Tourism demand prediction method and system based on large language model multi-agent cooperation

Abstract

The application relates to the technical field of intelligent agents, and provides a tourism demand prediction method based on a large language model and multi-agent cooperation, wherein through steps S1-S4, a data collection intelligent agent decomposes a prediction target into a plurality of task sequences by relying on a large language model, and outputs a data set after reasoning and planning a path, collecting action data, and observing and verifying data quality. The data set is converted into a standardized feature set by a data processing intelligent agent. A feature recommendation intelligent agent screens a core feature set. The core feature set is input into a prediction execution intelligent agent, a deep learning architecture is selected according to the data scale and the spatiotemporal correlation, training and evaluation are completed, and finally, a tourism demand prediction result is output. The application also discloses a system. The method and the system can work cooperatively through intelligent agents, each intelligent agent combines specific technologies and algorithms, the feature recommendation and the prediction accuracy are continuously optimized through a multi-agent cooperation mechanism, and meanwhile, the knowledge base is introduced to enhance the explainability of the prediction process.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent technology, and in particular to a method and system for predicting tourism demand based on multi-agent collaboration using a large language model. Background Technology

[0002] Tourism demand forecasting is a core basis for the recovery of the tourism industry, the optimization of resource allocation, and the formulation of industry policies. It is of great significance for improving operational efficiency, reducing market risks, and ensuring the high-quality development of the industry. With a significant increase in public awareness of health risks and structural changes in travel decisions and consumption behavior, tourism demand forecasting is gradually shifting towards a complex paradigm that integrates multi-source heterogeneous data such as search popularity, social media sentiment, public health risks, traffic and weather data, and market dynamics. This places higher demands on data integration efficiency, dynamic perception capabilities, and model interpretability.

[0003] CN118429144A discloses a tourism demand prediction method based on empirical mode decomposition and artificial intelligence models. The method includes the following steps: Step S1, using empirical mode decomposition to generate N modal components (IMFs), where N is a natural number greater than 1; Step S2, constructing a spatiotemporal graph convolutional model, which includes two spatiotemporal graph convolutional blocks and one output layer; each spatiotemporal graph convolutional block includes one spatial convolutional block and two temporal convolutional blocks; each temporal convolutional block includes one one-dimensional causal convolution and one gated linear unit; the temporal convolutional block uses a convolution with a kernel width of Kt to extract Kt-1 temporal features. For each node in the IMF, the temporal convolution searches the Kt neighbors of the input element. In step S2, the sequence length is shortened by Kt-1 each time to extract the second higher-order features in the time domain of the modal component IMF; the distance matrix W required for spatiotemporal graph convolution is calculated based on the N sets of modal component IMFs, and the spatial domain convolution block extracts the first higher-order features in the spatial domain of the modal component IMF respectively; in step S3, a gated recurrent unit is used to extract the low-dimensional temporal features in the N sets of modal component IMFs; in step S4, the first higher-order features, the second higher-order features and the low-dimensional temporal features in step S3 are input into the Transformer model to calculate the first predicted data, the first higher-order features, the second higher-order features and the low-dimensional temporal features in step S3 are input into the Long Short-Term Memory network model to calculate the second predicted data, and the first predicted data and the second predicted data are weighted to calculate the third predicted data.

[0004] To cope with complex, dynamic, and high-dimensional market environments, existing forecasting frameworks mostly use static rules for feature identification and factor selection, which cannot dynamically capture the real-time impact of sudden factors such as infectious disease fluctuations, public emergencies, public opinion hotspots, and policy adjustments on tourism demand, resulting in insufficient forecast timeliness and robustness. Summary of the Invention

[0005] Long-term practice has shown that tourism demand data is multi-source data, exhibiting a fragmented and heterogeneous distribution. The formats, definitions, and update frequencies of data from different sources vary significantly, making it difficult to achieve efficient real-time integration and unified data alignment, and compromising data availability and consistency. Data collection, cleaning, labeling, and normalization processes heavily rely on manual intervention, which is time-consuming, labor-intensive, and prone to introducing human error.

[0006] In view of this, the present invention provides a tourism demand prediction method based on multi-agent collaboration using a large language model, comprising: Step S1: In the data acquisition agent, the large language model decomposes the tourism demand prediction target into a task sequence that includes at least basic time series data, public sentiment and willingness, and external risk environment. During the reasoning phase, the operation path is planned based on the task context; During the action phase, structured and unstructured data are obtained by calling the API selector or automated crawling service through the model context protocol; During the observation phase, the built-in quality assessment module verifies the data's integrity, format consistency, and time coverage in real time. If the assessment is satisfactory, the dataset Draw is output. Step S2: The data processing agent transforms the dataset Draw into a standardized feature set and outputs a feature set Dprocessed that includes at least quantitative sentiment indicators, meteorological indicators, and public health-derived indicators. Step S3: The feature recommendation agent will use a dual mechanism of statistical algorithms and expert evaluation of large models to select the core feature set. ; Step S4, the core feature set Input the prediction execution agent, and the prediction execution agent predicts the core feature set. Based on the data scale and spatiotemporal correlation characteristics, we select a deep learning architecture from the built-in model library for training and evaluation to predict and output tourism demand.

[0007] Preferably, in step S1, if the built-in quality assessment module fails the assessment, the feedback result is used as an environmental variable to re-enter the reasoning stage, triggering the self-correction of the crawling strategy.

[0008] Preferably, in the data processing agent, the basic data preprocessing is completed by calling standardized tools to perform missing value imputation, deduplication, and frequency resampling; for large-scale unstructured text, the semantic analysis interface is called to convert sentiment into quantitative indicators based on custom evaluation prompts.

[0009] Preferably, the feature-based formula is retrieved from the knowledge base, and the logical transformation is achieved using the built-in code generation engine.

[0010] Preferably, in step S3, the maximum correlation minimum redundancy algorithm mRMR(Xi) is used to screen out features that have strong explanatory power for the target variable and are complementary to each other. in, For the first One characteristic variable whose impact needs to be evaluated; The core target variable; The size of the candidate feature set; For characteristic variables With target variable The correlation coefficient between them; For characteristic variables and The correlation coefficient between them.

[0011] Preferably, by utilizing the dependency between mutual information quantification features and the prediction target, nonlinear key features that strongly contribute to the target variable are identified by measuring the shared information I(X;Y) between variables; in, For variables and The joint probability density; For variables and The marginal probability density.

[0012] Preferably, a candidate feature subset Dcandidate is extracted from the feature set Dprocessed, and a comprehensive score is obtained by reasoning through a thought chain for each feature in the candidate feature subset Dcandidate. Here, SG represents the scoring dimension, R represents relevance, used to measure the fit between evidence and feature topic; S represents support, used to assess the strength of theoretical support; U represents utility, used to measure the information gain of the feature in improving model accuracy; the weight wG can be configured according to specific task scenarios, used to assign importance weights with theoretical support to each feature; the calculation results are mapped to the interval [0, 100] through a linear normalization function σ(x), and are prioritized from largest to smallest according to the calculation results, selecting a preset proportion of high-scoring items to form the core feature set. .

[0013] Preferably, in step S4, the Data Readiness Index (DCEI) is used to perform a self-check on data quality. in, For coverage, This is used to assess the coverage of core business dimensions; For novelty, This is used to quantify the ability to discover new risk characteristics; For redundancy, Used to identify the independence of feature sets. These are preset weights, used as criteria for determining data quality admission.

[0014] This invention also discloses a system for predicting tourism demand based on a large language model and multi-agent collaboration, as described above. The system includes... The data acquisition agent is used to decompose the tourism demand forecasting target into a task sequence that includes at least basic time series data, public sentiment and willingness, and external risk environment; it can use the ReAct framework to form a cyclical process of reasoning, action and observation phases. The inference phase is used to plan the operation path based on the task context; the action phase is used to call the API selector or automated crawling service through the model context protocol to obtain structured and unstructured data; the observation phase is used by the built-in quality assessment module to perform real-time verification of the data's integrity, format consistency and time coverage. If the assessment is qualified, the dataset Draw is output. A data processing agent is used to transform the dataset Draw into a standardized feature set, and output a feature set Dprocessed that includes at least quantitative sentiment indicators, meteorological indicators, and public health-derived indicators. Feature recommendation agents are used to select core feature sets through a dual mechanism of statistical algorithms and expert evaluation using large models. ; Predictive execution agents, used to predict based on core feature sets Based on the data scale and spatiotemporal correlation characteristics, we select a deep learning architecture from the built-in model library for training and evaluation to predict and output tourism demand.

[0015] The present invention also discloses a machine-readable storage medium storing instructions for causing a machine to execute the tourism demand prediction method based on a large language model and multi-agent collaboration as described in any of the preceding claims.

[0016] This invention provides a tourism demand prediction method based on a large language model and multi-agent collaboration. Through steps S1-S4, the data collection agent, relying on the large language model, decomposes the prediction target into task sequences such as basic time-series data, public sentiment and willingness, and external risk environment. After reasoning and planning the path, collecting data through action, and observing and verifying data quality, it outputs a dataset. The data processing agent transforms this dataset into a standardized feature set containing quantitative sentiment, meteorological, and public health-related indicators. A feature recommendation agent, combining statistical algorithms and expert evaluation of the large model, selects a core feature set. The core feature set is then input into the prediction execution agent, which selects an appropriate deep learning architecture based on the data scale and spatiotemporal relevance to complete training and evaluation, ultimately outputting the tourism demand prediction result. This invention also discloses a system, method, and system that, through the collaborative work of a data acquisition agent, a data processing agent, a feature recommendation agent, and a prediction execution agent, decomposes the complex task of tourism demand forecasting into four stages: perception, cognition, decision-making, and execution. This achieves automated processing of complex tourism demand forecasting tasks. Each agent, combined with specific technologies and algorithms, continuously optimizes feature recommendation and prediction accuracy through a multi-agent collaborative mechanism, automating the entire lifecycle from raw data capture to final prediction output. While improving prediction accuracy, the invention also enhances the interpretability of the prediction process by introducing a knowledge base, ensuring that the model can dynamically adapt to market fluctuations in the post-pandemic era and provide interpretable decision-making basis. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram illustrating the process of tourism demand prediction based on multi-agent collaboration using a large language model, according to one embodiment of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0019] To address the challenges of current tourism demand data being multi-source, fragmented, and heterogeneous, with significant differences in format, statistical definitions, and update frequency among different sources, efficient real-time integration and alignment are difficult to achieve, compromising data availability and consistency. Furthermore, data collection, cleaning, labeling, and normalization processes heavily rely on manual intervention, which is not only time-consuming and labor-intensive but also prone to introducing human error. This invention proposes a tourism demand prediction method based on a large language model and multi-agent collaboration, such as... Figure 1 As shown, the tourism demand prediction method based on multi-agent collaboration using a large language model includes: Step S1: In the data acquisition agent, the large language model decomposes the tourism demand prediction target into a task sequence that includes at least basic time series data, public sentiment and willingness, and external risk environment. During the reasoning phase, the operation path is planned based on the task context; During the action phase, structured and unstructured data are obtained by calling the API selector or automated crawling service through the model context protocol; During the observation phase, the built-in quality assessment module verifies the data's integrity, format consistency, and time coverage in real time. If the assessment is satisfactory, the dataset Draw is output. Step S2: The data processing agent transforms the dataset Draw into a standardized feature set and outputs a feature set Dprocessed that includes at least quantitative sentiment indicators, meteorological indicators, and public health-derived indicators. Step S3: The feature recommendation agent will use a dual mechanism of statistical algorithms and expert evaluation of large models to select the core feature set. ; Step S4, the core feature set Input the prediction execution agent, and the prediction execution agent predicts the core feature set. Based on the data scale and spatiotemporal correlation characteristics, we select a deep learning architecture from the built-in model library for training and evaluation to predict and output tourism demand.

[0020] This invention provides a tourism demand prediction method based on a large language model and multi-agent collaboration. Through steps S1-S4, the data collection agent, relying on the large language model, decomposes the prediction target into task sequences such as basic time-series data, public sentiment and willingness, and external risk environment. After reasoning and planning the path, collecting data through action, and observing and verifying data quality, it outputs a dataset. The data processing agent transforms this dataset into a standardized feature set containing quantitative sentiment, meteorological, and public health-related indicators. A feature recommendation agent, combining statistical algorithms and expert evaluation of the large model, selects a core feature set. The core feature set is then input into the prediction execution agent, which selects an appropriate deep learning architecture based on the data scale and spatiotemporal relevance to complete training and evaluation, ultimately outputting the tourism demand prediction result. By collaborating with data acquisition agents, data processing agents, feature recommendation agents, and prediction execution agents, the complex task of tourism demand forecasting is decomposed into four stages: perception, cognition, decision-making, and execution. This automates the complex tourism demand forecasting task. Each agent combines specific technologies and algorithms, and through a multi-agent collaborative mechanism, continuously optimizes feature recommendation and prediction accuracy. The entire lifecycle, from raw data capture to final prediction output, is automated. While improving prediction accuracy, the introduction of a knowledge base enhances the interpretability of the prediction process, ensuring that the model can dynamically adapt to market fluctuations in the post-pandemic era and provide interpretable decision-making basis.

[0021] The Data Collection Agent (DCA), acting as the system's perception unit, is responsible for analyzing and predicting task requirements and automatically acquiring multi-source heterogeneous data. To effectively address the problem of data acquisition failures caused by data source changes in complex network environments, in a more preferred embodiment of this invention, in step S1, if the built-in quality assessment module fails the evaluation, the feedback result is used as an environmental variable to re-enter the inference stage, triggering a self-correction of the data acquisition strategy. During the inference stage, the Large Language Model (LLM) plans the operation path based on the task context. During the action stage, structured and unstructured data are acquired by calling the API selector or automated data acquisition service through the MCP. During the observation stage, the built-in quality assessment module performs real-time verification of the data's completeness, format consistency, and time coverage. If the evaluation fails, the feedback result is re-inputted as an environmental variable into the inference stage, triggering a self-correction of the data acquisition strategy, ensuring that the output is high-quality data that can be directly used by subsequent algorithm models without further manual intervention. This process outputs the original heterogeneous dataset Draw, which contains structured numerical values ​​and unstructured text. Before output, the built-in quality assessment module verifies the data integrity. If the verification fails, the observation and correction steps in ReAct (Reasoning-Action-Observation) are triggered. The Draw that passes the verification is then passed as input to the data processing agent (DPA).

[0022] By collaborating with a Data Collection Agent (DCA) and a Data Processing Agent (DPA), this invention achieves automatic capture and standardized processing of official APIs, unstructured web page comments, and meteorological and health data. The Model Context Protocol (MCP) standardizes the interaction between the large language model and external data sources and tools, enabling plug-and-play integration. The framework can flexibly call external tools, effectively solving the bottleneck problems of low data collection efficiency and frequent manual intervention in traditional forecasting. The MCP can integrate official APIs, unstructured web page comments, and public health data in real time and comprehensively. Compared to the limitations of traditional models in handling heterogeneous data and lagging data updates, this invention ensures the timeliness and comprehensiveness of input data, enabling forecast results to quickly respond to market fluctuations.

[0023] The Data Collection Agent (DCA) employs the ReAct (Reasoning-Action-Observation) mechanism to decompose the prediction objective into a sequence of sub-tasks that include basic time-series data, public sentiment and willingness, and the external risk environment. As a client of the Model Context Protocol (MCP), the DCA interacts with external tool servers through a standardized JSON request format.

[0024] To integrate deterministic procedural tools with a flexible workflow for natural language operations, in a more preferred embodiment of this invention, the data processing agent performs preprocessing of basic data by invoking standardized tools to perform missing value imputation, deduplication, and frequency resampling. For large-scale unstructured text, a semantic analysis interface is invoked to convert sentiment into quantitative indicators based on custom evaluation prompts. The automated feature construction and semantic quantification data processing agent is responsible for converting the Draw into a standardized feature set.

[0025] For complex derived features within the domain, the breadth and processing efficiency of feature engineering are improved. In a more preferred embodiment of this invention, feature construction formulas are retrieved from a knowledge base, and logical transformation is achieved using a built-in code generation engine. For example, a large language model retrieves feature construction formulas from the knowledge base and dynamically generates executable script code to perform nonlinear fusion calculations on basic meteorological or public health indicators under controlled conditions. By converting natural language instructions into logically rigorous calculation scripts, the Data Processing Agent (DPA) can automatically derive derived features reflecting the deep correlation between the perceived environment and health risks. The results processed by the Data Processing Agent (DPA) generate a feature set containing quantitative sentiment indicators, meteorological indicators, and public health derived indicators. The set is then passed to the Feature Recommendation Agent (FRA) for scientific screening.

[0026] To ensure the scientific validity of the features input to the Forecasting Agent (FA), providing decision-makers with in-depth interpretive evidence and enhancing the credibility of prediction results, features with strong explanatory power for the target variable and complementary to each other are selected. Redundant features are initially eliminated by maximizing the correlation between features and the target and minimizing redundancy between features. In a more preferred embodiment of the invention, in step S3, the maximum correlation minimum redundancy algorithm mRMR(Xi) is used as follows: in, For the first One characteristic variable whose impact needs to be evaluated; The core target variable; The size of the candidate feature set; For characteristic variables With target variable The correlation coefficient between them; For characteristic variables and The correlation coefficient between them.

[0027] To identify nonlinear key features that strongly contribute to the target variable by measuring the shared information among variables, in a more preferred embodiment of the present invention, mutual information (MI) is used to quantify the dependency between features and the prediction target. By measuring the shared information I(X;Y) among variables, nonlinear key features that strongly contribute to the target variable are identified. in, For variables and The joint probability density; , For variables and The marginal probability density.

[0028] By employing the maximum relevance and minimum redundancy algorithm and shared information in the Data Processing Agent (DPA), variables with extremely low relevance or excessive redundancy are eliminated. In a more preferred embodiment of the invention, a candidate feature subset Dcandidate is extracted from the feature set Dprocessed. For each feature in the candidate feature subset Dcandidate, a thought chain is used to reason and obtain a comprehensive score S(feature). Here, SG represents the scoring dimension, R represents relevance, used to measure the fit between evidence and feature topic; S represents support, used to assess the strength of theoretical support; U represents utility, used to measure the information gain of the feature in improving model accuracy; the weight wG can be configured according to specific task scenarios, used to assign importance weights with theoretical support to each feature; the calculation results are mapped to the interval [0, 100] through a linear normalization function σ(x), and are prioritized from largest to smallest according to the calculation results, selecting a preset proportion of high-scoring items to form the core feature set. .

[0029] in, This is the weighted sum of the scores for the currently selected features. and These represent the theoretical minimum and maximum values ​​under this scoring system, respectively. Based on the scoring results, candidate features are prioritized and a recommendation scheme with detailed textual explanations is generated, transforming the feature selection process from traditional black-box statistical screening to a scientifically logical auxiliary decision-making process. The Feature Recommendation Agent (FRA) prioritizes features based on the normalized score of σ(x), selecting high-scoring items to form the core feature set. This set not only possesses statistical significance but is also linked to traceable theoretical evidence, ultimately being passed to the predictive executive agent (FA).

[0030] To better facilitate automated modeling and quality monitoring feedback, in a more preferred embodiment of the present invention, step S4 involves using the Data Readiness Assessment Index (DCEI) to perform a self-check on data quality. in, For coverage, This is used to assess the coverage of core business dimensions; For novelty, This is used to quantify the ability to discover new risk characteristics; For redundancy, Used to identify the independence of feature sets. These are preset weights, used as criteria for determining data quality admission.

[0031] The closed-loop feedback mechanism is as follows: The system sets the access threshold. : Logic for achieving the target: If If the data is ready, it is considered that the data is ready. The predictive execution agent (FA) automatically matches the optimal algorithm from the model library, such as LSTM and Informer models, to make predictions based on the principle of minimizing prediction error.

[0032] Non-compliance correction logic: If This indicates that the current feature set has deficiencies in coverage or novelty. At this point, the system triggers a feedback signal, instructing the data collection agent (DCA) to expand the data collection range or requiring the data processing agent (DPA) to adjust the feature calculation logic, thereby forming an intelligent closed loop of self-correction internally.

[0033] The prediction execution module is responsible for transforming the selected core feature set into the final tourism demand prediction output. The prediction execution agent (FA) automates the entire process from model selection and training to evaluation. The module has a built-in model library covering various deep learning architectures, which can automatically trigger hyperparameter tuning and algorithm adaptation based on the data scale and spatiotemporal correlation characteristics of the feature set. The prediction execution agent (FA) can automatically match the optimal algorithm for different prediction tasks and dynamically adjust the front-end agent's data collection and recommendation strategies based on prediction error feedback. This closed-loop optimization mechanism ensures the model's universality across different regions and time spans, achieving a balance between prediction accuracy and system stability.

[0034] The predictive execution agent (FA) automatically generates time-series visualization charts comparing actual and predicted values ​​while outputting predicted values. Based on the principle of minimizing prediction error, the system automatically matches algorithms from the model library using standardized metrics such as Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE). After prediction execution, the resulting performance deviation is fed back to the Feature Recommendation agent (FRA) as a feedback signal to dynamically adjust the weight factor wG in the feature (S) component, achieving intelligent closed-loop optimization throughout the entire process. This evaluation result not only serves as the final output but also as a feedback signal to adjust the feature weights of the Feature Recommendation agent (FRA) or the data collection range of the Data Collection agent (DCA), thus forming a continuously optimizing intelligent closed loop within the entire framework and ensuring that the system can provide highly reliable decision-making support in complex and dynamic environments.

[0035] This invention also discloses a system for predicting tourism demand based on a large language model and multi-agent collaboration, as described above. The system includes... The data acquisition agent is used to decompose the tourism demand forecasting target into a task sequence that includes at least basic time series data, public sentiment and willingness, and external risk environment; it can use the ReAct framework to form a cyclical process of reasoning, action and observation phases. The inference phase is used to plan the operation path based on the task context; the action phase is used to call the API selector or automated crawling service through the model context protocol to obtain structured and unstructured data; the observation phase is used by the built-in quality assessment module to perform real-time verification of the data's integrity, format consistency and time coverage. If the assessment is qualified, the dataset Draw is output. A data processing agent is used to transform the dataset Draw into a standardized feature set, and output a feature set Dprocessed that includes at least quantitative sentiment indicators, meteorological indicators, and public health-derived indicators. Feature recommendation agents are used to select core feature sets through a dual mechanism of statistical algorithms and expert evaluation using large models. ; Predictive execution agents, used to predict based on core feature sets Based on the data scale and spatiotemporal correlation characteristics, we select a deep learning architecture from the built-in model library for training and evaluation to predict and output tourism demand.

[0036] This invention discloses a system for tourism demand forecasting based on a multi-agent collaborative approach using a large language model. Addressing the pain points of traditional tourism demand data processing, such as multi-source heterogeneity, severe fragmentation, high reliance on manual intervention, and inconsistent data quality, the system utilizes a data acquisition agent based on the ReAct framework to achieve autonomous iterative cycles of reasoning, action, and observation. It automatically decomposes multi-dimensional data acquisition tasks using a large language model and efficiently calls external tools through a model context protocol to automate the capture of structured and unstructured data. A built-in quality assessment module enables real-time verification of data compliance, completely breaking down barriers to multi-source data acquisition, significantly reducing the cost of manual intervention, and ensuring the integrity, standardization, and timeliness of the dataset from the source. This solves the problems of low efficiency and difficulty in quality control in traditional data acquisition. Through the division of labor and closed-loop process of the data acquisition agent (DCA), data processing agent (DPA), feature recommendation agent (FRA), and prediction execution agent (FA), combined with a large language model (LLM), the ReAct inference framework, and the model context protocol (MCP), the system achieves intelligent upgrading and efficiency optimization of the entire tourism demand forecasting process.

[0037] The present invention also discloses a machine-readable storage medium storing instructions for causing a machine to execute the tourism demand prediction method based on a large language model and multi-agent collaboration as described above.

[0038] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A tourism demand prediction method based on large language model multi-agent collaboration, characterized in that, The tourism demand prediction method based on large language model multi-agent collaboration includes: Step S1: In the data acquisition agent, the large language model decomposes the tourism demand prediction target into a task sequence that includes at least basic time series data, public sentiment and willingness, and external risk environment. During the reasoning phase, the operation path is planned based on the task context; During the action phase, structured and unstructured data are obtained by calling the API selector or automated crawling service through the model context protocol; During the observation phase, the built-in quality assessment module verifies the data's integrity, format consistency, and time coverage in real time. If the assessment is satisfactory, the dataset Draw is output. Step S2: The data processing agent transforms the dataset Draw into a standardized feature set and outputs a feature set Dprocessed that includes at least quantitative sentiment indicators, meteorological indicators, and public health-derived indicators. Step S3, the feature recommendation agent screens out a core feature set through a statistical algorithm and a double mechanism of expert evaluation of a large model ; Step S4, selecting a deep learning architecture from the built-in model library according to the data scale and spatiotemporal correlation characteristics of the core feature set The input prediction execution agent selects a deep learning architecture from the built-in model library according to the data scale and spatiotemporal correlation characteristics of the core feature set The input prediction execution agent selects a deep learning architecture from the built-in model library according to the data scale and spatiotemporal correlation characteristics of the core feature set 2. The method of claim 1, wherein the method is characterized by, In step S1, if the built-in quality assessment module fails the assessment, the feedback result is used as an environmental variable to re-enter the reasoning stage, triggering the self-correction of the crawling strategy.

3. The method of claim 1, wherein the method is characterized by, In the data processing agent, basic data preprocessing is completed by calling standardized tools to perform missing value imputation, deduplication, and frequency resampling; for large-scale unstructured text, the semantic analysis interface is called to convert sentiment into quantitative indicators based on custom evaluation prompts.

4. The method of claim 3, wherein the method is characterized by, Formulas are constructed by retrieving features from the knowledge base, and logical transformations are achieved using the built-in code generation engine.

5. The method of claim 1, wherein the method is characterized by, In step S3, the maximum correlation minimum redundancy algorithm mRMR(Xi) is used to select features that have strong explanatory power for the target variable and are complementary to each other. wherein, is the th impact feature variable to be evaluated; is the core target variable; is the size of the candidate feature set; is the correlation coefficient between the feature variable and the target variable ; and is the correlation coefficient between the feature variable and .

6. The method of claim 5, wherein the method is characterized by, By leveraging the dependence between mutual information quantification features and prediction targets, and by measuring the shared information I(X;Y) among variables, nonlinear key features that strongly contribute to the target variable are identified. wherein is a variable the joint probability density of ; is a variable the marginal probability density of .

7. The method of claim 6, wherein the method further comprises: From the feature set Dprocessed, a subset of candidate features Dcandidate is extracted. For each feature in the subset Dcandidate, a reasoning chain is used to obtain a comprehensive score. Here, SG represents the scoring dimension, R represents relevance, used to measure the fit between evidence and feature topic; S represents support, used to assess the strength of theoretical support; U represents utility, used to measure the information gain of the feature in improving model accuracy; the weight wG can be configured according to specific task scenarios, used to assign importance weights with theoretical support to each feature; the calculation results are mapped to the interval [0, 100] through a linear normalization function σ(x), and are prioritized from largest to smallest according to the calculation results, selecting a preset proportion of high-scoring items to form the core feature set. .

8. The tourism demand forecasting method based on multi-agent collaboration using a large language model according to any one of claims 1-7, characterized in that, In step S4, the Data Readiness Index (DCEI) is used to perform a self-check on data quality. wherein, is coverage, for assessing the coverage of core business dimensions; is novelty, for quantifying the discovery ability of new risk features; is redundancy, for identifying the independence of the feature set, is a preset weight, used as a decision criterion for data quality access.

9. The system based on the large language model multi-agent collaboration tourism demand prediction method according to any one of claims 1-8, characterized in that, The system includes, The data acquisition agent is used to decompose the tourism demand forecasting target into a task sequence that includes at least basic time series data, public sentiment and willingness, and external risk environment; it can use the ReAct framework to form a cyclical process of reasoning, action and observation phases. The inference phase is used to plan the operation path based on the task context; the action phase is used to call the API selector or automated crawling service through the model context protocol to obtain structured and unstructured data; the observation phase is used by the built-in quality assessment module to perform real-time verification of the data's integrity, format consistency and time coverage. If the assessment is qualified, the dataset Draw is output. A data processing agent is used to transform the dataset Draw into a standardized feature set, and output a feature set Dprocessed that includes at least quantitative sentiment indicators, meteorological indicators, and public health-derived indicators. A feature recommendation agent is configured to filter out a core feature set through a dual mechanism of statistical algorithm and large model expert evaluation ; A prediction-executing agent for predicting tourist demand based on a core feature set of data size and spatio-temporal correlation characteristics, selecting a deep learning architecture from an in-built model library for training and evaluation.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute the tourism demand prediction method based on a large language model and multi-agent collaboration as described in any one of claims 1-8.

Images