A service facility dynamic scheduling method based on tourist flow heat prediction

Abstract

The application discloses a kind of based on tourist dynamic line heat forecast service facility dynamic scheduling method, fusion history passenger flow, real-time positioning, video analysis and weather and other multi-source data, construct unified space-time grid;Through trajectory clustering mining typical tour path, and combined with LSTM etc. Time series model rolling prediction future tourist density heat map;Further introduce crowd portrait label identification old people, special groups of parents and children, and differentially calculate facility demand in demand mapping, and set vulnerable area minimum service guarantee constraint;Then adopt multi-agent reinforcement learning (MARL) collaborative optimization dynamic scheduling of facilities such as feeder car, mobile toilet, give consideration to tourist waiting time, scheduling cost and fairness;Through model predictive control (MPC) framework rolling execution scheduling instruction, and combined with A / B test continuously correct prediction and strategy.The application realizes facility "pre-judgment-dispatch-feedback-evolution" closed loop, significantly improves service response efficiency, resource utilization and tourist satisfaction.

Description

Technical Field

[0001] This invention relates to the field of service facility management, and in particular to a method for dynamic scheduling of service facilities based on heat map prediction of tourist movement. Background Technology

[0002] Traditional scenic area service facility scheduling typically employs static configuration or simple threshold-triggered responses (e.g., "add more facilities when the queue for a toilet exceeds 10 people"), which presents several problems: Lag and passivity: Relying on current or historical data, it fails to anticipate changes in visitor flow, resulting in consistently slow facility deployment and poor visitor experience (e.g., long queues, difficulty finding shuttle buses); "One-size-fits-all" allocation: Allocating resources solely based on the total number of visitors ignores the needs of special groups such as the elderly, children, and disabled individuals who are more sensitive to service radius and waiting time, leading to superficial balance but actual unfairness; Isolated and inefficient scheduling: Independent decision-making by each facility, lacking coordination, easily results in resource overcrowding or coverage blind spots, while failing to consider operational costs (e.g., high empty-running rates) and sustainability goals. This method integrates multi-source data for flow heat prediction, fairness modeling driven by visitor profiles, multi-agent collaborative dynamic scheduling, and closed-loop optimization using MPC and A / B testing. It systematically solves the key challenges of "inaccurate allocation, slow deployment, and uneven distribution" of service facilities in smart scenic areas, supporting efficient, inclusive, and sustainable cultural tourism operation and management. Summary of the Invention

[0003] A method for dynamic scheduling of service facilities based on heat map prediction of tourist movement includes the following steps; S1. Data Acquisition and Preprocessing: Multi-source data acquisition, including historical data: historical visitor flow, facility usage records, ticketing data; real-time data: Wi-Fi / Bluetooth probes, camera visual analysis, mobile signaling, APP location, sensor data (such as gate counting); external data: weather, holidays, scenic area activity schedules, public transportation timetables, and data cleaning and fusion: handling missing values ​​and outliers, unifying spatiotemporal coordinates (such as mapping data to a scenic area grid map), and aligning multi-source data to a unified time series and spatial grid; S2. Visitor Flow Analysis and Heat Map Prediction: Flow pattern mining, identification of typical tour routes (such as "classic routes" and "family routes") through trajectory clustering (such as DBSCAN), analysis of visitor dwell points (POI attraction model) and movement speed, and construction of heat map prediction models. Short-term prediction (future 0.5-2 hours): Using time series models (such as LSTM, GRU) or spatial-temporal graph neural networks (ST-GNN), combined with real-time pedestrian flow data, rolling predictions are performed, outputting predicted visitor density heat maps for each time period and each grid area in the future; S3. Service facility demand mapping: Based on heat map, tourist density is converted into facility demand (e.g., 1 shuttle bus per 100 people, X portable toilets per area). Considering facility service capacity, service radius, and tourist queuing tolerance, the predicted demand is compared with the current facility configuration to identify resource gaps or surpluses in different time periods / areas in the future. S4. Dynamic scheduling optimization: Modeling the scheduling problem, transforming facility scheduling into a dynamic resource allocation problem (such as vehicle routing problem (VRP) and facility layout problem). The objective function includes: minimizing tourist waiting time / dissatisfaction; Minimize scheduling costs (such as vehicle empty mileage); Maximize the balance of facility utilization; Constraints: number of facilities, movement speed, road capacity, scheduling time window. Multi-agent reinforcement learning (MARL) is employed to enable collaborative decision-making among facility agents; Pre-scheduling scenario: Using mixed integer programming (MIP) to find the globally optimal solution; S5. Scheduling Execution and Feedback Iteration: Generation and issuance of scheduling instructions, including generating facility movement instructions (e.g., "Move shuttle bus No. 2 to Area A") or resource allocation plans (e.g., "Add temporary toilets in Area B"); issuing instructions to staff or automated equipment (e.g., autonomous shuttle buses) through the scheduling system; monitoring deviations between actual passenger flow and forecasts, setting thresholds to trigger rescheduling (e.g., sudden increase in passenger flow in a certain area); using a Model Predictive Control (MPC) framework to continuously optimize subsequent scheduling strategies; collecting actual scheduling effect data (e.g., changes in visitor queuing time) to correct the prediction model and optimize parameters; and comparing the effectiveness of different scheduling strategies through A / B testing. S6. Fairness and Vulnerable Group Protection Mechanism: Introduce population profile tags (from APP registration information, ticketing type, historical behavior, etc.) into heat map and demand mapping to identify special demand groups and set fairness constraints, such as "at least one portable toilet must be kept within 500 meters of each accessible passage". Construct a vulnerable area protection mechanism: set a minimum service guarantee threshold for low-traffic but high-sensitivity areas (such as near medical points and mother and baby rooms) and add a fairness penalty term (such as minimizing the Gini coefficient) to the scheduling objective function.

[0004] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S1, the spatiotemporal coordinates are unified: the data is mapped to the scenic area grid map, and the multi-source data is aligned to a unified time series and spatial grid. The specific steps are as follows; S11. Establish a spatiotemporal grid framework: Divide the scenic area map into grids with unique IDs (e.g., by 50m×50m or by functional areas), and divide the time axis into slices of equal length (e.g., every 5 minutes or every hour is a time period). S12. Map all data to the grid: Through coordinate transformation and point-to-surface judgment algorithms, assign each piece of data with location information (such as GPS points, Wi-Fi connections) to a specific spatial grid, align the timestamps of all data to the nearest standard time slice, and aggregate high-frequency data (such as summation, averaging), while interpolating or preserving low-frequency data. S13. Fusion and Output: Weighted fusion or conflict resolution of multi-source data within the same grid-time slice (e.g., calibrating Wi-Fi data with gate data) is performed, and the final output is a structured spatiotemporal data table. Each row represents: (grid ID, time slice, estimated number of visitors, other indicators, etc.), providing a unified input for subsequent predictions.

[0005] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S2, the movement pattern mining is carried out by identifying typical tour routes through trajectory clustering DBSCAN, analyzing tourist dwell points and movement speeds, and constructing a heat prediction model. The time series model LSTM is used in conjunction with real-time pedestrian flow data for rolling prediction. The specific steps are as follows: S21. Track Cleaning and Standardization: The original tourist tracks from the APP, Wi-Fi, and cameras are sorted by time, abrupt changes are removed, duplicate positions are merged, and short-term missing data is interpolated to form continuous and valid tracks; S22. Identify typical tour routes: Use the DBSCAN clustering algorithm to group all tourist trajectories. Each cluster result represents a common route (such as "classic loop" or "family route"). Output several typical route templates for subsequent analysis and prediction reference. S23. Identifying Stop Points and Calculating Movement Speed: Stop point detection: Staying in an area for ≥3 minutes is considered a stop and marked as a potential POI (e.g., observation deck, restroom). The number of people staying in each area and the duration of stay are counted. The attractiveness of each POI is quantified. For each grid / POI, the following calculations are performed: Dwell intensity = Average stay duration × Number of people staying; Access frequency = Number of times accessed per unit of time; Path coverage = the number of typical paths in which it appears; Construct an attractiveness score: ,in Indicates the first The overall attractiveness score of each POI, , Weight parameters, satisfying The relative importance of dwell time intensity and visit frequency is used to adjust the relative importance of these factors. This represents the average length of time tourists spend in the area multiplied by the number of people staying (reflecting "in-depth experience"). It represents the number of times a data point is accessed per unit of time (reflecting its popularity). The average passage speed between grids is calculated based on historical data and dynamically adjusted according to the current flow of people (the more people there are, the slower the passage). S24. Construct a short-term heat map prediction model: Using the LSTM time series model as the core, input the visitor density of each grid in the last 30–60 minutes, as well as auxiliary information such as weather, holidays, and activity of typical routes, and output a grid-level visitor density heat map of the scenic area every 5–10 minutes for the next 30 minutes to 2 hours. Rolling prediction is adopted: every 5–10 minutes, the prediction is updated with the latest real-time data to maintain accuracy.

[0006] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S4, multi-agent reinforcement learning (MARL) is used to enable the various facility agents to make collaborative decisions, as detailed below; S41. Define the granularity of intelligent agents: Each mobile service facility (such as a shuttle bus or a portable toilet) is regarded as an independent intelligent agent. Each intelligent agent has the ability to perceive, make decisions and execute, and can respond to local information while achieving global coordination through wired communication. S42. Constructing the State Space: Each agent receives a comprehensive state vector at each decision moment, including: its own attributes (such as current battery level, remaining service capacity, and grid location), the real-time and predicted visitor density of the grid, the average number of people in the queue and the waiting time, and perceives the flow demand and facility gaps of several neighboring grids (e.g., within a radius of 200 meters). It also introduces global context information, such as whether it is a holiday or the weather conditions, as a shared environmental signal. S43. Reward Function Design: To balance efficiency, cost, and fairness, the reward consists of three weighted components: First, a service benefit component, which uses the reduction in average waiting time for tourists in the area as a positive incentive; second, an operating cost component, which applies a negative reward to travel distance or empty driving behavior to suppress unnecessary scheduling; and third, a fairness component, which provides an additional reward when the agent responds to service requests from the "vulnerable areas" marked in S6 (such as areas with a high concentration of elderly people or around mother and baby rooms). The weights of these three components can be flexibly adjusted according to the scenic area's management objectives, for example, emphasizing waiting time during peak periods and energy consumption control during off-peak seasons. S44. Local communication mechanism: A dynamic communication graph is constructed based on spatial proximity. When the distance between two facilities is less than a set threshold (e.g., 300 meters), they can exchange encoded state summaries (e.g., by aggregating neighbor information through attention mechanism or graph neural network). S45. Algorithm Training: Using the MARL framework, training is conducted in a digital twin simulation environment built based on historical data to simulate facility interactions under different passenger flow scenarios. After the strategy converges, it can be deployed to the actual system and supports online fine-tuning. S46. Execution Coordination: During the execution phase, each agent outputs action suggestions in parallel. A lightweight central coordinator performs conflict detection and priority arbitration, generates the final scheduling instruction, and issues it to the automated equipment or personnel. The entire MARL scheduling module runs on a rolling basis with a cycle of 5–10 minutes, seamlessly connecting with the Model Predictive Control (MPC) framework in S5 to form a closed-loop optimization of "prediction-decision-execution-feedback".

[0007] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S5, the Model Predictive Control (MPC) framework is used, and the effectiveness of different scheduling strategies is compared through A / B testing, as detailed below; Model predictive control (MPC) is used to achieve rolling scheduling: every 5-10 minutes, based on the latest passenger flow forecast and facility status, the scheduling plan for the next 30-60 minutes is optimized, but only the first step is executed (such as issuing the instruction "move bus No. 2 to area A"). The next cycle is re-planned based on the new data, so as to dynamically adapt to changes in passenger flow. An embedded A / B testing mechanism is implemented: Under similar scenarios (such as weekend mornings), different scheduling strategies are deployed for different regions or time periods (e.g., strategy A focuses on efficiency, strategy B focuses on fairness). Actual metrics (such as queuing time, empty mileage, and tourist satisfaction) are automatically collected, and their performance is judged through statistical comparison. The test results are not only used to select a better strategy for deployment, but also used to correct the heat map prediction model S2 and adjust the scheduling reward function S4, forming a closed-loop iteration of "execution-evaluation-learning-optimization".

[0008] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S6, population profile tags are introduced into heat forecasting and demand mapping to identify special demand groups and set fairness constraints. The specific steps are as follows: S61. Constructing a crowd profile: Based on data such as ticketing, APP registration, and trajectory behavior, label them with tags such as "elderly", "family with children", and "people with disabilities"; for tourists without explicit information, infer group attributes through behavioral patterns (such as slow movement and frequent stops in mother and baby rooms); S62. Group-specific heat map prediction: Incorporate profile features into the model to predict the spatiotemporal distribution of different groups and output detailed heat maps; S63. Differentiated Demand Mapping: Set stricter service standards for special groups (such as smaller service radius and lower queuing tolerance), and calculate facility demand separately for group types; S64. Identify vulnerable areas: Mark areas with a high proportion of special needs people, areas near accessibility facilities, or areas with a high number of historical complaints as “vulnerable areas”; S65. Embedding fairness constraints: In scheduling optimization, enforce minimum service in vulnerable areas (e.g., "at least one portable toilet per area"), and add a fairness term to the reward function or objective function to guide resources toward disadvantaged groups.

[0009] The beneficial effects of this invention are as follows: By predicting pedestrian hotspots in advance and dynamically allocating resources, waiting times are significantly reduced, especially ensuring service accessibility for special groups such as the elderly, children, and people with disabilities, thereby enhancing satisfaction and a sense of security. It avoids resources being idle in low-demand areas and scarce in high-demand areas, enabling precise on-demand deployment of facilities such as shuttle buses and portable toilets, increasing overall utilization by 20%–40% (depending on the scenario). Through collaborative route planning and rolling optimization, it reduces ineffective empty runs and redundant scheduling, saving manpower, energy, and maintenance costs; it also supports the priority scheduling of new energy equipment, contributing to green operations. By explicitly modeling the needs of "vulnerable areas" and special groups and embedding scheduling constraints and reward mechanisms, it ensures that service allocation balances efficiency and social inclusion, preventing the systemic neglect of vulnerable groups. Attached Figure Description

[0010] Figure 1 A flowchart of a dynamic scheduling method for service facilities based on tourist flow heat map prediction; Detailed Implementation

[0011] A method for dynamic scheduling of service facilities based on tourist flow heat map prediction, flowchart as follows: Figure 1 As shown, it includes the following steps; S1. Data Acquisition and Preprocessing: Multi-source data acquisition, including historical data: historical visitor flow, facility usage records, ticketing data; real-time data: Wi-Fi / Bluetooth probes, camera visual analysis, mobile signaling, APP location, sensor data (such as gate counting); external data: weather, holidays, scenic area activity schedules, public transportation timetables, and data cleaning and fusion: handling missing values ​​and outliers, unifying spatiotemporal coordinates (such as mapping data to a scenic area grid map), and aligning multi-source data to a unified time series and spatial grid; S2. Visitor Flow Analysis and Heat Map Prediction: Flow pattern mining, identification of typical tour routes (such as "classic routes" and "family routes") through trajectory clustering (such as DBSCAN), analysis of visitor dwell points (POI attraction model) and movement speed, and construction of heat map prediction models. Short-term prediction (future 0.5-2 hours): Using time series models (such as LSTM, GRU) or spatial-temporal graph neural networks (ST-GNN), combined with real-time pedestrian flow data, rolling predictions are performed, outputting predicted visitor density heat maps for each time period and each grid area in the future; S3. Service facility demand mapping: Based on heat map, tourist density is converted into facility demand (e.g., 1 shuttle bus per 100 people, X portable toilets per area). Considering facility service capacity, service radius, and tourist queuing tolerance, the predicted demand is compared with the current facility configuration to identify resource gaps or surpluses in different time periods / areas in the future. S4. Dynamic scheduling optimization: Modeling the scheduling problem, transforming facility scheduling into a dynamic resource allocation problem (such as vehicle routing problem (VRP) and facility layout problem). The objective function includes: minimizing tourist waiting time / dissatisfaction; Minimize scheduling costs (such as vehicle empty mileage); Maximize the balance of facility utilization; Constraints: number of facilities, movement speed, road capacity, scheduling time window. Multi-agent reinforcement learning (MARL) is employed to enable collaborative decision-making among facility agents; Pre-scheduling scenario: Using mixed integer programming (MIP) to find the globally optimal solution; S5. Scheduling Execution and Feedback Iteration: Generation and issuance of scheduling instructions, including generating facility movement instructions (e.g., "Move shuttle bus No. 2 to Area A") or resource allocation plans (e.g., "Add temporary toilets in Area B"); issuing instructions to staff or automated equipment (e.g., autonomous shuttle buses) through the scheduling system; monitoring deviations between actual passenger flow and forecasts, setting thresholds to trigger rescheduling (e.g., sudden increase in passenger flow in a certain area); using a Model Predictive Control (MPC) framework to continuously optimize subsequent scheduling strategies; collecting actual scheduling effect data (e.g., changes in visitor queuing time) to correct the prediction model and optimize parameters; and comparing the effectiveness of different scheduling strategies through A / B testing. S6. Fairness and Vulnerable Group Protection Mechanism: Introduce population profile tags (from APP registration information, ticketing type, historical behavior, etc.) into heat map and demand mapping to identify special demand groups and set fairness constraints, such as "at least one portable toilet must be kept within 500 meters of each accessible passage". Construct a vulnerable area protection mechanism: set a minimum service guarantee threshold for low-traffic but high-sensitivity areas (such as near medical points and mother and baby rooms) and add a fairness penalty term (such as minimizing the Gini coefficient) to the scheduling objective function.

[0012] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S1, the spatiotemporal coordinates are unified: the data is mapped to the scenic area grid map, and the multi-source data is aligned to a unified time series and spatial grid. The specific steps are as follows; S11. Establish a spatiotemporal grid framework: Divide the scenic area map into grids with unique IDs (e.g., by 50m×50m or by functional areas), and divide the time axis into slices of equal length (e.g., every 5 minutes or every hour is a time period). S12. Map all data to the grid: Through coordinate transformation and point-to-surface judgment algorithms, assign each piece of data with location information (such as GPS points, Wi-Fi connections) to a specific spatial grid, align the timestamps of all data to the nearest standard time slice, and aggregate high-frequency data (such as summation, averaging), while interpolating or preserving low-frequency data. S13. Fusion and Output: Weighted fusion or conflict resolution of multi-source data within the same grid-time slice (e.g., calibrating Wi-Fi data with gate data) is performed, and the final output is a structured spatiotemporal data table. Each row represents: (grid ID, time slice, estimated number of visitors, other indicators, etc.), providing a unified input for subsequent predictions.

[0013] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S2, the movement pattern mining is carried out by identifying typical tour routes through trajectory clustering DBSCAN, analyzing tourist dwell points and movement speeds, and constructing a heat prediction model. The time series model LSTM is used in conjunction with real-time pedestrian flow data for rolling prediction. The specific steps are as follows: S21. Track Cleaning and Standardization: The original tourist tracks from the APP, Wi-Fi, and cameras are sorted by time, abrupt changes are removed, duplicate positions are merged, and short-term missing data is interpolated to form continuous and valid tracks; S22. Identify typical tour routes: Use the DBSCAN clustering algorithm to group all tourist trajectories. Each cluster result represents a common route (such as "classic loop" or "family route"). Output several typical route templates for subsequent analysis and prediction reference. S23. Identifying Stop Points and Calculating Movement Speed: Stop point detection: Staying in an area for ≥3 minutes is considered a stop and marked as a potential POI (e.g., observation deck, restroom). The number of people staying in each area and the duration of stay are counted. The attractiveness of each POI is quantified. For each grid / POI, the following calculations are performed: Dwell intensity = Average stay duration × Number of people staying; Access frequency = Number of times accessed per unit of time; Path coverage = the number of typical paths in which it appears; Construct an attractiveness score: ,in Indicates the first The overall attractiveness score of each POI, , Weight parameters, satisfying The relative importance of dwell time intensity and visit frequency is used to adjust the relative importance of these factors. This represents the average length of time tourists spend in the area multiplied by the number of people staying (reflecting "in-depth experience"). It represents the number of times a data point is accessed per unit of time (reflecting its popularity). The average passage speed between grids is calculated based on historical data and dynamically adjusted according to the current flow of people (the more people there are, the slower the passage). S24. Construct a short-term heat map prediction model: Using the LSTM time series model as the core, input the visitor density of each grid in the last 30–60 minutes, as well as auxiliary information such as weather, holidays, and activity of typical routes, and output a grid-level visitor density heat map of the scenic area every 5–10 minutes for the next 30 minutes to 2 hours. Rolling prediction is adopted: every 5–10 minutes, the prediction is updated with the latest real-time data to maintain accuracy.

[0014] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S4, multi-agent reinforcement learning (MARL) is used to enable the various facility agents to make collaborative decisions, as detailed below; S41. Define the granularity of intelligent agents: Each mobile service facility (such as a shuttle bus or a portable toilet) is regarded as an independent intelligent agent. Each intelligent agent has the ability to perceive, make decisions and execute, and can respond to local information while achieving global coordination through wired communication. S42. Constructing the State Space: Each agent receives a comprehensive state vector at each decision moment, including: its own attributes (such as current battery level, remaining service capacity, and grid location), the real-time and predicted visitor density of the grid, the average number of people in the queue and the waiting time, and perceives the flow demand and facility gaps of several neighboring grids (e.g., within a radius of 200 meters). It also introduces global context information, such as whether it is a holiday or the weather conditions, as a shared environmental signal. S43. Reward Function Design: To balance efficiency, cost, and fairness, the reward consists of three weighted components: First, a service benefit component, which uses the reduction in average waiting time for tourists in the area as a positive incentive; second, an operating cost component, which applies a negative reward to travel distance or empty driving behavior to suppress unnecessary scheduling; and third, a fairness component, which provides an additional reward when the agent responds to service requests from the "vulnerable areas" marked in S6 (such as areas with a high concentration of elderly people or around mother and baby rooms). The weights of these three components can be flexibly adjusted according to the scenic area's management objectives, for example, emphasizing waiting time during peak periods and energy consumption control during off-peak seasons. S44. Local communication mechanism: A dynamic communication graph is constructed based on spatial proximity. When the distance between two facilities is less than a set threshold (e.g., 300 meters), they can exchange encoded state summaries (e.g., by aggregating neighbor information through attention mechanism or graph neural network). S45. Algorithm Training: Using the MARL framework, training is conducted in a digital twin simulation environment built based on historical data to simulate facility interactions under different passenger flow scenarios. After the strategy converges, it can be deployed to the actual system and supports online fine-tuning. S46. Execution Coordination: During the execution phase, each agent outputs action suggestions in parallel. A lightweight central coordinator performs conflict detection and priority arbitration, generates the final scheduling instruction, and issues it to the automated equipment or personnel. The entire MARL scheduling module runs on a rolling basis with a cycle of 5–10 minutes, seamlessly connecting with the Model Predictive Control (MPC) framework in S5 to form a closed-loop optimization of "prediction-decision-execution-feedback".

[0015] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S5, the Model Predictive Control (MPC) framework is used, and the effectiveness of different scheduling strategies is compared through A / B testing, as detailed below; Model predictive control (MPC) is used to achieve rolling scheduling: every 5-10 minutes, based on the latest passenger flow forecast and facility status, the scheduling plan for the next 30-60 minutes is optimized, but only the first step is executed (such as issuing the instruction "move bus No. 2 to area A"). The next cycle is re-planned based on the new data, so as to dynamically adapt to changes in passenger flow. An embedded A / B testing mechanism is implemented: Under similar scenarios (such as weekend mornings), different scheduling strategies are deployed for different regions or time periods (e.g., strategy A focuses on efficiency, strategy B focuses on fairness). Actual metrics (such as queuing time, empty mileage, and tourist satisfaction) are automatically collected, and their performance is judged through statistical comparison. The test results are not only used to select a better strategy for deployment, but also used to correct the heat map prediction model S2 and adjust the scheduling reward function S4, forming a closed-loop iteration of "execution-evaluation-learning-optimization".

[0016] Furthermore, a dynamic scheduling method for service facilities based on tourist flow heat map prediction is proposed. In step S6, population profile tags are introduced into heat forecasting and demand mapping to identify special demand groups and set fairness constraints. The specific steps are as follows: S61. Constructing a crowd profile: Based on data such as ticketing, APP registration, and trajectory behavior, label them with tags such as "elderly", "family with children", and "people with disabilities"; for tourists without explicit information, infer group attributes through behavioral patterns (such as slow movement and frequent stops in mother and baby rooms); S62. Group-specific heat map prediction: Incorporate profile features into the model to predict the spatiotemporal distribution of different groups and output detailed heat maps; S63. Differentiated Demand Mapping: Set stricter service standards for special groups (such as smaller service radius and lower queuing tolerance), and calculate facility demand separately for group types; S64. Identify vulnerable areas: Mark areas with a high proportion of special needs people, areas near accessibility facilities, or areas with a high number of historical complaints as “vulnerable areas”; S65. Embedding fairness constraints: In scheduling optimization, enforce minimum service in vulnerable areas (e.g., "at least one portable toilet per area"), and add a fairness term to the reward function or objective function to guide resources toward disadvantaged groups.

[0017] Example 2. Taking Yunfeng Mountain Scenic Area, a large mountainous 5A-level scenic area, as an example: This method will be deployed in the Yunfeng Mountain scenic area; S1: The system integrates with the ticketing system (including visitor age and ticket type), APP location logs, 200 Wi-Fi probes, 50 AI cameras (identifying crowd flow and movement speed), and gate counters, and incorporates weather data from the meteorological bureau, holiday calendars, and the scenic area's daily performance schedule. After cleaning, all data is mapped to a 10m×10m grid map of the scenic area, with time aligned to a 10-minute granularity. S2: Using DBSCAN to cluster 100,000 historical trajectories, three typical routes were identified: "Eastern Uphill Express Line", "Western Family Loop Line", and "Cable Car-Viewing Platform Classic Line". A thermal prediction model was built based on LSTM, and the heat map of tourist density for the next 90 minutes was generated by inputting the pedestrian flow and weather data for each grid in the past hour. S3: Based on the heat map calculation, facility requirements are as follows: 1 shuttle bus is needed for every 100 people (service radius 300m), and 1 portable toilet is needed for every 200 people (queue tolerance threshold 8 minutes). The system detected that from 10:00 AM to 11:30 AM, there will be a shortage of 2 shuttle buses and 1 toilet in the mountaintop viewing area. S4: Initiate MARL scheduling, with 8 shuttle buses and 5 portable toilets acting as agents for collaborative decision-making. The objective function is weighted as follows: waiting time (0.5), empty-run cost (0.3), and fairness (0.2). Simultaneously, MIP pre-generates the baseline scheduling plan for the day. Optimization result: Shuttle buses 3 and 5 are instructed to assemble at the mountaintop 15 minutes earlier, and one spare toilet is moved from the warehouse to the east side of the viewing platform. S5: Instructions are issued to the autonomous shuttle vehicles and dispatcher terminals. At 10:20, the actual passenger flow is detected to be 15% higher than predicted, triggering the rescheduling threshold. MPC immediately replans the strategy for the next 60 minutes. Data collected that day, such as queue length and visitor ratings, are used for A / B testing, verifying that the new strategy reduces the average waiting time by 22% compared to the old rules. S6: Ticketing and behavioral tags identify that "senior citizen groups" are concentrated on the western route, and this area is marked as a vulnerable zone. The system mandates that there is always one available toilet within 200 meters of this area, and adds a fairness reward to the MARL rewards for agents that respond to requests in this area, ensuring that services are not ignored due to low traffic.

Claims

1. A method for dynamic scheduling of service facilities based on tourist flow heat map prediction, characterized in that, Includes the following steps; S1. Data Acquisition and Preprocessing: Multi-source data acquisition, including historical data: historical visitor flow, facility usage records, and ticketing data; Real-time data: Wi-Fi / Bluetooth probes, camera visual analysis, mobile signaling, APP location, sensor data; external data: weather, holidays, scenic area activity schedules, public transportation timetables, and data cleaning and fusion: handling missing values ​​and outliers, unifying spatiotemporal coordinates: mapping data to a scenic area grid map, aligning multi-source data to a unified time series and spatial grid; S2. Visitor Flow Analysis and Heat Map Prediction: Flow pattern mining, identification of typical tour routes through trajectory clustering DBSCAN, analysis of visitor dwell points and movement speed, and construction of a heat map prediction model. Using the time series model LSTM, combined with real-time pedestrian flow data, rolling prediction is performed to output the predicted visitor density heat map for each time period and each grid area in the future. S3. Service Facility Demand Mapping: Based on heat maps, tourist density is transformed into facility demand. Factors such as facility service capacity, service radius, and tourist queuing tolerance are considered. The predicted demand is compared with the current facility configuration to identify resource gaps and surpluses in different time periods / regions in the future. S4. Dynamic Scheduling Optimization: Modeling the scheduling problem, transforming facility scheduling into a dynamic resource allocation problem, including vehicle routing problem (VRP) and facility layout problem; The objective functions include: minimizing visitor waiting time / dissatisfaction, minimizing scheduling costs, and maximizing facility utilization balance. Constraints include: number of facilities, travel speed, road capacity, and scheduling time window. Multi-agent reinforcement learning (MARL) is employed to enable collaborative decision-making among facility agents. Pre-scheduling scenario: Using Mixed Integer Programming (MIP) to find the globally optimal solution; S5. Scheduling Execution and Feedback Iteration: Generation and issuance of scheduling instructions, generation of facility movement instructions and resource configuration plans, issuance to staff and automated equipment through the scheduling system, monitoring of deviations between actual passenger flow and forecasts, setting thresholds to trigger rescheduling, adopting the Model Predictive Control (MPC) framework, continuously optimizing subsequent scheduling strategies, collecting actual scheduling effect data to correct the prediction model and optimize parameters, and comparing the effectiveness of different scheduling strategies through A / B testing. S6. Fairness and Vulnerable Group Protection Mechanism: Introduce population profile tags in heat forecasting and demand mapping to identify special demand groups and set fairness constraints: Construct a vulnerable area protection mechanism, set a minimum service guarantee threshold for low-traffic but highly sensitive areas, and add a fairness penalty term to the scheduling objective function.

2. The method for dynamic scheduling of service facilities based on tourist flow heat map prediction as described in claim 1, characterized in that, In step S1, the spatiotemporal coordinates are unified: the data is mapped to the scenic area grid map, and the multi-source data is aligned to a unified time series and spatial grid. The specific steps are as follows; S11. Establish a spatiotemporal grid framework: Divide the scenic area map into grids with unique IDs, and divide the timeline into slices of equal length; S12. Map all data to the grid: Through coordinate transformation and point-to-surface judgment algorithms, assign each piece of data with location information to a specific spatial grid, align the timestamps of all data to the nearest standard time slice, and aggregate high-frequency data: summation and averaging, and interpolate and preserve low-frequency data; S13. Fusion and Output: Weighted fusion or conflict resolution is performed on multi-source data within the same grid-time slice, and the final output is a structured spatiotemporal data table. Each row represents: grid ID, time slice, and estimated number of tourists, providing a unified input for subsequent predictions.

3. The method for dynamic scheduling of service facilities based on tourist flow heat map prediction as described in claim 1, characterized in that, In step S2, the movement pattern mining is carried out by identifying typical tour routes through trajectory clustering DBSCAN, analyzing tourist dwell points and movement speeds, and constructing a heat prediction model. The time series model LSTM is used in conjunction with real-time pedestrian flow data for rolling prediction. The specific steps are as follows: S21. Track Cleaning and Standardization: The original tourist tracks from the APP, Wi-Fi, and cameras are sorted by time, abrupt changes are removed, duplicate positions are merged, and short-term missing data is interpolated to form continuous and valid tracks; S22. Identify typical tour routes: Use the DBSCAN clustering algorithm to group all tourist trajectories. Each clustering result represents a common route: classic loop, family route, etc. Output several typical route templates for subsequent analysis and prediction reference. S23. Identifying Stop Points and Calculating Movement Speed: Stop point detection; staying in an area for ≥3 minutes is considered a stop and marked as a potential POI: observation deck, restroom. Statistical analysis of the number of people and duration of stay in each area is performed, and POI attractiveness is quantified. For each grid / POI, the following calculations are made: Dwell intensity = Average stay duration × Number of people staying; Access frequency = Number of times accessed per unit of time; Path coverage = the number of typical paths in which it appears; Construct an attractiveness score: ,in Indicates the first The overall attractiveness score of each POI. , Weight parameters, satisfying The relative importance of dwell time intensity and visit frequency is used to adjust the relative importance of these factors. This represents the average length of time tourists spend in the area multiplied by the number of people staying, reflecting the depth of their experience. This indicates the number of times a data point is accessed per unit of time (reflecting popularity; the average passage speed between grids is calculated based on historical data and dynamically adjusted according to current pedestrian flow). S24. Construct a short-term heat map prediction model: Using the LSTM time series model as the core, input the visitor density of each grid in the last 30–60 minutes, as well as auxiliary information such as weather, holidays, and activity of typical routes, and output a grid-level visitor density heat map of the scenic area every 5–10 minutes for the next 30 minutes to 2 hours. Rolling prediction is adopted: every 5–10 minutes, the prediction is updated with the latest real-time data to maintain accuracy.

4. The method for dynamic scheduling of service facilities based on tourist flow heat map prediction as described in claim 1, characterized in that, In step S4, multi-agent reinforcement learning (MARL) is used to enable the various facility agents to make collaborative decisions, as detailed below; S41. Define the granularity of intelligent agents: Each mobile service facility is regarded as an independent intelligent agent. Each intelligent agent has the ability to perceive, make decisions and execute, and can respond to local information, while achieving global coordination through wired communication. S42. Constructing the State Space: Each agent receives a comprehensive state vector at each decision moment, including: its own attributes, the real-time and predicted tourist density of the grid it is in, the average number of people in the queue and the waiting time, and perceives the flow demand and facility gaps of several neighboring grids. It also introduces global context information as a shared environmental signal. S43. Reward Function Design: To balance efficiency, cost, and fairness, the reward consists of three weighted components: first, a service benefit component, which uses the reduction in average waiting time for tourists in the area as a positive incentive; second, an operating cost component, which applies a negative reward to travel distance or empty runs to suppress unnecessary scheduling; and third, a fairness component, which provides an additional reward when the agent responds to service requests from vulnerable areas marked in S6: areas with high concentrations of elderly people and areas around mother and baby rooms. The weights of these three components are flexibly adjusted according to the scenic area's management objectives. S44. Local communication mechanism: A dynamic communication graph is constructed based on spatial proximity. When the distance between two facilities is less than a set threshold of 300 meters, they can exchange encoded state summaries. S45. Algorithm Training: Using the MARL framework, training is conducted in a digital twin simulation environment built based on historical data to simulate facility interactions under different passenger flow scenarios. After the strategy converges, it can be deployed to the actual system and supports online fine-tuning. S46. Execution Coordination: During the execution phase, each agent outputs action suggestions in parallel. A lightweight central coordinator performs conflict detection and priority arbitration, generates the final scheduling instruction, and issues it to the automated equipment. The entire MARL scheduling module runs on a rolling basis with a cycle of 5–10 minutes, seamlessly connecting with the Model Predictive Control (MPC) framework in S5 to form a closed-loop optimization of "prediction-decision-execution-feedback".

5. The method for dynamic scheduling of service facilities based on tourist flow heat map prediction as described in claim 1, characterized in that, In step S5, the Model Predictive Control (MPC) framework is used, and the effectiveness of different scheduling strategies is compared through A / B testing, as detailed below; Model Predictive Control (MPC) is used to achieve rolling scheduling: every 5–10 minutes, based on the latest passenger flow forecast and facility status, the scheduling plan for the next 30–60 minutes is optimized, but only the first step is executed. The next cycle is re-planned based on new data, thereby dynamically adapting to changes in passenger flow. Embedded A / B testing mechanism: Under similar scenarios, different scheduling strategies are deployed for different regions or time periods, and actual indicators are automatically collected. The superiority or inferiority is judged by statistical comparison. The test results are not only used to select a better strategy to go online, but also used to correct the heat prediction model S2 and adjust the scheduling reward function S4, forming a closed-loop iteration of "execution-evaluation-learning-optimization".

6. The method for dynamic scheduling of service facilities based on tourist flow heat map prediction as described in claim 1, characterized in that, In step S6, population profile tags are introduced into heat forecasting and demand mapping to identify special demand groups and set fairness constraints. The specific steps are as follows: S61. Constructing a crowd profile: Based on data such as ticketing, APP registration, and trajectory behavior, label them as elderly, families with children, and people with disabilities; for tourists without explicit information, infer group attributes through behavioral patterns: slow movement and frequent stops in mother and baby rooms; S62. Group-specific heat map prediction: Incorporate profile features into the model to predict the spatiotemporal distribution of different groups and output detailed heat maps; S63. Differentiated Demand Mapping: Set stricter service standards for special groups and calculate facility requirements separately according to group type; S64. Identify vulnerable areas: Mark areas with a high proportion of special needs people, areas near accessibility facilities, or areas with a high number of historical complaints as vulnerable areas; S65. Embedding fairness constraints: In scheduling optimization, a minimum service level must be guaranteed for vulnerable areas: at least one portable toilet per area, and a fairness term is added to the reward function or objective function to guide resources toward vulnerable groups.

Images