An AI prediction-based tourist carrying capacity dynamic regulation method and system

Abstract

The application discloses a kind of tourist carrying capacity dynamic regulation method and system based on AI prediction, to solve the problems such as traditional scenic spot carrying capacity management static, lag, extensive etc..The method fuses tourist behavior, environmental facilities, external environment and other multi-source real-time data by constructing digital twin of scenic spot, establishes unified space-time data lake;Using space-time graph neural network and Transformer model, predict future passenger flow distribution heat map, identify congestion risk in advance;Break through the limitation of single maximum carrying capacity, comprehensive safety, experience, ecology and four types of threshold value of facility, dynamically calculate regional carrying capacity, and introduce risk resilience index evaluation system to assess the anti-interference ability, automatically generate hierarchical control scheme, finally through A / B test and effect feedback, continuously optimize model, form "perception-prediction-decision-making-execution-evaluation" closed loop.The system significantly improves the safety level of scenic spot, tourist experience and ecological sustainability, realizes intelligent, flexible, self-evolution passenger flow management.

Description

Technical Field

[0001] This invention relates to the field of dynamic regulation of scenic areas, and in particular to a method and system for dynamic regulation of tourist carrying capacity based on AI prediction. Background Technology

[0002] Most scenic areas rely on reactive measures: only implementing flow control and evacuation measures after congestion, complaints, or even safety incidents occur; they lack the ability to predict tourist flow trends in advance, making proactive intervention impossible before risks arise; passive responses not only affect the tourist experience but also increase emergency management costs and safety risks. Data sources within scenic areas are fragmented: ticketing, turnstiles, WiFi probes, cameras, environmental sensors, and external weather / traffic systems operate independently; data formats are inconsistent, and spatiotemporal benchmarks are chaotic, making fusion and analysis impossible; managers "can see the part but not the whole," further hindering intelligent decision-making. Common control practices involve a "one-size-fits-all" approach of flow control, closing entrances, or broadcasting reminders, ignoring individual tourist differences (e.g., family tourists vs. backpackers); they cannot achieve refined control by time, area, or group; supply-side measures (e.g., ticket distribution) are disconnected from on-site crowd management, limiting the effectiveness of control. An excessive focus on "safety" and "revenue" leads to a lack of quantitative constraints on tourist experience quality (e.g., queuing time, crowding) and ecological carrying capacity (e.g., vegetation trampling, water pollution); long-term overload operation results in resource degradation, declining reputation, and damage to the sustainable development of scenic areas. Summary of the Invention

[0003] A method for dynamic regulation of tourist carrying capacity based on AI prediction includes the following steps; S1. Multi-source data acquisition: Construct a digital twin of the scenic area to achieve full-dimensional, real-time data acquisition, including tourist-related data, environmental and facility data, external environment data, and historical data. Map the 3D model of the scenic area with real-time pedestrian flow, facility status, and environmental parameters to form a dynamic simulation environment that is synchronized between the virtual and the real. S2. Data Fusion and Processing: Align, clean, and standardize the heterogeneous data in time and space to build a unified spatiotemporal data lake. Based on tourist origin, age, and consumption preference tags, segment the customer groups to provide high-quality training and input data for AI models. S3. Tourist Flow and Distribution Prediction: The Spatiotemporal Graph Neural Network (STGNN) is used to model the topological relationships between attractions and the patterns of tourist flow. A Transformer-based prediction module is introduced to capture long-term dependencies: the periodicity of holidays. Based on real-time flow data, a heat map of the spatial distribution of tourists within the scenic area is predicted to identify potential congestion points in advance (such as the lower cable car station and the core viewing platform). S4. Dynamic Bearing Capacity Threshold Calculation: This method departs from the traditional single concept of maximum bearing capacity. Based on prediction results and real-time status, it calculates multi-level dynamic bearing capacity thresholds. The AI ​​model comprehensively considers: safety red line, experience threshold, ecological threshold, and facility load threshold. It also introduces the concept of resilience bearing capacity to assess the system's recovery capability under sudden disturbances (such as heavy rain or equipment failure) and constructs a risk resilience index. ; S5. Intelligent generation of control strategies: The system has a built-in "strategy knowledge base" which contains contingency plans for different levels of congestion, from early warning to different levels of congestion. Based on the predicted congestion level, location and cause, the AI ​​decision engine (rule engine + reinforcement learning) automatically generates and recommends the optimal combination of control strategies. S6. Implementation of Multi-Level Dynamic Control Measures: Supply-Side Control (Pre-Event / Remote), Dynamic Adjustment of Ticket / Reservation Inventory: Based on forecasts, release reservation quotas in different time periods, link with OTAs / navigation apps, issue visitor flow warnings, recommend off-peak routes and alternative destinations, predict peak visitor arrival times from source areas, and coordinate public transportation capacity in advance; In-Event Diversion Control (During Event / Nearby): Push the best tour routes and congestion alerts in real time through smart screens, mini-programs, and broadcasts; dynamically adjust one-way traffic routes, temporarily open backup channels, intelligently dispatch shuttle buses, and increase capacity in high-frequency congestion areas; dynamically adjust the number of guided tours for popular venues and implement dynamic time-sharing pricing to guide tourists to choose off-peak times using price signals; S7. Real-time evaluation of control effects: After the implementation of control measures, key indicators (such as the speed of congestion relief, tourist satisfaction questionnaires, social media sentiment, and the balance of facility usage) are continuously monitored. Through A / B testing, the actual effects of different control strategies are quantitatively analyzed. The control effect data and new tourist behavior data are fed back to the data lake of the first stage. The AI ​​prediction model and decision-making model are updated regularly or online to adapt to new tourist behavior patterns (such as the effect of popular tourist attractions) and changes in the external environment, forming a complete closed loop of "perception-prediction-decision-action-evaluation".

[0004] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. Step S1 achieves full-dimensional, real-time data collection, including tourist-related data, environmental and facility data, external environment data, and historical data. The specific data collected is as follows; Tourist-related data: real-time number of visitors, origin of visitors, visitor movement routes, online booking / ticketing data, social media sentiment and search popularity; Environmental and facility data: density of key areas, parking lot saturation, toilet usage frequency, queue length for sightseeing buses / cable cars, and dwell time at key attractions; External environmental data: weather, holiday information, schedules of major events, and surrounding traffic congestion index; Historical data: tourist volume, accident records, and complaint hotspots in the same period of previous years.

[0005] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S2, the heterogeneous data mentioned above are aligned, cleaned, and standardized in time and space to construct a unified time and space data lake. The specific steps are as follows: S21. Unified Time Standard: All data is first processed according to a unified time format. For example, timestamps from different sources, such as gate card swipe records, mobile phone location points, and sensor readings, are all converted into a standard format and based on Beijing time. High-frequency and low-frequency data are aggregated and interpolated according to fixed time windows (e.g., every 5 minutes) so that they can be aligned at the same time granularity. S22. Unified Spatial Coordinates: All location information of various devices and data points within the scenic area is uniformly converted to the same geographic coordinate system (such as WGS84 latitude and longitude), and the entire scenic area is divided into several small areas (such as a 50m × 50m grid). Each tourist location, sensor, and facility point is mapped to the corresponding grid to achieve spatial alignment. S23. Cleaning dirty data: Data from different sources often contains errors and noise. A sudden change in cell phone signal caused tourists to "instantly teleport" several hundred meters. A temperature and humidity sensor is stuck, outputting the same value for several consecutive hours; If the ticketing system misses the age of tourists, the system will automatically identify these anomalies, remove obviously erroneous records, and fill in the missing values ​​appropriately (such as by using the average value of surrounding sensors) to ensure that the data entering the subsequent stages is clean and reliable. S24. Standardization and Feature Extraction: Transform the cleaned data format into a consistent structure. Convert sunny / rainy / snowy weather descriptions into numerical codes; Tagging tourists by their place of origin, age group, and spending level creates customer profiles; Extract useful features such as dwell time, movement speed, and popular areas from the original trajectory; S25. Writing to the Spatiotemporal Data Lake: All processed data is organized according to a "time + spatial grid" and stored in a specially designed storage system—the Spatiotemporal Data Lake—which supports fast querying of data in any time period and any region.

[0006] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S3, the spatiotemporal graph neural network STGNN is used to model the topological relationships between attractions and the flow patterns of tourists. A Transformer-based prediction module is introduced to capture long-term dependencies. The specific steps are as follows: S31. Spatiotemporal graph structure construction: The scenic spots, main road intersections, transportation hubs and important facilities in the scenic area are abstracted as graph nodes. Based on the connection paths between scenic spots and the actual transfer probability of tourists, a weighted adjacency matrix is ​​constructed to quantify the spatial association strength. A multi-dimensional feature vector is constructed for each node, including historical passenger flow sequence, real-time status data, node attributes and external environmental factors. S32.STGNN Spatial Topology and Short-Term Dependency Modeling: By aggregating the features of each node and its neighboring nodes through a graph convolutional network, the spatial association between attractions and the effect of visitor flow diffusion are encoded. A temporal convolutional network is applied to the feature sequence of each node to extract local temporal patterns and short-term change trends, and outputs a node state sequence that integrates spatiotemporal information. The features of each time step contain its spatial context and recent evolution information. S33. Transformer Long-Term Temporal Dependency Modeling: The spatiotemporal feature sequence output by STGNN is used as input. Global dependency capture is performed. Through a multi-head self-attention mechanism, direct correlation between features at any time step in the sequence is established. Long-term periodic patterns spanning days, weeks, and holidays are identified. After processing by a feedforward network and layer normalization, a high-order spatiotemporal feature representation with global temporal context is output. S34. Multi-step prediction and heat map generation: Based on the comprehensive features output by Transformer, the system decodes the data through a fully connected network to generate predicted visitor flow values ​​for each node at multiple future time steps. The predicted node values ​​are then mapped to a geospatial grid. An interpolation algorithm is used to generate a refined heat map of visitor flow distribution across the entire scenic area. Based on the gradient changes in the heat map, potential congestion areas and their evolution directions are identified, enabling early warning.

[0007] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S4, the AI ​​model comprehensively considers: safety red line, experience threshold, ecological threshold, and facility load threshold, as detailed below; Safety red line: Maximum physical space capacity (absolute upper limit); Experience threshold: An alert is triggered when the predicted density in a certain area exceeds the "comfortable experience" threshold. Ecological threshold: For sensitive ecological areas, the number of people allowed to enter is dynamically adjusted based on environmental sensor data (such as noise, temperature and humidity); Facility load threshold: The carrying capacity of facilities such as transportation, toilets, and trash cans.

[0008] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S5, based on the predicted congestion level, location, and cause, the AI ​​decision engine (rule engine + reinforcement learning) automatically generates and recommends the optimal combination of control strategies. The specific steps are as follows: S51. Rule Matching: The rule engine quickly matches the preset emergency plan (such as "moderate congestion in area A → activate one-way traffic plan") based on the predicted congestion level, location and cause, to ensure that basic and safe control measures can be activated immediately; S52. Reinforcement Learning Optimization: Within the framework of the contingency plan, the reinforcement learning model conducts rapid simulation and deduction through a digital twin environment, fine-tunes the strategy parameters: increases or decreases the number of shuttle buses, adjusts the timing of information push, and evaluates the long-term effects of different strategy combinations (such as experience, safety, and efficiency), and selects the optimization scheme with the highest overall benefit. S53. Execution and Evolution: After the optimized strategy is manually confirmed, it is issued for execution. The actual effect data after execution (such as whether congestion is relieved faster) will be fed back to the reinforcement learning model, enabling it to continuously learn from experience and make smarter decisions in the future.

[0009] A dynamic control system for tourist carrying capacity based on AI prediction, wherein the system is used to implement any AI prediction-based dynamic control method for tourist carrying capacity; the system comprises: Digital twin and multi-source sensing module: integrates a 3D model of the scenic area and accesses real-time data streams such as visitor trajectories, environmental sensors, facility status, and external weather / traffic / ticketing information; Spatiotemporal data fusion and governance module: performs spatiotemporal alignment, cleaning, and standardization on heterogeneous data from different systems, formats, and frequencies, and adds customer group tags (such as "family with children", "senior citizen group", "high-spending independent travel"). Passenger flow prediction and risk warning module: Using spatiotemporal graph neural network + Transformer model, predict the heat map of passenger flow distribution in various areas of the scenic spot in the next 1-4 hours, and identify potential congestion points and overloaded areas in real time. Dynamic bearing capacity assessment and resilience analysis module: Integrating four threshold categories—safety, experience, ecology, and facilities—it dynamically calculates the upper limit of bearing capacity for each area and introduces a "risk resilience index" to assess the system's resistance to pressure and its ability to recover under sudden events (such as fires and rainstorms). Intelligent control decision-making and strategy generation module: Built-in strategy knowledge base (including historical effective contingency plans); Combines rule engine (to ensure bottom-line safety) and reinforcement learning (to optimize long-term effects) to automatically generate tiered control plans (such as "mild warning: push diversion suggestions" and "severe congestion: temporarily close entrances + dispatch shuttle buses"). The control implementation and effect feedback module: links the ticketing system, OTA platform, navigation app, in-venue broadcasting, mini-program, shuttle bus dispatching system, etc., to implement control measures; the evaluation layer: quantifies the effect of the strategy through A / B testing and monitoring of key indicators (such as average queuing time, visitor satisfaction, evacuation efficiency); the feedback layer: feeds newly generated behavioral data and evaluation results back to the data lake to drive continuous iteration and optimization of the model.

[0010] The beneficial effects of this invention are as follows: By predicting congestion hotspots in advance and dynamically calculating safety thresholds, it transforms passive response into proactive prevention, effectively reducing the risk of safety accidents such as stampedes and chaotic evacuations; the introduction of "resilience bearing capacity" assessment enables rapid judgment of system stress capacity under emergencies (such as extreme weather or equipment failure), improving the scientific nature of emergency response. It avoids "crowded" tourism by reducing queuing time and the feeling of congestion through intelligent diversion, off-peak recommendations, and optimal route pushes; personalized services are provided based on customer profiles (such as avoiding steep routes for senior groups), improving satisfaction and repeat visits. It breaks away from the "one-size-fits-all" flow control model, supporting dynamic regulation by time period, region, and group; resources such as ticket reservations, traffic scheduling, and shuttle bus operation are precisely allocated on demand, improving facility utilization and operational efficiency. Attached Figure Description

[0011] Figure 1 A flowchart of a dynamic regulation method for tourist carrying capacity based on AI prediction; Detailed Implementation

[0012] A dynamic control method for tourist carrying capacity based on AI prediction, the flowchart of which is as follows: Figure 1 As shown, it includes the following steps; S1. Multi-source data acquisition: Construct a digital twin of the scenic area to achieve full-dimensional, real-time data acquisition, including tourist-related data, environmental and facility data, external environment data, and historical data. Map the 3D model of the scenic area with real-time pedestrian flow, facility status, and environmental parameters to form a dynamic simulation environment that is synchronized between the virtual and the real. S2. Data Fusion and Processing: Align, clean, and standardize the heterogeneous data in time and space to build a unified spatiotemporal data lake. Based on tourist origin, age, and consumption preference tags, segment the customer groups to provide high-quality training and input data for AI models. S3. Tourist Flow and Distribution Prediction: The Spatiotemporal Graph Neural Network (STGNN) is used to model the topological relationships between attractions and the flow patterns of tourists. A Transformer-based prediction module is introduced to capture long-term dependencies: the periodicity of holidays. Based on real-time flow data, a heat map of the spatial distribution of tourists within the scenic area is predicted to identify potential congestion points in advance (such as the lower cable car station and the core viewing platform). S4. Dynamic Bearing Capacity Threshold Calculation: This method departs from the traditional single concept of maximum bearing capacity. Based on prediction results and real-time status, it calculates multi-level dynamic bearing capacity thresholds. The AI ​​model comprehensively considers: safety red line, experience threshold, ecological threshold, and facility load threshold. It also introduces the concept of resilience bearing capacity to assess the system's recovery capability under sudden disturbances (such as heavy rain or equipment failure) and constructs a risk resilience index. ; S5. Intelligent generation of control strategies: The system has a built-in "strategy knowledge base" which contains contingency plans for different levels of congestion, from early warning to different levels of congestion. Based on the predicted congestion level, location and cause, the AI ​​decision engine (rule engine + reinforcement learning) automatically generates and recommends the optimal combination of control strategies. S6. Implementation of Multi-Level Dynamic Control Measures: Supply-Side Control (Pre-Event / Remote), Dynamic Adjustment of Ticket / Reservation Inventory: Based on forecasts, release reservation quotas in different time periods, link with OTAs / navigation apps, issue visitor flow warnings, recommend off-peak routes and alternative destinations, predict peak visitor arrival times from source areas, and coordinate public transportation capacity in advance; In-Event Diversion Control (During Event / Nearby): Push the best tour routes and congestion alerts in real time through smart screens, mini-programs, and broadcasts; dynamically adjust one-way traffic routes, temporarily open backup channels, intelligently dispatch shuttle buses, and increase capacity in high-frequency congestion areas; dynamically adjust the number of guided tours for popular venues and implement dynamic time-sharing pricing to guide tourists to choose off-peak times using price signals; S7. Real-time evaluation of control effects: After the implementation of control measures, key indicators (such as the speed of congestion relief, tourist satisfaction questionnaires, social media sentiment, and the balance of facility usage) are continuously monitored. Through A / B testing, the actual effects of different control strategies are quantitatively analyzed. The control effect data and new tourist behavior data are fed back to the data lake of the first stage. The AI ​​prediction model and decision-making model are updated regularly or online to adapt to new tourist behavior patterns (such as the effect of popular tourist attractions) and changes in the external environment, forming a complete closed loop of "perception-prediction-decision-action-evaluation".

[0013] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. Step S1 achieves full-dimensional, real-time data collection, including tourist-related data, environmental and facility data, external environment data, and historical data. The specific data collected is as follows; Tourist-related data: real-time number of visitors, origin of visitors, visitor movement routes, online booking / ticketing data, social media sentiment and search popularity; Environmental and facility data: density of key areas, parking lot saturation, toilet usage frequency, queue length for sightseeing buses / cable cars, and dwell time at key attractions; External environmental data: weather, holiday information, schedules of major events, and surrounding traffic congestion index; Historical data: tourist volume, accident records, and complaint hotspots in the same period of previous years.

[0014] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S2, the heterogeneous data mentioned above are aligned, cleaned, and standardized in time and space to construct a unified time and space data lake. The specific steps are as follows: S21. Unified Time Standard: All data is first processed according to a unified time format. For example, timestamps from different sources, such as gate card swipe records, mobile phone location points, and sensor readings, are all converted into a standard format and based on Beijing time. High-frequency and low-frequency data are aggregated and interpolated according to fixed time windows (e.g., every 5 minutes) so that they can be aligned at the same time granularity. S22. Unified Spatial Coordinates: All location information of various devices and data points within the scenic area is uniformly converted to the same geographic coordinate system (such as WGS84 latitude and longitude), and the entire scenic area is divided into several small areas (such as a 50m × 50m grid). Each tourist location, sensor, and facility point is mapped to the corresponding grid to achieve spatial alignment. S23. Cleaning dirty data: Data from different sources often contains errors and noise. A sudden change in cell phone signal caused tourists to "instantly teleport" several hundred meters. A temperature and humidity sensor is stuck, outputting the same value for several consecutive hours; If the ticketing system misses the age of tourists, the system will automatically identify these anomalies, remove obviously erroneous records, and fill in the missing values ​​appropriately (such as by using the average value of surrounding sensors) to ensure that the data entering the subsequent stages is clean and reliable. S24. Standardization and Feature Extraction: Transform the cleaned data format into a consistent structure. Convert sunny / rainy / snowy weather descriptions into numerical codes; Tagging tourists by their place of origin, age group, and spending level creates customer profiles; Extract useful features such as dwell time, movement speed, and popular areas from the original trajectory; S25. Writing to the Spatiotemporal Data Lake: All processed data is organized according to a "time + spatial grid" and stored in a specially designed storage system—the Spatiotemporal Data Lake—which supports fast querying of data in any time period and any region.

[0015] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S3, the spatiotemporal graph neural network STGNN is used to model the topological relationships between attractions and the flow patterns of tourists. A Transformer-based prediction module is introduced to capture long-term dependencies. The specific steps are as follows: S31. Spatiotemporal graph structure construction: The scenic spots, main road intersections, transportation hubs and important facilities in the scenic area are abstracted as graph nodes. Based on the connection paths between scenic spots and the actual transfer probability of tourists, a weighted adjacency matrix is ​​constructed to quantify the spatial association strength. A multi-dimensional feature vector is constructed for each node, including historical passenger flow sequence, real-time status data, node attributes and external environmental factors. S32.STGNN Spatial Topology and Short-Term Dependency Modeling: By aggregating the features of each node and its neighboring nodes through a graph convolutional network, the spatial association between attractions and the effect of visitor flow diffusion are encoded. A temporal convolutional network is applied to the feature sequence of each node to extract local temporal patterns and short-term change trends, and outputs a node state sequence that integrates spatiotemporal information. The features of each time step contain its spatial context and recent evolution information. S33. Transformer Long-Term Temporal Dependency Modeling: The spatiotemporal feature sequence output by STGNN is used as input. Global dependency capture is performed. Through a multi-head self-attention mechanism, direct correlation between features at any time step in the sequence is established. Long-term periodic patterns spanning days, weeks, and holidays are identified. After processing by a feedforward network and layer normalization, a high-order spatiotemporal feature representation with global temporal context is output. S34. Multi-step prediction and heat map generation: Based on the comprehensive features output by Transformer, the system decodes the data through a fully connected network to generate predicted visitor flow values ​​for each node at multiple future time steps. The predicted node values ​​are then mapped to a geospatial grid. An interpolation algorithm is used to generate a refined heat map of visitor flow distribution across the entire scenic area. Based on the gradient changes in the heat map, potential congestion areas and their evolution directions are identified, enabling early warning.

[0016] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S4, the AI ​​model comprehensively considers: safety red line, experience threshold, ecological threshold, and facility load threshold, as detailed below; Safety red line: Maximum physical space capacity (absolute upper limit); Experience threshold: An alert is triggered when the predicted density in a certain area exceeds the "comfortable experience" threshold. Ecological threshold: For sensitive ecological areas, the number of people allowed to enter is dynamically adjusted based on environmental sensor data (such as noise, temperature and humidity); Facility load threshold: The carrying capacity of facilities such as transportation, toilets, and trash cans.

[0017] Furthermore, a method for dynamic regulation of tourist carrying capacity based on AI prediction is proposed. In step S5, based on the predicted congestion level, location, and cause, the AI ​​decision engine (rule engine + reinforcement learning) automatically generates and recommends the optimal combination of control strategies. The specific steps are as follows: S51. Rule Matching: The rule engine quickly matches the preset emergency plan (such as "moderate congestion in area A → activate one-way traffic plan") based on the predicted congestion level, location and cause, to ensure that basic and safe control measures can be activated immediately; S52. Reinforcement Learning Optimization: Within the framework of the contingency plan, the reinforcement learning model conducts rapid simulation and deduction through a digital twin environment, fine-tunes the strategy parameters: increases or decreases the number of shuttle buses, adjusts the timing of information push, and evaluates the long-term effects of different strategy combinations (such as experience, safety, and efficiency), and selects the optimization scheme with the highest overall benefit. S53. Execution and Evolution: After the optimized strategy is manually confirmed, it is issued for execution. The actual effect data after execution (such as whether congestion is relieved faster) will be fed back to the reinforcement learning model, enabling it to continuously learn from experience and make smarter decisions in the future.

[0018] A dynamic control system for tourist carrying capacity based on AI prediction, wherein the system is used to implement any AI prediction-based dynamic control method for tourist carrying capacity; the system comprises: Digital twin and multi-source sensing module: integrates a 3D model of the scenic area and accesses real-time data streams such as visitor trajectories, environmental sensors, facility status, and external weather / traffic / ticketing information; Spatiotemporal data fusion and governance module: performs spatiotemporal alignment, cleaning, and standardization on heterogeneous data from different systems, formats, and frequencies, and adds customer group tags (such as "family with children", "senior citizen group", "high-spending independent travel"). Passenger flow prediction and risk warning module: Using spatiotemporal graph neural network + Transformer model, predict the heat map of passenger flow distribution in various areas of the scenic spot in the next 1-4 hours, and identify potential congestion points and overloaded areas in real time. Dynamic bearing capacity assessment and resilience analysis module: Integrating four threshold categories—safety, experience, ecology, and facilities—it dynamically calculates the upper limit of bearing capacity for each area and introduces a "risk resilience index" to assess the system's resistance to pressure and its ability to recover under sudden events (such as fires and rainstorms). Intelligent control decision-making and strategy generation module: Built-in strategy knowledge base (including historical effective contingency plans); Combines rule engine (to ensure bottom-line safety) and reinforcement learning (to optimize long-term effects) to automatically generate tiered control plans (such as "mild warning: push diversion suggestions" and "severe congestion: temporarily close entrances + dispatch shuttle buses"). The control implementation and effect feedback module: links the ticketing system, OTA platform, navigation app, in-venue broadcasting, mini-program, shuttle bus dispatching system, etc., to implement control measures; the evaluation layer: quantifies the effect of the strategy through A / B testing and monitoring of key indicators (such as average queuing time, visitor satisfaction, evacuation efficiency); the feedback layer: feeds newly generated behavioral data and evaluation results back to the data lake to drive continuous iteration and optimization of the model.

[0019] Example 2: Visitor flow control during the National Day Golden Week at Yunfeng Mountain, a 5A-level mountain scenic area. On the eve of the National Day holiday in 2025, Yunfeng Mountain Scenic Area deployed a dynamic control system based on this method; In the S1 phase, the system accesses gate card swipe records, mobile phone signaling, WiFi probes, IoT environmental sensors, meteorological APIs, and historical passenger flow data from the past three years, and integrates them with a high-precision 3D real-scene model to build a digital twin platform; In the S2 phase, all data were aligned according to 5-minute time slices and 50-meter H3 grids, abnormal trajectories were cleaned up, and three main customer groups were identified: "families with children", "senior travel groups", and "young independent travelers". During the S3 phase, the STGNN-Transformer model predicted that the crowd density in the main peak viewing area would exceed 8 people / ㎡ from 10:00 to 12:00 on October 2nd, and issued a red congestion warning 4 hours in advance. In the S4 phase, the system calculates the dynamic carrying capacity of the area: the safety threshold (6 people / ㎡) and the experience threshold (4 people / ㎡), and the current upper limit is 4.5 people / ㎡. At the same time, the risk resilience index is low (due to narrow exits and insufficient emergency supplies), triggering an enhanced response. In the S5 phase, the AI ​​decision engine recommends the following combined solutions from the strategy library: ① Suspend online reservations for park entry slots from 10:00 to 12:00; ② Push a "suggested to avoid peak hours in the afternoon" prompt to Gaode Map; ③ Activate one-way loop traffic; ④ Deploy 2 additional shuttle buses for cyclical transportation. In Phase S6, the above measures will be implemented simultaneously through OTA platforms, scenic area mini-programs, and entrance screens. In the S7 phase, the system compared data before and after implementation: the average density of the observation deck decreased to 3.8 people / ㎡, the average visitor satisfaction increased by 22%, and the effects of the relevant strategies were recorded and used to optimize subsequent models.

Claims

1. A method for dynamic regulation of tourist carrying capacity based on AI prediction, characterized in that, Includes the following steps; S1. Multi-source data acquisition: Construct a digital twin of the scenic area to achieve full-dimensional, real-time data acquisition, including tourist-related data, environmental and facility data, external environment data, and historical data. Map the 3D model of the scenic area with real-time pedestrian flow, facility status, and environmental parameters to form a dynamic simulation environment that is synchronized between the virtual and the real. S2. Data Fusion and Processing: Align, clean, and standardize the heterogeneous data in time and space to build a unified spatiotemporal data lake. Based on tourist origin, age, and consumption preference tags, segment the customer groups to provide high-quality training and input data for AI models. S3. Tourist Flow and Distribution Prediction: The Spatiotemporal Graph Neural Network (STGNN) is used to model the topological relationships between attractions and the flow patterns of tourists. A Transformer-based prediction module is introduced to capture long-term dependencies: the periodicity of holidays. Based on real-time flow data, a heat map of the spatial distribution of tourists within the scenic area is predicted to identify potential congestion points in advance. S4. Dynamic Bearing Capacity Threshold Calculation: This method departs from the traditional single concept of maximum bearing capacity. Based on prediction results and real-time status, it calculates multi-level dynamic bearing capacity thresholds. The AI ​​model comprehensively considers: safety red line, experience threshold, ecological threshold, and facility load threshold. It also introduces the concept of resilience bearing capacity to assess the system's recovery capability under sudden disturbances and constructs a risk resilience index. ; S5. Intelligent generation of control strategies: The system has a built-in strategy knowledge base, which includes contingency plans for different levels of congestion, from early warning to different levels of congestion. Based on the predicted congestion level, location and cause, the AI ​​decision engine (rule engine + reinforcement learning) automatically generates and recommends the optimal combination of control strategies. S6. Implementation of multi-level dynamic control measures: Supply-side control and dynamic adjustment of ticket / reservation inventory: Based on forecasts, release reservation quotas in different time periods, link with OTA / navigation apps, issue visitor flow warnings, recommend off-peak routes and alternative destinations, predict peak visitor arrival times from source areas, coordinate public transportation capacity in advance, and implement on-site diversion control: Push the best tour routes and congestion alerts in real time through smart screens, mini-programs, and broadcasts, dynamically adjust one-way traffic routes, temporarily open backup channels, intelligently dispatch shuttle buses, and increase capacity in high-frequency congestion areas; S7. Real-time evaluation of control effects: After the implementation of control measures, key indicators are continuously monitored. The actual effects of different control strategies are quantitatively analyzed through A / B testing. The control effect data and new tourist behavior data are fed back to the data lake in the first stage. The AI ​​prediction model and decision-making model are updated regularly to adapt to new tourist behavior patterns and changes in the external environment, forming a complete closed loop of perception-prediction-decision-action-evaluation.

2. The method for dynamic adjustment of tourist carrying capacity based on AI prediction as described in claim 1, characterized in that, Step S1 achieves full-dimensional, real-time data collection, including tourist-related data, environmental and facility data, external environment data, and historical data. The specific data collected is as follows; Tourist-related data: real-time number of visitors, origin of visitors, visitor movement routes, online booking / ticketing data, social media sentiment and search popularity; Environmental and facility data: density of key areas, parking lot saturation, toilet usage frequency, queue length for sightseeing buses / cable cars, and dwell time at key attractions; External environmental data: weather, holiday information, schedules of major events, and surrounding traffic congestion index; Historical data: tourist volume, accident records, and complaint hotspots in the same period of previous years.

3. The method for dynamic adjustment of tourist carrying capacity based on AI prediction as described in claim 1, characterized in that, In step S2, the heterogeneous data mentioned above are aligned, cleaned, and standardized in time and space to construct a unified time and space data lake. The specific steps are as follows: S21. Unified Time Standard: All data is first processed according to a unified time format. For example, timestamps from different sources, such as gate card swipe records, mobile phone location points, and sensor readings, are all converted into a standard format and based on Beijing time. High-frequency and low-frequency data are aggregated and interpolated according to fixed time windows so that they can be aligned at the same time granularity. S22. Unified Spatial Coordinates: All location information of various equipment and data points in the scenic area is uniformly converted to the same geographic coordinate system, and the entire scenic area is divided into several small areas. Each tourist location, sensor, and facility point is mapped to the corresponding grid to achieve spatial alignment. S23. Cleaning dirty data: Data from different sources often contains errors and noise. The system identifies these anomalies, removes obviously erroneous records, and fills in missing values ​​appropriately to ensure that the data entering subsequent stages is clean and reliable. S24. Standardization and Feature Extraction: The cleaned data format is transformed into a consistent structure. Convert sunny / rainy / snowy weather descriptions into numerical codes; Tagging tourists by their place of origin, age group, and spending level creates customer profiles; Extract useful features such as dwell time, movement speed, and popular areas from the original trajectory; S25. Writing to the Spatiotemporal Data Lake: All processed data is organized according to a "time + spatial grid" and stored in a specially designed storage system—the Spatiotemporal Data Lake. This system supports quick querying of data in any time period and any region, while preserving complete data lineage and update records for easy traceability and management.

4. The method for dynamic adjustment of tourist carrying capacity based on AI prediction as described in claim 1, characterized in that, In step S3, the spatiotemporal graph neural network STGNN is used to model the topological relationships between attractions and the flow patterns of tourists. A Transformer-based prediction module is introduced to capture long-term dependencies. The specific steps are as follows: S31. Spatiotemporal graph structure construction: The scenic spots, main road intersections, transportation hubs and important facilities in the scenic area are abstracted as graph nodes. Based on the connection paths between scenic spots and the actual transfer probability of tourists, a weighted adjacency matrix is ​​constructed to quantify the spatial association strength. A multi-dimensional feature vector is constructed for each node, including historical passenger flow sequence, real-time status data, node attributes and external environmental factors. S32.STGNN Spatial Topology and Short-Term Dependency Modeling: By aggregating the features of each node and its neighboring nodes through a graph convolutional network, the spatial association between attractions and the effect of visitor flow diffusion are encoded. A temporal convolutional network is applied to the feature sequence of each node to extract local temporal patterns and short-term change trends, and outputs a node state sequence that integrates spatiotemporal information. The features of each time step contain its spatial context and recent evolution information. S33. Transformer Long-Term Temporal Dependency Modeling: The spatiotemporal feature sequence output by STGNN is used as input. Global dependency capture is performed. Through a multi-head self-attention mechanism, direct correlation between features at any time step in the sequence is established. Long-term periodic patterns spanning days, weeks, and holidays are identified. After processing by a feedforward network and layer normalization, a high-order spatiotemporal feature representation with global temporal context is output. S34. Multi-step prediction and heat map generation: Based on the comprehensive features output by Transformer, the system decodes the data through a fully connected network to generate predicted visitor flow values ​​for each node at multiple future time steps. The predicted node values ​​are then mapped to a geospatial grid. An interpolation algorithm is used to generate a refined heat map of visitor flow distribution across the entire scenic area. Based on the gradient changes in the heat map, potential congestion areas and their evolution directions are identified, enabling early warning.

5. The method for dynamic adjustment of tourist carrying capacity based on AI prediction as described in claim 1, characterized in that, In step S4, the AI ​​model comprehensively considers: safety red line, experience threshold, ecological threshold, and facility load threshold, as detailed below; Safety red line: Maximum physical space capacity; Experience threshold: An alert is triggered when the predicted density in a certain area exceeds the threshold for a comfortable experience; Ecological threshold: For sensitive ecological areas, the number of people allowed to enter is dynamically adjusted based on environmental sensor data; Facility load threshold: The carrying capacity of related transportation, toilet, and garbage bin facilities.

6. The method for dynamic adjustment of tourist carrying capacity based on AI prediction as described in claim 1, characterized in that, In step S5, based on the predicted congestion level, location, and cause, the AI ​​decision engine (rule engine + reinforcement learning) automatically generates and recommends the optimal combination of control strategies. The specific steps are as follows: S51. Rule Matching: The rule engine quickly matches the preset emergency plan based on the predicted congestion level, location and cause to ensure that basic and safe control measures can be activated immediately; S52. Reinforcement Learning Optimization: Within the framework of the contingency plan, the reinforcement learning model conducts rapid simulation and deduction through a digital twin environment, fine-tunes policy parameters: increases or decreases the number of shuttle buses, adjusts the timing of information push, and evaluates the long-term effects of different policy combinations: experience, safety, and efficiency, and selects the optimization scheme with the highest overall benefit. S53. Execution and Evolution: After the optimized strategy is manually confirmed, it is issued for execution. The actual effect data after execution will be fed back to the reinforcement learning model, enabling it to continuously learn from experience and make more intelligent decisions in the future.

7. A dynamic control system for tourist carrying capacity based on AI prediction, characterized in that, The AI-based predictive dynamic control system for tourist carrying capacity is used to implement any one of the AI-based predictive dynamic control methods for tourist carrying capacity as described in claims 1-6; the AI-based predictive dynamic control system for tourist carrying capacity includes: Digital twin and multi-source sensing module: integrates a 3D model of the scenic area and accesses real-time data streams such as visitor trajectories, environmental sensors, facility status, and external weather / traffic / ticketing information; Spatiotemporal data fusion and governance module: performs spatiotemporal alignment, cleaning, and standardization on heterogeneous data from different systems, formats, and frequencies, and adds customer group tags; Passenger flow prediction and risk warning module: Using a spatiotemporal graph neural network + Transformer model, predict the heat map of passenger flow distribution in various areas of the scenic spot in the future, and identify potential congestion points and overloaded areas in real time; Dynamic bearing capacity assessment and resilience analysis module: Integrating four threshold categories—safety, experience, ecology, and facilities—it dynamically calculates the upper limit of bearing capacity for each area and introduces a "risk resilience index" to assess the system's resistance to pressure and its ability to recover under sudden events. Intelligent regulation decision-making and strategy generation module: Built-in strategy knowledge base; Combines rule engine and reinforcement learning to automatically generate hierarchical regulation schemes; The control implementation and effect feedback module: links the ticketing system, OTA platform, navigation app, in-venue broadcasting, mini-program, shuttle bus dispatching system, etc., to implement control measures; the evaluation layer: quantifies the effect of the strategy through A / B testing and key indicator monitoring; the feedback layer: feeds newly generated behavioral data and evaluation results back to the data lake to drive continuous iteration and optimization of the model.

Images