flow demand predictions for each future time period, calculating the real-time remaining saleable capacity based on the dynamic effective carrying capacity and the distribution of tourist dwell time in the scenic area, and optimizing the saleable ticket quota for each future time period based on the real-time remaining saleable capacity with the objective of minimizing the load rate variance for each future time period.

[0035] It should be noted that the scenic area ticketing scheduling method provided in this embodiment is applied to the field of smart tourism and digital management of scenic areas, and is especially suitable for time-sharing reservations and online ticketing scenarios for scenic area tickets and related products, such as sightseeing vehicles, cruise ships, and guided tours.

[0036] The scenic area ticketing scheduling method provided in this embodiment takes a smart scenic area ticketing system based on artificial intelligence big data model and vector retrieval as the execution subject. Specifically, the system can be a software service cluster deployed on the scenic area's cloud server or local data center, and is connected to the scenic area's gate equipment, ticketing terminals, user terminals, and external data sources such as meteorological services, social media platforms, and traffic data interfaces through network communication.

[0037] During operation, the smart scenic area ticketing system builds time-series data streams and vector indexes based on the Elasticsearch 9.x search engine or its subsequent compatible versions. It utilizes Large Language Models (LLM) to provide inference prediction and structured output capabilities, uses Redis in-memory database to ensure quota counting and distributed locks under high concurrency, and leverages Kafka message queues and Flink stream processing engine to achieve real-time data access. It also has a configurable LLM adapter to shield the interface differences between different model vendors.

[0038] The following will provide a detailed description of steps S110-S130 and related steps.

[0039] S110, acquire numerical time-series data and text data of the scenic area, and obtain statistical baseline prediction results for each future time period based on the numerical time-series data, and obtain text evidence for each future time period based on the text data.

[0040] In this step, the numerical time-series data consists of continuously collected scenic area operation indicators over time, including historical and real-time minute-level visitor numbers. Number of people leaving the park This includes ticket sales and refunds for different time periods. Text-based data refers to unstructured data in the form of natural language text, including public opinion data, environmental text, rule text, and ticketing product databases.

[0041] The data includes public opinion data such as the number of views, interactions, and text related to scenic spots on short videos or social media platforms; environmental texts such as weather forecasts, meteorological warnings, and temporary traffic control announcements; rule texts such as holiday and work schedule adjustments and ticket policies; and a ticketing product database including structured attributes of products such as tickets, sightseeing buses, cruise ships, guided tours, cultural and creative packages, and parking vouchers, such as price, age range, time of day, and accessibility attributes, as well as natural language descriptions of the products.

[0042] The process is straightforward: collect numerical time-series data and text-based data, and write them to Elasticsearch 9.4.x. Specifically, numerical time-series data is written to a Time Series Data Stream (TSDS), with the index mode set to `time_series`, and gate numbers used as time-series dimensions. Historical data retains hourly and daily granularity using a sampling strategy. Public opinion data is written to the `semantic_text` field, automatically segmented and vectorized via the Inference API during the write process; environmental text and rule text are also written to the `semantic_text` field; structured attributes of ticketing products are stored in inverted indexes or keyword fields, while natural language descriptions of ticketing products are stored in both the `semantic_text` and `dense_vector` fields.

[0043] The vector fields mentioned above all use dense_vector with a dimension d of 1024. They are used in conjunction with HNSW (Hierarchical Navigable Small World) graph indexes. The index option type is either bbq_hnsw (in-memory binary quantization) or bbq_disk (DiskBBQ) for scenarios where the vector set exceeds memory. The HNSW parameters are set to m = 16 and ef_construction = 200.

[0044] Feature engineering is performed using ES|QL (Elasticsearch Query Language). The TSDS time-series data stream (i.e., the collected numerical time-series data) is combined with the weather dimension table and the holiday dimension table using a lookup join. This time-series aggregation produces feature frames for each time period and solidifies them into an ES|QL View, i.e., a virtual index. Subsequent passenger flow demand prediction tasks can directly read from this view to ensure global consistency of feature definitions.

[0045] After obtaining the numerical time-series data, the statistical baseline prediction results for each future period are obtained based on the numerical time-series data. Specifically, an additive exponential smoothing model that includes intraday and intraweek dual seasonal terms can be used to fit the historical numerical time-series data, and the output passenger flow baseline prediction values ​​for each future period that reflect the regular historical patterns are obtained. This is the statistical baseline prediction result.

[0046] Simultaneously, textual evidence for future time periods is obtained based on textual data. Specifically, textual data is vectorized and indexed using an embedding model. When predictions for a target time period (future time periods) are needed, a query request is constructed, and candidate texts are retrieved in parallel from the index through a hybrid retrieval method. This hybrid retrieval combines lexical relevance retrieval and vector similarity retrieval. After inverse ranking fusion and semantic reordering, diverse selection is performed according to the maximum marginal relevance rule, ultimately yielding a set of textual evidence related to the target time period. This textual evidence includes case information on similar weather, holidays, and public opinion characteristics in history, as well as currently active warnings or announcements.

[0047] Next, proceed to step S120.

[0048] S120, based on the large language model, according to the numerical time series data, the statistical baseline prediction results and the text evidence, outputs the passenger flow prediction values ​​for each future time period, and integrates the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain the passenger flow demand prediction for each future time period.

[0049] Specifically, the numerical characteristics of recent time periods, including the current number of people in the park and the number of reservations, as well as the statistical baseline prediction results and textual evidence obtained from step S110, are constructed into prompt words according to a predefined structured output format, such as JSONSchema, and input into the large language model. The large language model utilizes its logical reasoning capabilities to output the predicted visitor flow values ​​for future time periods and the corresponding reasoning basis.

[0050] Subsequently, the passenger flow forecasts for each future time period are integrated with the statistical baseline forecasts to obtain the passenger flow demand forecasts for each future time period. Specifically, the confidence level of the large language model prediction results is calculated, for example, by obtaining the coefficient of variation through multiple temperature samplings, and the recent relative errors of the two predictions are statistically analyzed. Based on the aforementioned confidence level and relative error, the online adaptive fusion weights are calculated using the Softmax function, and the large language model predictions and the statistical baseline forecasts are convexly combined to obtain the final passenger flow demand forecast.

[0051] The passenger flow demand forecast obtained in this step not only includes the regularity of numerical time series, but also incorporates the impact of sudden events in text data, which can more comprehensively reflect the actual ticket purchase demand in different time periods in the future.

[0052] Next, proceed to step S130.

[0053] S130, based on the predicted passenger flow demand for each future time period, calculate the real-time remaining saleable capacity according to the dynamic effective carrying capacity and the distribution of tourist stay duration in the scenic area, and optimize the solution of the saleable ticket quota for each future time period based on the real-time remaining saleable capacity with the goal of minimizing the variance of the load rate for each future time period.

[0054] In this step, the dynamic effective carrying capacity is the effective value of the scenic area's maximum instantaneous carrying capacity after dynamic correction based on real-time environmental factors. Specifically, the basic carrying capacity is multiplied by a weather attenuation factor, an emergency event factor, and a facility opening rate factor. These factors can be automatically generated by parsing warning or announcement texts using a large language model, thus allowing the carrying capacity to expand and contract in real time according to factors such as weather and facility status.

[0055] The distribution of visitor dwell time within a scenic area is a statistical distribution function that determines the duration of visitor stays, such as a log-normal distribution obtained by fitting historical gate pairing data. The load factor is the ratio of the actual number of visitors allowed to the dynamic effective carrying capacity for each future time period.

[0056] In practice, the distribution of visitor dwell time within the scenic area is used to calculate the conditional departure probability of visitors currently in the park, and then the summation is used to obtain the predicted number of visitors leaving the park within a future time window. Combining the current number of visitors in the park, the predicted number of visitors leaving the park, and the dynamic effective carrying capacity, the number of tickets sold but not yet redeemed and expected to arrive is deducted to calculate the real-time remaining available capacity. After obtaining the real-time remaining available capacity, an optimization model is constructed with the objective of minimizing the sum of squares of the differences between the load rate of each time period and the average load rate of the whole day, and total quantity constraints and box constraints are applied, with the quota upper limit not exceeding the real-time remaining available capacity and physical passage capacity.

[0057] The optimization model is a strictly convex quadratic programming problem, which can be solved in a closed loop using the water injection algorithm or by using the bisection method to determine the water level parameters, thereby obtaining the available ticket quota for each time period. This quota allocation scheme can smoothly distribute passenger flow demand across different time periods, achieving peak shaving and valley filling.

[0058] In this embodiment, numerical time-series data and text data of the scenic area are acquired. Statistical baseline predictions for future time periods are obtained based on the numerical time-series data, and textual evidence for future time periods is obtained based on the text data. Based on a large language model, passenger flow predictions for future time periods are output based on the numerical time-series data, statistical baseline predictions, and textual evidence. These predictions are then fused with the statistical baseline predictions to obtain passenger demand predictions for future time periods. Based on these passenger demand predictions, the real-time remaining saleable capacity is calculated according to the dynamic effective carrying capacity and the distribution of visitor dwell time within the scenic area. Finally, based on the real-time remaining saleable capacity, the available ticket quota for future time periods is optimized by minimizing the variance of the load rate for each future time period. This method improves the accuracy of passenger flow demand forecasting and the ability to respond to text signals such as sudden public opinion events by integrating numerical time-series data and textual data and using a large language model for logical reasoning prediction. By combining dynamic effective carrying capacity and the distribution of tourist dwell time in real time to calculate the remaining saleable capacity, the capacity assessment is upgraded from static setting to dynamic prediction. Finally, the ticket quota is optimized with the goal of minimizing the variance of the load rate in each time period, which can smoothly distribute passenger flow demand to all time periods of the day, effectively achieving peak shaving and valley filling and balanced utilization of the scenic area's carrying capacity.

[0059] Based on the above embodiments, the following will further describe in detail the process of constructing textual evidence in step S110.

[0060] The process involves obtaining textual evidence for future time periods based on textual data, including: constructing vector and lexical indexes for the textual data; performing lexical relevance retrieval on the textual data based on the lexical index and vector similarity retrieval on the textual data based on the vector index for each future time period; fusing the lexical relevance retrieval results and vector similarity retrieval results, and performing semantic rearrangement on the fused results; and diversifying the selection of semantic rearrangement results according to the maximum marginal relevance rule to obtain textual evidence for each future time period.

[0061] It is easy to understand that, firstly, a vector index and a lexical index are constructed for text data. Specifically, any text d is mapped to a unit vector through the embedding endpoints of the inference API, as shown in the following formula (1).

[0062] (1).

[0063] In equation (1), This is an embedded function.

[0064] To maintain millisecond-level retrieval at a vector scale of tens of millions, the vectors are binarized dimension by dimension using BBQ (Better Binary Quantization) technology. The i-th quantized bit is shown in the following formula (2).

[0065] (2).

[0066] In equation (2), The quantization threshold for this dimension is obtained from training samples. The storage cost after quantization is approximately 1 / 32 of the original float32.

[0067] Based on the above vectorization results, a vector index is constructed, and a lexical index, i.e., an inverted index, is also built.

[0068] Then, for each future time period, lexical relevance retrieval is performed on the text data based on the lexical index, and vector similarity retrieval is performed on the text data based on the vector index. For vector similarity retrieval, the query input for each future time period is the query vector. With document vectors The cosine similarity is shown in the following formula (3).

[0069] (3).

[0070] A two-stage retrieval scoring method is adopted. In the first stage, candidates are quickly recalled in the quantization space by bitwise operation with approximate similarity. The quantization distance is based on the XOR bit count, as shown in the following formula (4).

[0071] (4).

[0072] In the second stage, the original vectors of the Top-R candidates are accurately re-scored according to formula (3) to balance recall and precision. The KNN retrieval with structured filtering (K-Nearest Neighbors) adopts a filtering-aware graph traversal strategy to avoid recall loss caused by searching first and then filtering.

[0073] In terms of lexical relevance retrieval, the same query vector is scored on the inverted index according to BM25 (Best Match 25), as shown in the following formula (5).

[0074] (5).

[0075] In equation (5), (Inverse Document Frequency) is the inverse document frequency; For word frequency; The average document length =1.2, =0.75.

[0076] Next, the lexical relevance retrieval results and the vector similarity retrieval results are fused, and the fused results are semantically rearranged. Specifically, through the retrieval framework of the retrieval engine, BM25 and KNN retrieval are executed in parallel within a single request, and the reciprocal rank fusion is performed by the RRF (Reciprocal Rank Fusion) retrieval tool, as shown in the following formula (6).

[0077] (6).

[0078] The fusion results are then refined by a semantic reordering retrieval system, which is a reordering inference endpoint, and a text block rescoring system.

[0079] Finally, the semantic rearrangement results are diversified according to the maximum marginal relevance rule to obtain textual evidence for each future time period. Specifically, the fine-ranking results are selected iteratively by the maximum marginal relevance (MMR) rule, as shown in the following formula (7).

[0080] (7).

[0081] In equation (7), The best candidate document selected in the current iteration step is one of the final selected textual evidences; The set of candidate documents, i.e., the semantic re-ranking results / fine-tuning results; The selected set of evidence; The correlation-diversity trade-off coefficient is set to 0.7 in this embodiment; This is the vector representation of the selected document s.

[0082] Ultimately, the Top-K values ​​are selected. RAG evidence set constituting the large model This refers to textual evidence, which includes historical cases that are most similar to the target prediction day in terms of weather, holidays, and public opinion trends, the original text of currently active public opinion hotspots, and effective control or early warning announcements.

[0083] Accordingly, Figure 2 A flowchart illustrating the construction process of textual evidence provided in an embodiment of the present invention is shown. Figure 2As shown, the process begins with text / product / case data being entered into the database. After automatic segmentation and vectorization, a sparse vector with a quantized spatial structure and XOR bit counts are constructed using a dense vector index. Combining query input and structured filtering conditions, BM25 keyword retrieval and KNN vector retrieval are performed in parallel. The results from both retrieval paths are then fused using RRF (Relational Ranking Framework). The fused Top-R original vectors undergo semantic rearrangement and MMR (Multi-Rate Rank) diversification to ultimately generate a RAG (Relational Aggregate) evidence set, including similar historical daily cases, active public opinion focal points, and effective early warning or regulatory announcements.

[0084] Based on the above embodiments, further, Figure 3 This paper presents a flowchart of passenger flow demand prediction based on a large language model, as provided in an embodiment of the present invention. The following will be combined with... Figure 3 The process of predicting passenger flow in future time periods using a large language model in step S120 is described in detail.

[0085] Structured prompts are constructed, comprising numerical time-series data, statistical baseline predictions for each future time period, textual evidence for each future time period, and output constraints. These structured prompts are then input into a large language model, which outputs passenger flow predictions for each future time period and corresponding natural language inference chains. Based on the exponential smoothing relative error and confidence level of the large language model's predictions, and the exponential smoothing relative error of the statistical baseline predictions, fusion weights are determined online. Based on these fusion weights, the passenger flow predictions for each future time period and the statistical baseline predictions for each future time period are weighted and fused to obtain passenger demand predictions for each future time period.

[0086] The process is straightforward: obtain numerical time-series data / features from the ES|QL feature view, including recent visitor flow, current park attendance, and reservations, and then acquire the statistical baseline prediction results generated through bi-seasonal exponential smoothing. Specifically, the statistical baseline prediction results are obtained by fitting historical numerical time-series data to an additive exponential smoothing model that includes intraday and weekly bi-seasonal terms. This model outputs baseline predictions of future visitor flow for each time period, reflecting typical historical patterns; this is the statistical baseline prediction result. As shown in equation (8).

[0087] (8).

[0088] In equation (8), , These are horizontal and trend components, respectively. , These are the intraday seasonal items and the weekly seasonal items, respectively; period. Number of ticket sales periods within the day; Get the corresponding number of time periods within a week; The current moment is the end point in time for known historical data; To predict the step size, corresponding to future time periods; in engineering implementation, TBATS, Prophet, or multi-seasonal state-space models can also be used to obtain the same two-period baseline.

[0089] Meanwhile, RAG text evidence, after being diversified by MMR, is introduced, covering public opinion, weather, announcements, and historical cases. Structured cue words are constructed, which include numerical time series data, statistical baseline prediction results for each future time period, text evidence for each future time period, and output constraints (JSON Schema). Subsequently, the structured cue words are input into the multimodal large language model for inference and prediction. The large language model outputs the passenger flow prediction values ​​for each future time period and the corresponding natural language inference chain, specifically satisfying the following formula (9).

[0090] (9).

[0091] In equation (9), These are the predicted passenger flow values ​​for different time periods in the future. For traceable natural language reasoning chains, such as "the number of views of short videos related to the scenic spot increased by 320% month-on-month in 24 hours. Based on similar historical cases, the number of visitors is expected to increase by 1.8 times during the 10:00–12:00 period"; It is the sum of numerical time series data and the statistical baseline prediction results for each future time period; It is a set of textual evidence.

[0092] In the above process, self-consistent sampling and confidence level calculation are also performed. Specifically, K predictions are generated independently based on temperature sampling (K=5 in this embodiment), and the median is taken. As a point prediction, and with the coefficient of variation characterizing the confidence level. It satisfies the following formula (10).

[0093] (10).

[0094] In equation (10), Forecasting passenger flow demand for future time periods; Confidence level; The coefficient of variation; The standard deviation of the predicted values ​​from multiple samplings; This is the average of the predicted values ​​from multiple samplings.

[0095] Next, the passenger flow forecasts for each future time period are integrated with the statistical baseline forecasts to obtain the passenger flow demand forecasts for each future time period. Specifically, the integration weights are determined online based on the exponential smoothing relative error and confidence level of the large language model forecast and the exponential smoothing relative error of the statistical baseline forecast. Among them, the errors of the large language model forecast and the statistical baseline forecast are taken as the exponential smoothing relative errors (MAPE type), which satisfy the following formula (11).

[0096] (11).

[0097] In equation (11), This is the exponentially smoothed relative error (MAPE type) at the current time t. This is the error smoothing coefficient; These are actual observed values; These are predicted values.

[0098] The fusion weights are updated online by Softmax using the recent errors and confidence scores of the large language model prediction and the statistical baseline prediction, satisfying the following formula (12).

[0099] (12).

[0100] In equation (12), For weighting; The confidence level of predictions made by the large language model; For temperature parameters; The exponentially smoothed relative error predicted by the large language model; The exponentially smoothed relative error of the statistical baseline prediction is used.

[0101] It should be noted that the fusion weight The recent errors and confidence scores of the large language model predictions and the statistical baseline predictions are updated online using Softmax. When the confidence score of the large language model... High and prediction error Hour, Increase the weight of the large language model automatically, such as on days with sudden public opinion events or extreme weather; conversely, decrease the weight of the large language model, focusing on low-cost statistical baselines, such as on regular days. The entire process of integrating weight adjustments is completely automatic.

[0102] Finally, based on the fusion weights Passenger flow forecasts for different time periods in the future And the statistical baseline forecast results for each future period Weighted fusion is performed to satisfy the following formula (13).

[0103] (13).

[0104] In this embodiment, by constructing structured prompts containing numerical time-series data, statistical baseline prediction results, textual evidence, and output constraints, and inputting them into a large language model, the model can output passenger flow predictions for future time periods and traceable natural language inference chains, enhancing its responsiveness to textual signals such as sudden public opinion events and extreme weather, as well as its predictive interpretability. Simultaneously, prediction confidence is obtained through self-consistent sampling, and the fusion weights are determined online adaptively based on the exponentially smoothed relative errors and confidence levels of the two predictions. On regular days, the low-cost statistical baseline is primarily used, while on days with sudden events, the large model weight is automatically increased, achieving an intelligent balance between large model prediction and statistical baseline prediction. This optimizes system computational overhead while ensuring prediction accuracy.

[0105] Based on the above embodiments, further, Figure 4 This diagram illustrates the optimized process for allocating ticket quotas according to an embodiment of the present invention. The following will be combined with... Figure 4 The process of quota optimization in step S130 is described in detail.

[0106] Based on the distribution of visitor dwell time within the scenic area, the conditional departure probability of each visitor is calculated. Based on this probability, the predicted number of visitors leaving the scenic area is determined. The real-time remaining available capacity is calculated based on the dynamic effective carrying capacity, the current number of visitors in the scenic area, the predicted number of visitors leaving, and the portion of sold but unredeemed tickets expected to arrive during the specified time period from the predetermined arrival volume. The actual admission volume for each future time period is determined as the sum of the predetermined arrival volume and the corresponding available ticket quota for that time period. With the objective of minimizing the variance of the load rate between the actual admission volume and the dynamic effective carrying capacity for each future time period, the available ticket quota for each future time period is solved. The sum of the available ticket quotas for each future time period must not exceed the real-time remaining available capacity and must not exceed the controllable demand for the corresponding time period.

[0107] It is easy to understand that the dynamic effective carrying capacity in this embodiment is calculated based on the basic carrying capacity, weather attenuation factor, emergency event factor and facility opening rate factor, and satisfies the following formula (14).

[0108] (14).

[0109] In equation (14), This represents the maximum instantaneous carrying capacity of the scenic area. This is a weather attenuation factor, which is adjusted downwards during heavy rain / high temperature warnings. This is a contingency factor that will accumulate when restrictions are in place or the performance ends. This is a facility openness factor, which is calculated based on the open area when some areas are under maintenance.

[0110] The weather attenuation factor is obtained by parsing weather warning texts, the emergency event factor is obtained by parsing traffic control announcement texts, and the facility opening rate factor is determined based on the real-time opening status of scenic area facilities. These three factors are parsed from the warning or announcement texts by a large language model and then written into the decision engine through function calls, realizing the automatic conversion of text into carrying capacity parameters.

[0111] After obtaining the passenger flow demand forecast for each future time period, the passenger flow demand forecast for each future time period is decomposed into a predetermined arrival volume and a controllable demand, depending on whether it is subject to quota control, satisfying the following formula (15).

[0112] (15).

[0113] In equation (15), The predetermined arrival volume includes the portion of sold but unredeemed tickets expected to arrive during that period, as well as the arrival volume of free tickets, annual passes, and group tickets that do not occupy the available ticket quota / this release quota. This refers to controllable demand, which consists of individual tourists who have not yet purchased tickets and whose demand can be absorbed by the available ticket quota / this release quota. These two categories are mutually exclusive and exhaustive, and are predicted by visitor flow demand. (Obtained by combining formula (13)) This is obtained by combining the historical composition ratio of sold orders and free ticket redemption.

[0114] Since the quota allocation cannot exceed the demand that can be absorbed in that period, the upper limit of the quota for each period is taken as follows (16).

[0115] (16).

[0116] In equation (16), This serves as the upper limit for quotas in future time periods; The physical limits for gate throughput capacity, security screening throughput, etc. For controllable demand.

[0117] Subsequently, based on the forecast of passenger flow demand for different time periods, and according to the dynamic effective carrying capacity and the distribution of visitor dwell time within the scenic area, the real-time remaining saleable capacity is calculated. Specifically, firstly, based on the visitor count within the scenic area... Duration of stay Distribution, calculate the conditional probability of each tourist leaving the scenic area. Then, based on the probability of each tourist leaving the park under certain conditions, the predicted number of people leaving the park is determined as follows (17).

[0118] (17).

[0119] In equation (17), For the future The predicted number of visitors leaving the park within a given time period; To gather tourists within the scenic area The first in One tourist; The function representing the length of stay for tourists within the scenic area is obtained by fitting historical gate pairing data (entry to exit) to different tourist types (e.g., log-normal distribution).

[0120] Next, based on the dynamic effective carrying capacity Current number of people in the park Predicted number of visitors leaving the park And the portion of sold but unredeemed tickets expected to arrive during that time period out of the predetermined arrival volume. The real-time remaining salable capacity is calculated as shown in the following formula (18).

[0121] (18).

[0122] In formula (18), the current number of people in the park It refers to existing users rather than new users, based on actual access volume. The predicted number of children leaving the park as determined by equation (17) The formula (19) is obtained by recursively applying it to each time period.

[0123] (19).

[0124] Actual admission quantity required for recursion The actual number of people entering the scenic area is determined by the following formula (20).

[0125] (20).

[0126] In equation (20), This refers to the allocation of tickets available for sale in future time slots.

[0127] Finally, based on the real-time remaining available capacity, the available ticket quota for each future time period is optimized and solved with the goal of minimizing the load factor variance for each future time period. Specifically, the actual admission volume for each future time period is determined as the sum of the predetermined arrival volume and the available ticket quota for the corresponding time period (i.e., formula (20)). The available ticket quota for each future time period is solved with the goal of minimizing the load factor variance between the actual admission volume and the dynamic effective carrying capacity for each future time period. See formula (21) below for details.

[0128] (twenty one).

[0129] (twenty two).

[0130] In equations (21)-(22), The target average load factor; The remaining total available quantity for the day is calculated from the real-time remaining available capacity. With adjustable demand To be determined jointly; Indicates future time periods; This represents the total quantity for the time period. It should be noted that this refers to the total remaining available quantity for the day. The quota should not exceed the remaining capacity in any future period (supply-side constraint) nor the sum of all controllable demand (demand-side constraint), so as to avoid releasing ineffective quotas that cannot be sold.

[0131] The available ticket quota for each future time period can be obtained by solving formula (21). Because... and For mutually exclusive groups, Only one term enters the target, formula (21) about It is a strictly convex quadratic programming problem with a single equality constraint plus a box constraint, thus maintaining the KKT closed-form solution in the form of water injection, as shown in the following formula (23).

[0132] (twenty three).

[0133] In equation (23), Indicates truncation to interval Water level Obtained by the bisection method, such that .

[0134] It should be noted that whenever passenger flow demand is predicted... Rolling updates; in this embodiment, updates are set to occur every 10 minutes. , The available ticket quota is recalculated after being updated synchronously by formula (15), which means that the purchase limit is dynamic rather than a fixed number of tickets per day.

[0135] In addition, the available ticket quotas for each future time period obtained from the solution are adjusted upwards according to the historical cancellation rate, as shown in the following formula (24).

[0136] (twenty four).

[0137] In equation (24), To ensure effective allocation of quotas; These are prior parameters; , This represents the historical number of orders and the number of cancelled orders for this period. Cancellation rate. Bayesian smoothing estimation is used to avoid estimation jitter in small sample periods.

[0138] In this embodiment, a target load rate is set and the water level is solved. After allocating the available quota for each time period according to the water injection solution, a constraint check is performed, that is, to determine whether the total quota is conserved and meets the upper and lower limits. If it does not meet the requirements, the water level is solved again. If it meets the requirements, the quota for each time period is output, and finally the peak shaving and valley filling effect is achieved by releasing less water during peak periods and more water during off-peak periods.

[0139] In this embodiment, by decomposing passenger flow demand forecasting into predetermined arrival volume and controllable demand, and calculating the remaining saleable capacity in real time based on dynamic effective carrying capacity and conditional departure probability, the quota allocation can accurately match the actual controllable demand, avoiding ineffective quota allocation. With the goal of minimizing the variance of load rate in each time period, the quota optimization model is solved using a water-filling closed-loop solution, realizing the peak-shaving and valley-filling of passenger flow and the balanced utilization of the scenic area's carrying capacity. Combined with Bayesian smoothing estimation of the cancellation rate and a predictive rolling update mechanism, a prediction-driven dynamic purchase restriction closed loop is formed, further improving the system's adaptability and ticketing efficiency in case of emergencies.

[0140] Based on the above embodiments, the following will further describe in detail the congestion-driven adaptive flow limiting process.

[0141] The system performs heat statistics on public opinion data in text-based data to obtain the public opinion heat index for each time period; it calculates the congestion level for each time period in the future based on the predicted number of people in the park, the dynamic effective carrying capacity, the number of reservations, and the public opinion heat index; and it adaptively adjusts the real-time release rate of ticket purchase requests for the corresponding time period based on the congestion level for each time period in the future.

[0142] What's easy to understand is that it extracts the number of views of content related to the scenic spot within window t. and interaction volume The interaction volume includes likes, comments, reposts and collections, and is first logarithmically standardized as shown in the following formula (25).

[0143] (25) In equation (25), These are the rolling mean and standard deviation over the past 90 days, generated in real time by ES|QL time series aggregation.

[0144] Subsequently, the public opinion heat index is calculated. This index is obtained by combining the level term and the growth rate term and compressing them using the Sigmoid function, as shown in the following formula (26).

[0145] (26).

[0146] In equation (26), The corresponding influence weights for view count level, interaction level, and view count growth rate are as follows: influence weight of growth rate. Used to capture leading signals of content suddenly going viral; This is the Sigmoid activation function.

[0147] Next, based on the predicted number of park visitors, dynamic effective carrying capacity, number of reservations, and public opinion intensity index, the congestion level for each time period in the future is calculated. Specifically, time periods are defined by comprehensively considering the predicted park load, reservation saturation, and public opinion intensity. The degree of congestion is shown in the following formula (27).

[0148] (27).

[0149] In equation (27), The predicted number of people in the park during this period (stock) is obtained by decomposition and recursion of the passenger flow demand forecast, as shown in formula (19) above. This refers to the dynamic effective carrying capacity. This represents the number of reservations already made. This represents the current available ticket quota for this time period; A minimal constant to prevent the denominator from becoming invalid when the quota is zero; This refers to the aforementioned public opinion heat index. This congestion level... This is the core metric for the display and ranking of the ticket purchase interface; These are the weighting coefficients for each crowding factor.

[0150] Finally, based on the congestion levels in future time periods, the real-time release rate of ticket purchase requests for the corresponding time periods is adaptively adjusted. Specifically, ticket purchase requests are rate-limited via a token bucket, and the release rate is adjusted according to the time period. Real-time release rate Its crowding level is adaptively reduced by the Logistic function, as shown in the following formula (28).

[0151] (28).

[0152] In equation (28), This is the maximum allowable passage rate of the system; The target congestion threshold; To control the steepness of the decay. When there is a sudden surge in instantaneous requests and congestion during a certain period. Real-time release rate when approaching saturation The current decreases smoothly and triggers peak-shaving guidance, thereby achieving dynamic current limiting rather than hard circuit breaking.

[0153] In this embodiment, the congestion level for future time periods is calculated by comprehensively predicting the park's load, reservation saturation, and public opinion intensity index. This allows the congestion assessment to fully reflect the real-time operational status of the scenic area and capture the traffic pressure caused by sudden public opinion events. Based on this, the real-time release rate of ticket purchase requests is adaptively adjusted according to the congestion level. When a sudden surge in requests causes the congestion level to approach saturation, the release rate smoothly decreases and peak-shifting guidance is triggered, avoiding the experience interruption caused by traditional hard circuit breakers. This achieves peak shaving and valley filling of ticket purchase traffic and ensures system stability.

[0154] Based on the above embodiments, further, Figure 5 This diagram illustrates the timing of multi-turn dialogue profile extraction and package combination generation provided by an embodiment of the present invention. The following will combine... Figure 5 The process of generating and recommending personalized packages to users is described in detail.

[0155] Tourist profile slots are extracted from multi-turn dialogues using a large language model, and these slots are vectorized to obtain user profile vectors. In response to user query text, candidate products are retrieved from a pre-built product vector index based on the user profile vectors. Using the matching degree between the user profile vectors and each candidate product as the optimization objective, combined with mutual exclusion constraints between products and semantic redundancy penalties, personalized packages are generated through combinatorial optimization. Among them, the discount rate of time-based products in the personalized packages is inversely linked to the congestion level of the corresponding time period.

[0156] It's easy to understand that products such as tickets, sightseeing buses, cruises, guided tours, cultural and creative products, and parking are sold in fixed packages, which fails to understand the true composition and preferences of tourists, such as "elderly people with mobility issues," "traveling with a 5-year-old child," and "wanting to photograph the sunset." Furthermore, it fails to link product combinations with the park's real-time demand and availability.

[0157] In this embodiment, tourists interact with the ticketing agent using natural language. For example, a user's query text might be "Bringing my parents, the elderly have difficulty walking, my child is 5 years old, and we want to see the sunset." The large language model extracts profile slots according to a predefined JSON Schema, such as "{ "party": {"elderly": 2, "child": 1, "adult": 2}, "mobility":"limited", "interests": ["sunset", "photography"], "pace": "slow", "budget_hint":null, "time_window": "afternoon"}.

[0158] No. The state update of the round image is formalized as an incremental state tracking that inherits the state of the previous round, as shown in equation (29).

[0159] (29).

[0160] In equation (29), In portrait mode; For users to query text; This is a summary of the previous system actions. The slot system is designed based on user demographics and experience preferences, serving the generation of product portfolios.

[0161] After the user profile slots are serialized into descriptive text, they are vectorized using the Inference API to obtain user profile vectors. ; Description vector of ticketing products Already stored in the product vector index. Based on user profile vectors. Perform filtered KNN retrieval on the product vector index (with structured filtering based on price, time period, and age, and efficiency guaranteed by filter-aware HNSW graph traversal), recalling Top- Candidate products are selected and their matching degree is calculated as follows (30).

[0162] (30).

[0163] In equation (30), For hard constraint gating, such as mandatory filtering of pure hiking routes for "people with mobility difficulties" and inclusion of sightseeing bus / boat tours, only... Products included in the viable candidate product set To avoid ranking anomalies caused by multiplication gating in scenarios with negative similarity.

[0164] The recalled candidate product set is first dehomogenized using the MMR rule of formula (7).

[0165] Assume decision variables Indicates product Whether or not a personalized package is selected, the optimization goal is the total matching benefit minus the redundancy penalty, as shown in the following formula (31).

[0166] (31).

[0167] In equation (31), The revenue of the i-th product; To reduce semantic redundancy between products, avoid pushing two homogeneous projects; This refers to the abandonment rate; For a set of mutually exclusive product pairs, such as two conflicting travel lines; The upper limit of the budget; Based on the selected entry time slot The time feasibility verification shows that the project opening hours can be linked and the walking / connecting time is accessible.

[0168] For candidate sizes K greater than 50, the preferred option is the above-mentioned package selection method. The candidate size represents the feasible candidate product set. The number of candidate products.

[0169] Candidate size Time-bounded exact solution; in online scenarios, a greedy algorithm with marginal gain is used, and at each step, products selected for the personalized package are chosen using the following formula (32). Then stop.

[0170] (32).

[0171] In equation (32), The optimal product vector index; A set of feasible candidate products; Given a set of selected products S, we want to determine the marginal gain resulting from choosing the i-th product. This represents the conflict penalty coefficient.

[0172] when At that time, the above revenue function with respect to the selected product set It exhibits diminishing returns with submodulus. In online scenarios, a greedy algorithm with marginal gain is used to obtain feasible approximate solutions. If the constraints degenerate into cardinal constraints or standard knapsack constraints and the corresponding submodulus maximization algorithm is used, classical approximation can be applied to guarantee the solution. Under complex mutual exclusion and time feasibility constraints, the system guarantees online executability with a branch-and-bound or feasibility-fixing greedy algorithm, with a single combination taking milliseconds.

[0173] For a specific personalized package, the discount rate for time-based products within the package. It is inversely related to the congestion level during that period, as shown in the following formula (33).

[0174] (33)

[0175] In equation (33), To maximize the discount rate, it is predicted that the more idle the time period, the greater the discount. Finally, a large language model generates an explained package deal script, and the transaction is completed by sequentially calling check_quota, lock_inventory, and create_order. A Redis distributed lock ensures the atomicity of quota deduction.

[0176] In this embodiment, a large language model incrementally extracts tourist profiles from multi-turn dialogues and performs filtered KNN retrieval and MMR dehomogenization to achieve accurate understanding of users' personalized needs and efficient recall of candidate products. Based on this, with matching degree and semantic redundancy penalty as optimization objectives, combined with mutual exclusion constraints and time feasibility verification, a millisecond-level solution using a submodal greedy algorithm or branch-and-bound algorithm generates personalized packages, avoiding homogenized recommendations while ensuring online executability. Simultaneously, the discount rate of time-slot products within the package is inversely linked to the congestion level of that time slot, using price leverage to guide tourists to off-peak periods, achieving synergy between demand-side regulation and peak-shaving at the scenic area. Finally, the large language model generates explained dialogue and automatically completes the transaction loop, improving the intelligence level and conversion efficiency of the ticketing experience.

[0177] Based on the above embodiments, the following will provide a detailed description of the staggered peak guidance process.

[0178] Obtain the user's original intended time period; when the original intended time period is saturated, calculate the guidance utility score of each candidate time period based on the user profile vector, time period feature vector, and the crowding of candidate time periods; determine the recommendation probability of each candidate time period based on the guidance utility score; and recommend alternative time periods to the user based on the recommendation probability of each candidate time period.

[0179] Specifically, when users Original intention period of the request When the conditions are not met or are close to saturation, for candidate time periods The guiding utility score is calculated using the following formula (34).

[0180] (34).

[0181] In equation (34), the three terms are the congestion penalty, the user profile vector, and the user profile vector, respectively. With time period feature vector The semantic matching benefits of embedding attributes such as lighting, project openness, and age-friendliness, and the cost of deviation from the user's original intention time period; , , These are the crowding penalty coefficient, the preference matching weight coefficient, and the time deviation penalty coefficient, respectively.

[0182] Subsequently, the recommendation probability for each candidate time period is determined based on the guiding utility score. The recommendation probability is generated by temperature Softmax, as shown in the following formula (35).

[0183] (35)

[0184] In equation (35), the temperature parameter Dynamically adjusts according to the overall load of the park: the higher the load, The smaller the size, the more focused the guidance will be on periods of low congestion.

[0185] Finally, the large language model translates the ranking results into natural language explanations, such as "The 10:00 session you selected is almost fully booked, while the 11:30 session has better lighting and is expected to have 40% fewer people," in order to improve the acceptance rate of the guidance.

[0186] In this embodiment, the guidance utility score is calculated by combining congestion penalty, user profile matching, and time deviation cost, and the recommendation probability is dynamically adjusted using temperature parameters. Personalized off-peak guidance is achieved when the original intention time period is saturated, effectively diverting users to less congested time periods. At the same time, the temperature parameters are adaptively adjusted according to the overall park load, and the guidance is more concentrated when the load is higher, further enhancing the peak shaving and valley filling effect of the scenic area.

[0187] Based on the above embodiments, the workflow of the operation agent and the maintenance agent will be described in detail below.

[0188] The scenic area ticketing management is carried out collaboratively by the operation intelligence agent and the maintenance intelligence agent; the operation intelligence agent is responsible for business-side control, while the maintenance intelligence agent is responsible for system-side support; the operation intelligence agent and the maintenance intelligence agent achieve linkage by sharing state indexes and exchanging events.

[0189] In one specific embodiment Figure 6 This diagram illustrates the workflow of the operational intelligent agent provided in an embodiment of the present invention. The following will be combined with... Figure 6 The workflow for operating intelligent agents is explained.

[0190] The operational intelligence agent operates in a perception-decision-action cycle: the perception layer subscribes to the ES shared state index (predicted values, actual passenger flow, load factor, revenue, public opinion heat, coupon redemption flow); the decision layer is completed by LLM combined with the following mathematical model; the action layer calls the operational toolset (coupon issuance, price adjustment suggestions, message push, report generation, parameter recalibration) through function calls.

[0191] Regarding prediction bias monitoring and recalibration triggering: For each completed period, the standardized prediction residual is calculated as the prediction bias, and the variance is estimated using an exponentially weighted sliding method, as shown in equation (36).

[0192] (36)

[0193] In equation (36), For the stability term, the standardized residuals are first calculated using the volatility estimate from the previous period, and then updated with the current residuals to prevent outliers from artificially inflating the denominator and masking warnings. When two or more consecutive periods... When the operation agent automatically triggers the recalibration of the fusion weight of formula (12) and the crowding weight of formula (27), and writes the deviation scenario (feature + inference chain + deviation magnitude) back to the case library for subsequent RAG retrieval to avoid similar errors.

[0194] Optimization of campaign targeting for peak-hour promotions. When predicting load factor for a specific time period... The operational intelligence agent automatically generates off-peak promotion plans. Assuming a specific time period... Targeted delivery Zhang's discount rate is Off-peak coupons, coupon redemption conversion rate Modeling based on Logistic elasticity is as follows (37).

[0195] (37)

[0196] In equation (37), Use the Sigmoid activation function; For bias terms; The weighting coefficient for the impact of the discount rate; The weighting coefficient for the impact of crowding; For congestion level.

[0197] Expected incremental passenger flow is With the goal of maximizing marginal contribution versus subsidy cost, the solution is obtained under the constraints of load ceiling and marketing budget, as shown in equation (38). (38).

[0198] In equation (38), This is the marginal contribution per customer (including secondary consumption). The face value; For marketing budget; The set of off-peak periods. Formula (38) is a nonlinear constraint optimization with integer variables. When implementing it, the discount rate can be discretized and solved by branch and bound, dynamic programming or heuristic search. This mechanism complements the off-peak pricing of the package in formula (33) and together constitutes a demand-side valley filling tool. Formula (21) shaving the peak and formula (38) filling the valley, the passenger flow adjustment is bidirectional.

[0199] Regarding Natural Language Operational Queries and Automated Reporting (Text-to-ES|QL): Operations personnel ask questions in natural language, such as "Why were the number of visitors to the park through the West Gate higher than predicted last Saturday afternoon?" The operations agent translates the questions into ES|QL pipeline queries, executes them on the feature view and case library, and combines them with the inference chain for the day. Generate attribution explanations; daily / weekly reports (customer flow, revenue, forecast accuracy, promotional effectiveness) are automatically generated according to planned tasks, without the need for manual query writing throughout the process.

[0200] In another specific embodiment, Figure 7 This diagram illustrates the workflow of the operation and maintenance intelligent agent provided in an embodiment of the present invention. The following will be combined with... Figure 7 The workflow of the operation and maintenance intelligent agent is explained.

[0201] The operation and maintenance intelligent agent ensures the availability of the closed-loop system itself, and the objects of perception are cluster and service telemetry (ES query latency, JVM heap, thread pool queue, LLM inference endpoint latency and error rate, Redis rate limiting queue length) and system logs (templated and vectorized into ES via semantic_text).

[0202] First, regarding anomaly detection for each telemetry indicator. Maintain the index-weighted moving baseline and volatility, and calculate the standardized deviation as shown in equation (39).

[0203] (39)

[0204] Standardized deviation If the threshold is exceeded, a candidate alarm will be generated. Similarly, after alarm calculation, alarms are estimated and updated using an exponential weighting method. Multiple indicator alarms are aggregated and deduplicated within a time window to avoid immediately polluting the baseline by current anomalies.

[0205] Secondly, regarding log novelty detection. After template extraction and vectorization, the logs ( ), and normal template library For comparison, the novelty is taken as the complement of the maximum similarity, as shown in the following formula (40).

[0206] (40)

[0207] High-novelty logs (error patterns that have never appeared before) go directly into the root cause analysis queue, and this detection does not rely on pre-defined keyword rules.

[0208] Secondly, regarding root cause ranking. Candidate component set. The service call topology and its dependent facilities derived from the ticket purchase link are automatically maintained by service registration and link tracing data, specifically including: 1) access gateway and load balancer; 2) ticketing service; 3) inventory and quota service (including Redis quota counting, distributed lock and token bucket rate limiting components); 4) order and payment service (including third-party payment / channel callback); 5) verification and gate access service; 6) Kafka message queue and Flink stream processing job; 7) Elasticsearch search and vector retrieval cluster (including inference embedding / reordering endpoints); 8) LLM inference endpoint and configurable LLM adapter; 9) prediction service and decision engine; 10) AI operation intelligence agent and AI operation and maintenance intelligence agent itself; as well as the host nodes, container instances and network links on which the above components depend.

[0209] alarm timing Candidate components in front and back windows The scores are based on a combination of temporal proximity, call topology distance, and log semantic relevance, as shown in equation (41).

[0210] (41).

[0211] Top candidates, along with indicator curves and log fragments, are constructed as prompt words. The LLM outputs root cause conclusions and handling suggestions, and executes the contingency plan through function calls (restarting instances, switching traffic, clearing queues, and rolling back configurations). High-risk actions are transferred to manual approval, while the rest are executed automatically and recorded for auditing.

[0212] Regarding predictive scaling. Predicting the ticket purchase request rate using the main closed loop. (Similar to passenger flow forecasting, the signal of a surge in public opinion is also transmitted to this point in advance), and the number of copies of the ticket purchase service is adjusted in advance as follows (42).

[0213] (42).

[0214] in Average resource cost per request; The resource capacity that a single copy can provide per unit of time; The target utilization rate is [not specified]. Capacity expansion occurs before, rather than after, peak loads, directly reflecting how business forecasting supports system operations and maintenance.

[0215] Regarding health-driven automatic degradation: The health of the large model inference link is defined as the normalized latency. Error rate Timeout rate The weighted complement is as follows (43).

[0216] (43)

[0217] when At that time, the operation and maintenance intelligent agent automatically performs degradation: the fusion weights of equation (12) are... Zeroing (predicted rollback statistical baseline prediction), dialogue package rollback rule template, staggered peak guidance rollback static sorting; After recovery, the system gradually increases according to the hysteresis interval to avoid jitter. Therefore, the large-scale model failure does not affect the main ticketing process.

[0218] Regarding the dual-agent blackboard collaboration: The operation agent and the maintenance agent exchange events through the shared state index (blackboard mechanism) of ES: the maintenance agent's degradation event will notify the operation agent to suspend promotional campaigns (to avoid increasing volume during the degradation period); the overall operation's promotion / valley filling plan will inform the maintenance agent in advance to trigger (42) pre-expansion, forming a two-way linkage between business and system.

[0219] In this embodiment, the operations intelligence agent automatically triggers model recalibration based on prediction deviation monitoring and optimizes passenger flow during off-peak hours using peak-hour promotions. Simultaneously, it achieves zero-code data analysis through natural language query conversion, effectively improving operational efficiency. The maintenance intelligence agent quickly locates faults through anomaly detection, log novelty analysis, and root cause sorting. It also performs pre-emptive predictive scaling based on business forecasts, combined with a health-driven automatic degradation mechanism, ensuring that the main ticketing process remains unaffected by large model failures. The two intelligence agents collaborate via a shared state index, enabling bidirectional linkage between business-side promotional plans and system-side degradation events. This avoids ineffective traffic generation during degradation periods and reserves resources in advance for marketing activities, ensuring system stability and business continuity.

[0220] Based on the above embodiments, further, Figure 8 This diagram illustrates the closed-loop feedback and online calibration of the scenic area ticketing scheduling method provided in this embodiment of the invention. The following will combine... Figure 8 The closed-loop feedback and online calibration process for system parameters is described in detail.

[0221] It is easy to understand that all the key parameters of the system are self-calibrated online, forming a data flywheel.

[0222] Firstly, regarding forecast calibration: at the end of each time period, actual passenger flow will be used. The error of updated equation (11) and the fusion weight of equation (12) Crowding weight To minimize predicted crowding Compared with the measured load The squared loss of the deviation is used for small-step gradient updates, as shown in equation (44).

[0223] (44)

[0224] in, The congestion weight is to be updated; To project onto the probability simplex, ensuring that the weights are non-negative and normalized; The learning rate; This is the gradient operator.

[0225] Secondly, regarding guidance calibration: the acceptance / rejection of off-peak guidance is used as a label, and the utility coefficient is updated by online binary logistic regression. As shown in equation (45).

[0226] (45)

[0227] in, The conversion rate for redemption of off-peak coupons; It is the sigmoid activation function; The guidance effectiveness score for user i during time period j; The utility coefficient is to be updated; The learning rate; Label the actual conversion behavior of user i during time period j, such as accepting / rejecting off-peak guidance; This is the gradient operator.

[0228] Thirdly, regarding package calibration: Package conversion rate and cancellation / modification rate are fed back to ES, and product vectors and redundancy penalty coefficients are periodically recalculated. .

[0229] Fourthly, regarding the increase in the case library: at the end of each operating day, the day's features + actual customer flow + large model inference chain will be added. Prediction bias is written back to the ES case library (the semantic_text field is automatically vectorized and entered into the library) as subsequent RAG text evidence. The predictive ability of the large model continues to improve with the accumulation of operational data, without the need to retrain the model ontology.

[0230] In other embodiments, Figure 9 The diagram shows the overall architecture of the scenic area ticketing scheduling method provided in this embodiment of the invention.

[0231] like Figure 9As shown, the data acquisition layer gathers data from multiple sources, including turnstiles, ticketing records, weather and traffic APIs, and short video public opinion, and transmits it to the storage and indexing layer. The storage and indexing layer, based on an Elasticsearch 9.4.x cluster, utilizes time-series data streams, inverted indexes, and vector indexes to achieve efficient data storage and management. The vectorized and hybrid retrieval layer processes query requests through inference API embedding and a hybrid BM25 and KNN retrieval technology, then passes the data to the LLM inference layer for multimodal passenger flow prediction, congestion inference, and function calls. Building upon the inference layer, the operations intelligence agent is responsible for monitoring prediction deviations and optimizing promotions, the decision engine layer executes dynamic carrying capacity models and quota optimization, and the operations and maintenance intelligence agent performs system anomaly detection and predictive scaling. Finally, the interactive application layer provides services to users through a ticketing client, a multi-turn dialogue intelligence agent, and a personalized package generator. All layers collaborate through data interaction and feedback mechanisms, combined with off-peak guidance and dynamic pricing strategies, to achieve intelligent regulation of scenic area resources.

[0232] Based on the daily reception limit Taking a 5A-level scenic area with 150,000 visitors as an example, this paper provides an example of the parameters involved in implementing the scenic area ticketing scheduling method.

[0233] Regarding the Elasticsearch cluster: Version 9.4.3; 3 master nodes + 6 data nodes. Gate / ticket sales data is written to TSDS (index.mode:time_series), downsampled to hourly granularity after 30 days, and to daily granularity after 180 days; public opinion / announcements / product descriptions use the semantic_text field (default inference endpoint, BFLOAT16 encoding); product and case vectors are dense_vector, d=1024, index_options.type:bbq_hnsw (for case libraries exceeding memory size, shards use bbq_disk), HNSW m=16, ef_construction=200, ef_search=100; mixed retrieval uses a retriever tree: RRF (standard BM25+KNN) → text_similarity_reranker; evidence top- RRF MMR .

[0234] Regarding the feature pipeline: ES|QL View fixed feature queries (passenger flow TSDS LOOKUP JOIN weather dimension table, holiday dimension table), and the prediction service retrieves frames from the view every 10 minutes.

[0235] Regarding prediction: LLM self-consistent sampling Temperature 0.7; Fusion parameters , Dual-seasonal index smoothing cycle (Ticket sales period for 30 minutes within the day) (Corresponding time period within a week).

[0236] Regarding crowding: Initial weights Online calibration using formula (27); target threshold .

[0237] Regarding quotas: Time granularity 30 minutes ( ); Target load factor A priori rejection (Priority abandonment rate 5%); Traffic limiting Requests / second .

[0238] Regarding staggered peak hours guidance: initial temperature .

[0239] Regarding packages: Top Candidates Redundancy penalty Maximum discount .

[0240] Regarding operational intelligent agents: Deviation triggering conditions Two consecutive time periods ( ); Lower limit of valley filling Marketing budget 10,000 yuan / day; elasticity coefficient Daily online logical regression updates; daily reports are automatically generated at 22:00 every day.

[0241] Regarding the Operation and Maintenance Intelligent Agent: EWMA Metric Alarm threshold Log novelty threshold: 0.35; Root cause weighting ; Lead time for expansion and contraction Minutes, target utilization Health (Hysteresis recovery band 0.6–0.75).

[0242] Typical operation process: A short video went viral that night (views increased by 320% in 6 hours) → Equation (26) heat index rises to 0.91 → Mixed search recalls similar historical cases and hot original texts → LLM predicts that demand will increase by 1.8 times and confidence level of 0.83 from 10:00 to 12:00 the next day → Equation (13) automatically increases λ to 0.71 → Equation (23) recalculates quota, reduces quota by 35% during peak hours, and shifts total to the period after 14:00 → When ticket purchase requests surge the next morning, Equation (28) limits speed and guides 41% of users to reschedule for the afternoon according to Equation (35) → Dialogue with ticket purchase users to get an 80% off off-peak package including afternoon cruise + explanation → The actual peak load rate on the day is 0.78, and the total number of visitors increases by 12% throughout the day. During the period, the dual-agent action was as follows: the maintenance agent predicted the ticket purchase service copy from 8 to 20 at 09:45 based on the request rate, and there was zero flow restriction timeout during the peak period; the operation agent detected that the predicted load of 0.31 after 16:00 was lower than the valley filling lower limit, and after solving according to formula (38), 12,000 15% discount off-peak coupons were targeted to users matched with the profile, with a redemption rate of 38%, and the valley load rose to 0.52; at noon, the large model endpoint jittered once, causing the health to drop to 0.55, and the overall maintenance automatically downgraded within 3 seconds and recovered according to the hysteresis band after 3 minutes, with zero interruption of the main ticketing link. During the downgrade period, the operation agent automatically suspended the promotion delivery.

[0243] The scenic area ticketing scheduling method provided by this invention has the following advantages.

[0244] First, prediction accuracy and leading nature: Multimodal reasoning prediction can understand public opinion, weather and announcement signals in text form and has the ability to capture sudden passenger flow; the adaptive fusion of formula (13) automatically balances cost and accuracy between regular days and sudden days; Second, passenger flow balance: the quota injection + adaptive flow restriction + utility-based staggered peak guidance of formulas (21), (28), (34)-(35) will bring the control action forward to the time when the ticket purchase occurs, and reduce the peak and fill the valley from the source. Thirdly, experience and benefits: Conversational profiling, combined optimization, and off-peak pricing simultaneously improve per capita consumption and utilization during off-peak hours; Fourthly, the project is feasible and cost-controllable: TSDS+ downsampling reduces time-series storage costs; BBQ / DiskBBQ reduces vector memory usage to about 1 / 32, supporting tens of millions of product and case vectors; semantic_text eliminates the need for external vectorization pipelines; the retriever framework completes hybrid retrieval and reordering with a single request; closed-loop water-filling solution and greedy combination are both millisecond-level; large models are only called during prediction scrolling cycles and dialogue rounds. Fifth, explainability and traceability: Reasoning Chain E enables each quota adjustment and recommendation to be explained to regulators and operators, and is stored as a searchable case asset; Sixth, autonomous operation and fault self-healing: The operation intelligence agent not only enables peak shaving but also has the ability to actively fill valleys and automatically protect the prediction quality; the maintenance intelligence agent uses predictive scaling to deal with ticket purchase peaks in advance and uses health degradation to ensure zero interruption of the main ticketing link when the large model fails, so that the system can operate autonomously without human intervention.

[0245] Corresponding to the scenic area ticketing scheduling methods described in the above embodiments, the present invention also provides a scenic area ticketing scheduling system.

[0246] Specifically, Figure 10 A schematic diagram of the scenic area ticketing scheduling system provided in an embodiment of the present invention is shown.

[0247] like Figure 10 As shown, the system includes: a multi-source data acquisition and text evidence generation module 1010, used to acquire numerical time-series data and text data of the scenic area, and to obtain statistical baseline prediction results for each future time period based on the numerical time-series data, and to obtain text evidence for each future time period based on the text data; a large model prediction and passenger flow demand fusion module 1020, used to output passenger flow prediction values ​​for each future time period based on a large language model, according to the numerical time-series data, the statistical baseline prediction results and the text evidence, and to fuse the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain passenger flow demand predictions for each future time period; and a dynamic capacity calculation and quota optimization module 1030, used to calculate the real-time remaining saleable capacity based on the passenger flow demand predictions for each future time period, according to the dynamic effective carrying capacity and the distribution of tourist dwell time in the scenic area, and to optimize the saleable ticket quota for each future time period based on the real-time remaining saleable capacity with the goal of minimizing the load rate variance for each future time period.

[0248] In this embodiment, the multi-source data acquisition and text evidence generation module 1010 acquires numerical time-series data and text data of the scenic area, and obtains statistical baseline prediction results for each future time period based on the numerical time-series data, and obtains text evidence for each future time period based on the text data; the large model prediction and passenger flow demand fusion module 1020, based on a large language model, outputs passenger flow prediction values ​​for each future time period based on the numerical time-series data, statistical baseline prediction results, and text evidence, and fuses the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain passenger flow demand predictions for each future time period; the dynamic capacity calculation and quota optimization module 1030, based on the passenger flow demand predictions for each future time period, calculates the real-time remaining saleable capacity based on the dynamic effective carrying capacity and the distribution of tourist dwell time in the scenic area, and optimizes the saleable ticket quota for each future time period based on the real-time remaining saleable capacity with the goal of minimizing the load rate variance for each future time period. This system improves the accuracy of passenger flow demand forecasting and the ability to respond to text signals such as sudden public opinion events by integrating numerical time-series data and textual data and using large language models for logical reasoning prediction. By combining dynamic effective carrying capacity and the distribution of visitor dwell time in real time to calculate the remaining saleable capacity, the capacity assessment is upgraded from static setting to dynamic prediction. Finally, the system optimizes the ticket quota solution with the goal of minimizing the variance of the load rate in each time period, which can smoothly distribute passenger flow demand throughout the day and effectively achieve peak shaving and valley filling and balanced utilization of the scenic area's carrying capacity.

[0249] Figure 11 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 11 As shown, the electronic device may include a processor 1110, a communications interface 1120, a memory 1130, and a communication bus 1140. The processor 1110, communications interface 1120, and memory 1130 communicate with each other via the communication bus 1140. The processor 1110 can call logical instructions from the memory 1130 to execute a scenic area ticketing scheduling method.

[0250] Furthermore, the logical instructions in the aforementioned memory 1130 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part 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 the present invention. 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.

[0251] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the scenic spot ticketing scheduling method provided in the above embodiments.

[0252] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the scenic area ticketing scheduling method provided in the above embodiments.

[0253] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0254] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0255] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for scheduling ticketing at a scenic area, characterized in that, include: The system acquires numerical time-series data and text data of the scenic area, and obtains statistical baseline prediction results for each future time period based on the numerical time-series data, and obtains text evidence for each future time period based on the text data. Based on the large language model, according to the numerical time series data, the statistical baseline prediction results and the text evidence, the predicted passenger flow values ​​for each future time period are output, and the predicted passenger flow values ​​for each future time period are fused with the statistical baseline prediction results to obtain the predicted passenger flow demand for each future time period. Based on the predicted passenger flow demand for each future time period, the real-time remaining available capacity is calculated according to the dynamic effective carrying capacity and the distribution of visitor dwell time in the scenic area. Based on the real-time remaining available capacity, the available ticket quota for each future time period is optimized and solved with the goal of minimizing the variance of the load rate for each future time period.

2. The scenic area ticketing scheduling method according to claim 1, characterized in that, The method, based on a large language model, outputs predicted passenger flow values ​​for future time periods according to the numerical time series data, the statistical baseline prediction results, and the textual evidence, including: Construct structured cue words, which include the numerical time series data, the statistical baseline prediction results for each future time period, the textual evidence for each future time period, and output constraints; The structured prompts are input into the large language model, which outputs the predicted passenger flow values ​​for each future time period and the corresponding natural language inference chain.

3. The scenic area ticketing scheduling method according to claim 1, characterized in that, By integrating the predicted passenger flow values ​​for each future time period with the statistical baseline prediction results, passenger flow demand forecasts for each future time period are obtained, including: Based on the exponential smoothing relative error and confidence level predicted by the large language model, and the exponential smoothing relative error predicted by the statistical baseline, the fusion weights are determined online. Based on the fusion weights, the predicted passenger flow values ​​for each future time period and the statistical baseline prediction results for each future time period are weighted and fused to obtain the predicted passenger flow demand for each future time period.

4. The scenic area ticketing scheduling method according to claim 1, characterized in that, The dynamic effective carrying capacity is calculated based on the basic carrying capacity, weather attenuation factor, emergency event factor, and facility opening rate factor. The weather attenuation factor is obtained by parsing weather warning texts, the emergency event factor is obtained by parsing control announcement texts, and the facility opening rate factor is determined based on the real-time opening status of scenic area facilities.

5. The scenic area ticketing scheduling method according to claim 1, characterized in that, The process of obtaining passenger flow demand forecasts for future time periods then includes: Depending on whether it is subject to quota control, the predicted passenger flow demand for each future time period is decomposed into a predetermined arrival volume and a controllable demand. The predetermined arrival volume includes the portion of sold but unredeemed tickets expected to arrive during that period, as well as the arrival volume of free tickets, annual passes, and group tickets that do not occupy the available ticket quota. The adjustable demand refers to the demand of individual tourists who have not yet purchased tickets and can be absorbed by the available ticket quota.

6. The scenic area ticketing scheduling method according to claim 5, characterized in that, The calculation of the real-time remaining saleable capacity based on the predicted passenger flow demand for each future time period, according to the dynamic effective carrying capacity and the distribution of visitor dwell time within the scenic area, includes: Based on the distribution of visitor stay duration within the scenic area, calculate the conditional departure probability of each visitor within the scenic area; Based on the probability of each tourist leaving the scenic area under certain conditions, the predicted number of people leaving the scenic area is determined. The real-time remaining available capacity is calculated based on the dynamic effective carrying capacity, the current number of people in the park, the predicted number of people leaving the park, and the portion of the sold but unredeemed tickets in the predetermined arrival volume that is expected to arrive during that time period.

7. The scenic area ticketing scheduling method according to claim 6, characterized in that, Based on the real-time remaining available ticket capacity, and with the objective of minimizing the variance of the load rate for each future time period, the available ticket quota for each future time period is optimized and solved, including: The actual number of admissions for each future time period will be determined as the sum of the predetermined arrival volume and the available ticket quota for the corresponding time period; With the objective of minimizing the variance of the load rate between the actual admission volume and the dynamic effective carrying capacity in each future time period, the available ticket quota for each future time period is solved. The sum of the available ticket quotas for each time period in the future shall not exceed the real-time remaining available capacity, and shall not exceed the controllable demand for the corresponding time period.

8. The scenic area ticketing scheduling method according to claim 1, characterized in that, Also includes: The public opinion data in the text data is statistically analyzed to obtain the public opinion heat index for each time period; Based on the predicted number of people in the park, the dynamic effective carrying capacity, the number of reservations, and the public opinion heat index, the congestion level for each time period in the future is calculated. Based on the congestion levels of each future time period, the real-time release rate of ticket purchase requests for the corresponding time period is adaptively adjusted.

9. The scenic area ticketing scheduling method according to claim 1, characterized in that, The scenic area ticketing management is carried out collaboratively by an operational intelligent agent and an operation and maintenance intelligent agent; wherein, the operational intelligent agent is responsible for business-side control, and the operation and maintenance intelligent agent is responsible for system-side support; The operational intelligence agent and the maintenance intelligence agent achieve linkage by exchanging events through shared state indexes.

10. A scenic area ticketing and dispatching system, characterized in that, include: The multi-source data acquisition and text evidence generation module is used to acquire numerical time-series data and text data of the scenic area, and to obtain statistical baseline prediction results for future time periods based on the numerical time-series data, and to obtain text evidence for future time periods based on the text data. The large model prediction and passenger flow demand fusion module is used to output passenger flow prediction values ​​for each future time period based on the large language model, the numerical time series data, the statistical baseline prediction results and the text evidence, and to fuse the passenger flow prediction values ​​for each future time period with the statistical baseline prediction results to obtain the passenger flow demand prediction for each future time period. The dynamic capacity calculation and quota optimization module is used to calculate the real-time remaining saleable capacity based on the predicted passenger flow demand for each future time period, according to the dynamic effective carrying capacity and the distribution of tourist stay duration within the scenic area, and optimize the saleable ticket quota for each future time period based on the real-time remaining saleable capacity with the goal of minimizing the load rate variance for each future time period.

Images