Tourist city traffic hub-attraction linkage scheduling method based on spatiotemporal big data

Abstract

The application discloses a tourism city traffic hub-attraction linkage scheduling method based on space-time big data, relates to the technical field of smart city and tourism traffic management, and first collects and fuses multi-source heterogeneous space-time data; secondly, strong driving factors are screened through Granger causality test, a causality Informer model embedded with external event Token is constructed, long-time sequence prediction of multi-hub passenger throughput is realized, a dynamic hypergraph neural network is constructed in combination with a meta-learning framework, space-time distribution prediction of tourist flow is realized, a multi-task joint learning framework is introduced, a space-time feature extractor is shared, a five-dimensional stakeholder utility function and a hierarchical deep reinforcement learning are used to generate a multi-objective collaborative scheduling strategy; and finally, continuous iteration optimization is realized through instruction issuing and closed-loop feedback. The tourism city traffic hub-attraction linkage scheduling method effectively improves the tourism passenger flow prediction accuracy and scene adaptability, realizes global integrated linkage scheduling, has decision-making interpretability and continuous optimization capability, and is suitable for landing application of multi-hub cluster of tourism city.

Description

Technical Field

[0001] This invention relates to the fields of smart city and urban tourism traffic management technology, specifically a method for coordinated scheduling of urban transportation hubs and scenic spots based on spatiotemporal big data. Background Technology

[0002] Against the backdrop of the construction of the Hainan Free Trade Port, air transport, as a core channel connecting the island province to the outside world, has formed a deeply mutually reinforcing relationship with the tourism industry. Haikou Meilan, Sanya Phoenix, and Qionghai Bo'ao airports together constitute the Hainan airport cluster, respectively serving as the provincial capital hub, tourism core, and convention and business center. Their passenger throughput directly determines the scale and pace of tourism development across the island. Tourist cities like Hainan face severe passenger flow impacts during holidays and peak tourist seasons. Large numbers of tourists arrive at transportation hubs such as airports and high-speed rail stations and then rapidly converge on core scenic spots, causing a chain reaction of problems such as hub congestion, road network paralysis, overcrowding at scenic spots, and a decline in tourist experience.

[0003] The closest existing technology mainly involves matching and scheduling highway networks by combining the integration degree of the nearest roads to scenic spots with the resource level of the scenic spots. For example, the Chinese patent document, application number 202310289974.5, entitled "A Method for Evaluating the Potential of All-Area Tourism Based on Highway Spatial Structure," mentions: Vectorizing regional roads and establishing topological relationships to form a highway network; then, using spatial syntax to calculate the global integration degree of highways, and combining the integration degree of the nearest roads to scenic spots with the resource level of the scenic spots, using the factor product method to calculate the attractiveness of the scenic spots; simultaneously, considering factors such as road level, direction, and speed to establish a cost matrix, and using the Floyd algorithm to solve for the shortest path between scenic spots; furthermore, based on the path length classification and the attractiveness of endpoint scenic spots, calculating the highway linkage capability between scenic areas; finally, using the linkage capability between all scenic spots as weights, using kernel density analysis to convert the linear path influence into regional area influence, generating the evaluation result of the potential of all-area tourism, providing a spatial quantitative basis for the layout of tourism industry and infrastructure planning.

[0004] Currently, there are also hybrid models used for passenger throughput prediction, such as GWO-SARIMA-XGBoost described in the publicly available literature "Research on Passenger Throughput Prediction and Development Strategies of Hainan Free Trade Port Airport". However, the above-mentioned existing technologies still have the following problems in practical applications: First, regardless of whether it is a single model or a hybrid model of traditional statistics and machine learning, most of them use a single time series model, which is difficult to effectively cope with the strong seasonality of tourist flow. They do not take into account multiple dimensions such as regional tourism industry, competing transportation modes, and macro policies, resulting in insufficient overall prediction and scheduling accuracy. Secondly, transportation hubs, urban traffic management departments, and scenic area management entities are basically operating independently, with no data sharing or business coordination, making it impossible to achieve timely early warning and coordinated dispatch. Moreover, dispatch plans are mostly optimized based on traffic efficiency, without fully considering factors such as passenger experience, commercial benefits, and the workload of regional service personnel. Finally, for regions like the Hainan Free Trade Port, which have multiple airports and various tourist destinations, existing designs are mostly designed for a single hub or a single scenic spot. This makes it difficult to achieve passenger flow diversion between multiple hubs and coordinated control between multiple scenic spots, thus failing to fully unleash the overall efficiency of regional tourism. Summary of the Invention

[0005] The purpose of this invention is to provide a tourism city transportation hub-scenic spot linkage scheduling method based on spatiotemporal big data, so as to solve the problems of low accuracy of passenger throughput prediction for multiple hubs in tourism cities, insufficient integration of tourism industry and policy and other multi-dimensional factors, single traditional scheduling objectives, and inability to achieve passenger flow diversion between multiple hubs and coordinated control between multiple scenic spots in the above-mentioned background technology.

[0006] To this end, this invention provides a tourism city transportation hub-scenic spot linkage scheduling method based on spatiotemporal big data, including the following steps: S1, Multi-source heterogeneous spatiotemporal data collection and fusion: collecting multi-transportation hub operation data, tourism industry operation data, urban road network traffic data, scenic spot passenger flow data, external environment data, and macroeconomic policy data within the region, and preprocessing the collected data; the collected data includes six categories, the first category being multi-transportation hub operation data, including flight / train arrival and departure plans and actual data, takeoff and landing times, on-time departure rate, passenger flow density within the hub, and real-time location and status of connecting transport capacity; the second... The data is categorized into six types: First, tourism industry data, including reservation and ticket sales data for various scenic spots, overnight visitor numbers, tourism revenue, and duty-free consumption data; second, urban road network data, including road network topology GIS data, floating car GPS data, traffic light timing data, and road congestion status; third, external environment data, including meteorological data, holiday arrangements, information on major events, and social media tourism popularity; fourth, macroeconomic policy data, including regional GDP, transportation CPI, air passenger market index, and free trade port policy event indicators; and fifth, competitive transportation data, including railway passenger volume and port passenger throughput.

[0007] S2. Joint Prediction of Passenger Throughput at Multiple Hubs: Based on stakeholder theory, a five-dimensional passenger flow impact indicator system is constructed. Strong causal driving factors are screened through Granger causality tests. A causal Informer model embedding external event tokens is used to jointly predict the time-segmented passenger throughput of multiple transportation hubs within the region, outputting the passenger flow scale, seasonal fluctuations, and temporal distribution characteristics of each hub within a preset period. The five-dimensional passenger flow impact indicator system includes: Tourist metrics: number of overnight visitors to the scenic area, tourism revenue, and transportation price index; Business partner dimensions: number of flight takeoffs and landings, flight on-time departure rate, number of cooperative routes, and ground service provider capacity; Staffing dimensions: number of airport and scenic area service personnel, staff skill levels, and service quality ratings; Social stakeholder dimensions: regional GDP, railway passenger volume, port passenger throughput, policy event indicators; Investor perspectives: Air passenger market index, investment scale in the cultural and tourism industry, and duty-free consumption quota.

[0008] S3, Integrated Spatiotemporal Digital Twin Modeling of Hubs, Road Networks, and Scenic Spots: This system constructs a four-layer digital twin system comprising a geometric and physical layer, a data perception layer, a model analysis layer, and an interactive application layer. It generates dynamic GIS attributes and builds a spatiotemporal knowledge graph for synchronous virtual-real mapping and scenario-based simulation of multiple transportation hubs, the entire road network, and multiple core scenic spots. The geometric and physical layer constructs BIM models of transportation hub buildings, high-precision GIS models of urban road networks, and 3D terrain and building models of scenic spots, integrating physical rule engines such as traffic flow cellular automata models and pedestrian flow social force models. The data perception layer receives real-time multi-source data streams from S1, driving the twin to synchronize with the real world, achieving virtual-real mapping of passenger flow, vehicle flow, and facility status. The model analysis layer embeds passenger flow prediction, path calculation, and scheduling optimization algorithms, endowing the digital twin with analytical and predictive capabilities. The interactive application layer provides a 3D visualization dashboard, supporting real-time monitoring, early warning display, and "What-if" simulation.

[0009] S4. Spatiotemporal distribution prediction of tourist flow: Based on the passenger throughput prediction results of each hub, combined with historical tourist migration patterns, real-time road network operation status and real-time popularity of scenic spots, a dynamic hypergraph neural network combined with a meta-learning framework is used to predict the number of tourists flowing from each transportation hub to each scenic spot, the distribution of arrival time and the route selection preferences at a preset time step in the future; and to predict the number of tourists flowing from each transportation hub to each core scenic spot, the distribution of arrival time and the main route selection at multiple time steps such as 15 minutes, 30 minutes, 1 hour and 3 hours in the future.

[0010] S5. Dual-task joint training and verification: Construct a multi-task joint learning framework for throughput prediction and flow direction prediction, share the underlying spatiotemporal feature extractor, set dual-task prediction heads, and use an uncertainty-weighted loss function to balance the training weights of the tasks; realize the closed-loop error correction of macro total quantity constraining micro distribution and micro distribution verifying macro total quantity.

[0011] S6. Multi-objective collaborative scheduling optimization: Based on the five-dimensional stakeholder value system, a quantifiable multi-objective optimization utility function is constructed. A hierarchical deep reinforcement learning framework is adopted to generate an integrated collaborative scheduling strategy that includes hub connection capacity allocation, dynamic control and route guidance of traffic lights on the road network, and staggered peak reservation and diversion guidance for scenic spots. S7. Spatiotemporal attribution interpretability analysis: Spatiotemporal attribution interpretability analysis specifically involves decoupling the attention weights of the model in both spatiotemporal dimensions: in the spatial dimension, the contribution of each hub, road segment, and scenic spot to the final prediction result is calculated, and a heat map of the impact of passenger flow across the entire region is generated; in the temporal dimension, the key time windows that have the greatest impact on future predictions in historical periods are located.

[0012] When abnormal passenger flow peaks or troughs are predicted, attribution backtracking is automatically triggered, traversing external event factors and spatiotemporal node states, identifying core driving factors by ranking feature importance, and generating an attribution report containing "key events, transmission paths, and degree of impact," providing dispatchers with clear decision-making basis.

[0013] S8. Dispatch instruction issuance and closed-loop feedback: The optimal dispatch strategy is transformed into standardized execution instructions and issued to various transportation operation units, traffic management departments and scenic area management entities. Real-time data on passenger flow, road conditions and scenic area operation are collected after dispatch execution. Through error quantification calculation, the throughput prediction model, passenger flow distribution prediction model and dispatch optimization algorithm are calibrated in reverse.

[0014] Preferably, the specific operation flow for preprocessing the collected data in step S1 is as follows: S11. Perform format parsing, topic classification, and metadata registration on the multi-source raw data collected from various interfaces to establish a unified data index; S12. Identify and correct missing and outlier values ​​in the original data to eliminate data noise; S13. Unify the spatial coordinate system and time reference of all data; S14. Encode and unify the units of different types of features to make them suitable for the input requirements of prediction models and scheduling algorithms; S15. Using spatial and temporal units as indices, associate and concatenate all features to construct a standardized model input dataset that integrates time, space, and features, which will then be used to input the model for prediction.

[0015] Preferably, in step S2, the construction steps of the causal Informer model are as follows: S21. Granger causality test was performed on the passenger throughput time series and five-dimensional influencing indicators. The significance level was set at 0.05. Strong causal driving factors that passed the significance test were screened out, and redundant correlation features were removed. S22. Construct the Informer network backbone, use a probabilistic sparse self-attention mechanism to capture the temporal dependence of long-sequence passenger flow, and compress the feature dimension through a self-attention distillation mechanism to focus on core temporal features. S23. Supplement the external event embedding layer by encoding sudden events such as policy adjustments, large-scale events, and extreme weather into independent tokens, which are then concatenated with time-series features and input into the decoder to achieve early perception and fitting of sudden passenger flow impacts. S24. With the goal of minimizing prediction error, the AdamW optimization algorithm is used to train the model, and the mean absolute error (MAE) is used as the basic loss function to output the multi-hub time-segmented throughput prediction results.

[0016] Preferably, in step S3, the digital twin system is specifically configured as follows: Geometric Physical Layer: Constructs BIM models of transportation hubs, high-precision GIS models of road networks, and 3D terrain models of scenic spots, embedding traffic flow cellular automata models and social force models of pedestrian flow to support physical rule simulation; Data perception layer: Real-time access to multi-source data streams from step S1, updates the twin status at a fixed frequency, and completes the virtual-real synchronization mapping of passenger flow, vehicle flow, and facility status; Model Analysis Layer: Embedded with throughput prediction model, hypergraph flow prediction model, multi-task collaboration module and scheduling optimization algorithm, giving the twin the ability to analyze and extrapolate; Interactive application layer: Provides a 3D visualization cockpit, supporting full-domain situational monitoring, early warning information display, and multi-scenario simulation.

[0017] The steps for generating dynamic GIS attributes and constructing a spatiotemporal knowledge graph are as follows: Based on static GIS road network, hub, and scenic area layers, real-time floating car data, hub passenger flow data, and scenic area ticketing data are integrated to dynamically calculate and add dynamic attributes such as real-time road speed, predicted travel time, congestion index, number of visitors in scenic areas, saturation, and queuing time; the formula for calculating the congestion index is as follows: In the formula For the road segment at time t The congestion index, For the road segment at time t The actual travel time. For the road segment at time t Free passage time; By abstracting transportation hubs, vehicles, road sections, scenic spots, and tourist groups into entities, and abstracting carrying, connecting, going to, and operating status into relationships, a time-stamped four-tuple spatiotemporal knowledge network is constructed for passenger flow reasoning and path association analysis.

[0018] Preferably, in step S5, the specific steps for the dual-task joint training and verification are as follows: A shared underlying spatiotemporal feature extraction backbone network is used to extract periodic trend features in the time dimension and correlation and interaction features in the spatial dimension, respectively. Set up two parallel task heads to output the hub throughput prediction value and the hub and scenic area OD flow direction prediction matrix, respectively; Using a homoscedastic uncertainty-weighted loss function, the loss weights for the two tasks are automatically learned. The weight calculation formula is as follows: In the formula For throughput prediction loss, To predict the loss of the flow, These are two artificially determined observation noise standard deviations, which are dynamically updated and adjusted during the training process.

[0019] Preferably, in step S4, the construction steps of the dynamic hypergraph neural network and the meta-learning framework are as follows: S41. Abstract the H transportation hubs, M key road nodes, and S scenic spots within the region as graph nodes to construct a spatiotemporal hypergraph structure. , where the set of nodes There are a total of N = H + M + S superedges; It dynamically updates over time, connects physically adjacent nodes, and dynamically aggregates non-adjacent nodes with similar functions and related traffic patterns to express the warning coupling relationship between multiple nodes; S42. Dynamically adjust the superedge weights and topology based on real-time road conditions, scenic spot popularity, and passenger flow tides to adapt to the second-level dynamic changes in the road network and passenger flow. S43. Initialize the new scenario model parameters based on the model parameters of historical similar scenarios, and quickly fine-tune them with a small number of samples to complete the passenger flow prediction for newly opened scenic spots, new hubs, or extreme scenarios during holidays.

[0020] Preferably, in step S6, the expression for the multi-objective optimization utility function is: In the formula - Let be the weighting coefficient, satisfying , For passenger experience utility, the expression is: ,in Average waiting time for connecting to the hub The average travel time from the hub to the scenic area This represents the average queuing time at the scenic area. , and The corresponding time weights; For the benefit of business partners, the expression is: In the formula The average load factor for connecting transport capacity. To improve the on-time departure rate, To improve the conversion rate of tourist spending at scenic spots, , and Weighted by benefits; For employee effectiveness, the expression is: In the formula Specifically, this refers to the average number of customers served per person. The service load coefficient for each position, and These are the weighting coefficients; For social security and regional synergy benefits, the expression is: In the formula The average congestion coefficient of the road network. Forecast visitor flow for scenic area S This represents the maximum instantaneous carrying capacity of the scenic area. The indicator function is a negative penalty when the tourist flow in the scenic area exceeds its carrying capacity, and a positive value for the overall traffic efficiency of the road network. For investment value utility, the expression is: In the formula For passenger flow growth rate, To improve the utilization rate of public transportation resources, and These are the weighting coefficients.

[0021] Preferably, a hierarchical deep reinforcement learning framework is used to solve for the optimal scheduling strategy. The framework consists of two layers: The high-level meta controller, based on long-term passenger flow forecast results, performs global resource allocation and subsystem objective decomposition, setting optimization sub-objectives and constraints for the hub end, road network end, and scenic area end respectively; The low-level execution controller, comprising a hub scheduling agent, a road network management agent, and a scenic area diversion agent, generates specific scheduling actions based on short-cycle real-time prediction data under high-level objective constraints. The hub scheduling agent is used to decide the departure frequency of tourist shuttle buses from various airports to different scenic spots, the capacity tilt coefficient of ride-hailing services, and the airport bus scheduling plan. The road network management agent is used to decide the traffic light timing adjustment plan at key intersections, the route guidance strategy of navigation apps, and temporary control measures for congested road sections. The scenic area diversion agent is used to decide the ticket reservation release schedule of various scenic spots, the timing of issuing passenger flow warnings, and the strategy of recommending alternative scenic spots in the surrounding area.

[0022] Preferably, in step S8, the specific steps for error quantization calculation to reverse-calibrate the throughput prediction model, passenger flow distribution prediction model, and scheduling optimization algorithm are as follows: The system collects operational data on actual throughput of hubs, actual vehicle speeds on road sections, and actual park entry volumes within a fixed time window. It calculates prediction deviations and uses Mean Absolute Percentage Error (MAPE) to quantify prediction accuracy. The quantification expression is as follows: In the formula For predicted values, These are actual observed values. This represents the number of data samples.

[0023] Preferably, before issuing the scheduling instruction in step S8, a scheduling strategy verification step is also included, specifically: The initial scheduling strategy is input into the integrated spatiotemporal digital twin system of hubs, road networks and scenic spots. Through the built-in traffic flow cellular automata model and human flow social force model, the system simulates and extrapolates the operation status of passenger flow and vehicle flow in the whole area after the scheduling strategy is executed, and calculates the multi-objective utility value in the simulation scenario. If the utility value is lower than the preset threshold, the multi-objective collaborative scheduling optimization module will be returned to regenerate the strategy; if the utility value meets the preset threshold requirement, the strategy will be determined as the optimal scheduling strategy and issued for execution.

[0024] The proposed method for coordinated scheduling of urban transportation hubs and scenic spots based on spatiotemporal big data has the following advantages: This invention eliminates redundant factors through Granger causality screening and combines it with an Informer model with event embedding to effectively capture long-term trends and sudden event impacts. It solves the limitations of single time series models in taking into account linear trends, strong seasonality, and nonlinear external shocks, as well as the problem of traditional hybrid models lagging in response to nonlinear factors such as policies and activities. It is particularly suitable for regions like the Hainan Free Trade Port, where tourism is highly seasonal and events such as free trade port policies and duty-free shopping on the island have a significant impact. It can output high-precision multi-hub time-sharing throughput forecasts, providing a reliable advance prediction basis for coordinated scheduling and avoiding the passive scheduling problems caused by large deviations and strong lags in traditional forecasting methods. By using dynamic hypergraph neural networks to predict tourist flow, the problem that traditional ordinary graph networks can only express pairwise relationships is solved. Combined with the meta-learning framework, it can achieve rapid adaptation of small samples in new scenarios, accurately identify tourist identities, predict specific destinations and arrival times, and provide refined spatiotemporal basis for differentiated transportation capacity allocation and staggered peak flow in scenic spots. This invention transforms stakeholder theory into a quantifiable five-dimensional multi-objective utility function, which dynamically adjusts weights according to peak and off-peak seasons to balance the concerns of five stakeholders: passenger experience, business partner benefits, employee efficiency, social security, and investment value. Coupled with a hierarchical deep reinforcement learning framework, a high-level meta-controller is responsible for overall resource coordination and objective decomposition, while low-level agents handle refined execution at hubs, road networks, and scenic areas, ensuring both overall synergy and local flexibility. Before issuing scheduling strategies, the effects are simulated using traffic flow and pedestrian flow models; if the utility is not met, the system iterates again, avoiding the secondary congestion risks caused by blind scheduling and improving the scientific nature and feasibility of scheduling decisions. Transforming previously fragmented and isolated passenger flow management into a holistic and systematic approach can effectively mitigate peak passenger flow during peak seasons, alleviate congestion at hubs and road networks, and reduce the safety risks of overloading at scenic spots. Simultaneously, through precise capacity matching and resource allocation, it improves connecting transport efficiency, reduces resource idleness, and drives growth in scenic spots and duty-free consumption. Ultimately, it achieves multiple improvements in tourist experience, operational efficiency, industrial benefits, and safety assurance. This approach is particularly suitable for implementation in island-type tourist destinations like the Hainan Free Trade Port, which features multiple hub clusters and scenic spots, and possesses strong promotional value and socio-economic benefits. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the overall process of the linkage scheduling method of the present invention. Figure 2 This is a schematic diagram of the multi-task joint learning framework structure of the present invention; Figure 3 This is a schematic diagram of the dynamic spatiotemporal hypergraph topology of the present invention; Figure 4 This is a schematic diagram of the causal Informer passenger throughput prediction model of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described in this specification are merely for explaining the invention and are not intended to limit the invention.

[0028] Example: Please see Figure 1-4Taking the Hainan Free Trade Port as an example, the region includes three major aviation hubs: Haikou Meilan International Airport, Sanya Phoenix International Airport, and Qionghai Boao International Airport. It covers eight core scenic spots, including Haikou Qilou Old Street, Sanya Yalong Bay, and Qionghai Boao Conference Site. It has typical tourist city characteristics such as strong seasonality of passenger flow, multiple hub distribution, and significant policy influence.

[0029] First, data is continuously collected through multiple interfaces, including hub data, tourism data, road network data, environmental and policy data, and competitive transportation data; Hub data is directly connected to the A-CDM systems of the three major airports to obtain flight schedules / actual arrival times, takeoffs and landings, on-time departure rates, and terminal passenger density; it also connects with public transportation, taxi, and ride-hailing platforms to obtain real-time GPS data and passenger status of shuttle vehicles. Tourism data is directly connected to the ticketing systems of the Hainan Provincial Department of Culture and Tourism and various scenic spots to obtain data on overnight visitors, tourism revenue, scenic spot reservations, real-time number of visitors, and queue times for the entire province and its cities and counties; it also connects with the duty-free shop system to obtain data on duty-free consumption on the island. Road network data is directly accessed from floating car GPS data, checkpoint data, and traffic light timing data of the Hainan Ring Expressway and urban main roads; a GIS vector layer of the entire provincial road network is loaded. Environmental and policy data directly obtains information on weather forecasts, holiday arrangements, and major events such as the Boao Forum; it collects macroeconomic prosperity indices, transportation CPI, and air passenger market indices; and it marks events such as the implementation of free trade port policies and adjustments to duty-free policies on the island. Competitive traffic data was obtained by collecting passenger volume data from Hainan Railway and port passenger throughput data. All data were preprocessed, unified into the CGCS2000 coordinate system and UTC timestamp, and spatiotemporal alignment was completed. Outliers were removed using the 3σ principle. Missing values ​​within 1 hour were filled in using linear interpolation, and missing values ​​beyond 1 hour were filled in using historical data from the same period. Weather levels and holiday markers were specially coded. Numerical features such as passenger flow and vehicle speed were standardized using min-max. All features were stitched together using a 500m grid and 15min as the window unit to construct a three-dimensional standardized dataset for subsequent model calls. Based on stakeholder theory, 18 passenger flow influencing indicators were constructed across five dimensions: passengers, business partners, employees, society, and investors. First, Granger causality tests were performed on the 18 indicators and the passenger throughput series of the three major airports, with a significance level set at 0.05. Twelve strong causal driving factors, including flight takeoffs and landings, offshore duty-free allowance, and policy event markers, were selected, while weakly correlated redundant features were eliminated.

[0030] See Figure 4A causal Informer prediction model was constructed. The encoder inputs historical throughput time-series data, filtered causal factor features, and event embedding tokens. A probabilistic sparse self-attention mechanism was employed for efficient feature extraction from long sequences. Redundant information was compressed through a self-attention distillation layer, and the decoder fused historical time-series features with features of known future events to output the prediction results. The training set consisted of data from 2012 to 2023, the validation set consisted of data from 2024, and the test set consisted of data from 2025. The AdamW optimizer was used with an initial learning rate of 0.001 and a batch size of 32. Testing showed that the model achieved R² values ​​of 0.942, 0.978, and 0.915 on the test sets of the three major airports, respectively, representing improvements of 2.2%, 0.5%, and 1.8% compared to the traditional GWO-SARIMA-XGBoost model. Particularly noteworthy were the reduction in prediction bias of over 30% for events such as the Boao Forum and adjustments to the offshore duty-free policy. The model outputs monthly trend forecasts for the next 12 months and hourly time-sharing forecasts for the next 72 hours, accurately identifying peak passenger flow during key periods such as the Spring Festival, summer vacation, and National Day.

[0031] Next, modeling and simulation are performed. At this point, the geometric and physical layer loads BIM models of the three major airport terminals, high-precision GIS data of the entire road network of Hainan Province, and 3D models of eight core scenic spots. It has a built-in SUMO traffic simulation engine and a social force model of pedestrian flow, supporting micro-level simulation of vehicle and pedestrian movement. The data perception layer receives real-time data streams every 5 minutes, displaying hub passenger flow density and road congestion status in the 3D scene using heat maps, and displaying the number of people in the scenic area and queuing time using numerical labels, achieving synchronization between virtual and real status. At the same time, dynamic GIS attributes are calculated: the actual travel time of each road segment is calculated based on real-time floating car data, and the congestion index CI is obtained by combining the free-flow travel time. A spatiotemporal knowledge graph is constructed, abstracting hubs, road segments, scenic spots, and tourists into entities, and abstracting carrying, connecting, and going into relationships, forming a four-tuple knowledge network with timestamps to support passenger flow reasoning and path association analysis.

[0032] Based on airport throughput prediction results, 23 nodes—3 airports, 12 main road intersections, and 8 core scenic areas—were constructed as hypergraph nodes. Hyperedges were divided into two categories: physically connected hyperedges, connecting adjacent nodes on the same road, and functionally related hyperedges, dynamically aggregating scenic areas and hub-linked nodes with similar passenger flow patterns. A gating update unit was then set up to update hyperedge weights every 15 minutes based on real-time traffic time and scenic area saturation, dynamically adjusting the hypergraph topology. During the training phase, four scenarios—peak season, off-season, holidays, and large-scale events—were used as meta-tasks, and the MAML algorithm was used to learn general initialization parameters. The model outputs an OD (Original Demand) flow matrix for the next 3 hours at 15-minute granularity, accurately predicting the arrival time and scale of passenger flow at each scenic area.

[0033] The temporal branch employs a 3-layer dilated causal convolution to extract multi-scale temporal patterns, while the spatial branch uses a 2-layer hypergraph convolution to extract higher-order spatial correlations. The fusion of these two branches yields general spatiotemporal features. The upper layer consists of two parallel task heads: a throughput prediction head outputting time-series throughput for three airports, with the mean squared error (MSE) as the loss function; and a flow direction prediction head outputting a 3×8 OD flow matrix, with the mean absolute percentage error (MAPE) as the loss function. Homoscedastic uncertainty-weighted loss automatically balances the weights of the two tasks, eliminating the need for manual parameter tuning. During training, the global constraints of total volume prediction guide the flow direction distribution to conform to passenger flow conservation logic, while the local spatiotemporal patterns of flow direction prediction improve the precision of total volume prediction; the two iterate collaboratively. Comparison shows that multi-task joint training reduces throughput prediction MAPE by 4.2% and flow direction prediction MAPE by 6.7% compared to independent training of the two models, effectively suppressing error accumulation in a serial architecture.

[0034] A multi-objective utility function is constructed based on a five-dimensional stakeholder value system, with weights dynamically adjusted according to peak and off-peak seasons: during peak seasons... , Focusing on passenger experience and safety; during the off-season , The focus is on operational efficiency.

[0035] Based on 3-hour forecasts, the high-level meta-controller sets global objectives: ensuring scenic spots do not exceed their maximum capacity, average waiting time at transportation hubs is ≤20 minutes, and average vehicle speed on main roads is ≥30km / h during the Spring Festival peak season. It decomposes the constraints of the three subsystems. In the lower-level three agents, the hub scheduling agent decides the frequency of tourist bus departures and the capacity allocation coefficient for ride-hailing services; the road network management agent decides traffic light timing and route guidance strategies; and the scenic spot diversion agent decides the timing of ticket reservations and recommends alternative scenic spots. The three agents perform strategy search in a digital twin simulation environment, and through hundreds of iterations, select the scheduling combination that maximizes the overall utility U.

[0036] Under normal circumstances, the attribution module generates a heat map of the overall passenger flow impact, intuitively displaying the contribution of each hub, road segment, and scenic spot to the overall passenger flow and marking the core impact nodes. When an abnormal passenger flow peak is predicted, attribution backtracking is automatically triggered: it traverses external event factors and spatiotemporal node states, calculates feature importance through SHAP values, identifies core driving factors, and generates an attribution report including transmission paths and the degree of impact. After the scheduling plan is generated, an attribution analysis of the plan's effectiveness is output simultaneously, illustrating the contribution percentage of each scheduling action to alleviating congestion and improving the experience, providing a clear basis for management decision-making.

[0037] Finally, the overall operational status is simulated, and multi-objective utility values ​​are calculated. If the targets are met, the system is implemented; otherwise, it iterates again. Standardized interfaces are used to send additional bus services to bus companies, traffic light timing adjustment suggestions to traffic management platforms, reservation quota adjustment instructions to scenic spot ticketing systems, and passenger flow warnings and detour suggestions to map and tourism platforms.

[0038] Actual operational data is collected every 30 minutes to calculate the deviation between the predicted MAPE and the scheduling effect. If the deviation exceeds 10%, online incremental fine-tuning of the model is triggered. At the same time, the actual scheduling effect is fed back to the reinforcement learning reward function to continuously optimize the strategy. For example, if the diversion effect of a certain scenic spot is found to be lower than expected, the reward weight of the scenic spot diversion action will be increased in the next iteration to continuously improve the scheduling effect.

[0039] The above-mentioned end-to-end solution enables integrated, intelligent, and interpretable coordinated management and control of transportation hubs and scenic spots in tourist cities. It effectively mitigates peak tourist flows during peak seasons, alleviates congestion, reduces safety risks, and improves operational efficiency and industrial benefits. It has significant promotional value and socio-economic benefits.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data, characterized by: Includes the following steps: S1. Multi-source heterogeneous spatiotemporal data collection and fusion: Collect operational data of multiple transportation hubs, tourism industry operation data, urban road network traffic data, scenic spot visitor flow data, external environment data and macroeconomic policy data within the region, and preprocess the collected data; S2. Joint prediction of passenger throughput of multiple hubs: Based on stakeholder theory, a five-dimensional passenger flow impact index system is constructed. Strong causal driving factors are screened through Granger causality test. A causal Informer model with embedded external event token is used to jointly predict the passenger throughput of multiple transportation hubs in the region at different times. The passenger flow scale, seasonal fluctuation and time distribution characteristics of each hub within the preset period are output. S3, integrated spatiotemporal digital twin modeling of hubs, road networks and scenic spots: construct a four-layer digital twin system including a geometric physical layer, a data perception layer, a model analysis layer and an interactive application layer, generate dynamic GIS attributes and construct a spatiotemporal knowledge graph for virtual and real synchronous mapping and scenario-based simulation of multiple transportation hubs, the entire road network and multiple core scenic spots; S4. Spatiotemporal distribution prediction of tourist flow: Based on the passenger throughput prediction results of each hub, combined with historical tourist migration patterns, real-time road network operation status and real-time popularity of scenic spots, a dynamic hypergraph neural network combined with a meta-learning framework is used to predict the number of tourists flowing from each transportation hub to each scenic spot, the distribution of arrival time and the route selection preference under a preset time step in the future. S5. Dual-task joint training verification: Construct a multi-task joint learning framework for throughput prediction and flow direction prediction, share the underlying spatiotemporal feature extractor, set up dual-task prediction heads, and use an uncertainty weighted loss function to balance the task training weights. S6. Multi-objective collaborative scheduling optimization: Based on the five-dimensional stakeholder value system, a quantifiable multi-objective optimization utility function is constructed. A hierarchical deep reinforcement learning framework is adopted to generate an integrated collaborative scheduling strategy that includes hub connection capacity allocation, dynamic control and route guidance of traffic lights on the road network, and staggered peak reservation and diversion guidance for scenic spots. S7. Interpretability Analysis: Embedded attribution analysis module, used to output a heatmap of the contribution of each influencing factor to the prediction results in real time; S8. Dispatch instruction issuance and closed-loop feedback: The optimal dispatch strategy is transformed into standardized execution instructions and issued to various transportation operation units, traffic management departments and scenic area management entities. Real-time data on passenger flow, road conditions and scenic area operation are collected after dispatch execution. Through error quantification calculation, the throughput prediction model, passenger flow distribution prediction model and dispatch optimization algorithm are calibrated in reverse.

2. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: The specific operation flow for preprocessing the collected data in step S1 is as follows: S11. Perform format parsing, topic classification, and metadata registration on the multi-source raw data collected from various interfaces to establish a unified data index; S12. Identify and correct missing and outlier values ​​in the original data to eliminate data noise; S13. Unify the spatial coordinate system and time reference of all data; S14. Encode and unify the units of different types of features to make them suitable for the input requirements of prediction models and scheduling algorithms; S15. Using spatial and temporal units as indices, associate and concatenate all features to construct a standardized model input dataset that integrates time, space, and features, which will then be used to input the model for prediction.

3. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: In step S2, the steps for constructing the causal Informer model are as follows: S21. Perform Granger causality tests on the passenger throughput time series and five-dimensional influencing indicators, screen out strong causal driving factors that pass the significance test, and eliminate redundant correlation features. S22. Construct an Informer network architecture, use a probabilistic sparse self-attention mechanism to capture the temporal dependence of long-sequence passenger flow, and compress the feature dimension through a self-attention distillation mechanism. S23. Encode policy adjustments, large-scale events, and extreme weather events into independent tokens, concatenate them with time-series features, and input them into the decoder; S24. With the goal of minimizing prediction error, an adaptive learning rate optimization algorithm is used to train the model and output the multi-hub time-segmented throughput prediction results.

4. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: In step S3, the specific structure of the digital twin system is as follows: Geometric Physical Layer: Constructs BIM models of transportation hubs, high-precision GIS models of road networks, and 3D terrain models of scenic spots, embedding traffic flow cellular automata models and social force models of pedestrian flow to support physical rule simulation; Data perception layer: Real-time access to multi-source data streams from step S1, updates the twin status at a fixed frequency, and completes the virtual-real synchronization mapping of passenger flow, vehicle flow, and facility status; Model Analysis Layer: Embedded with throughput prediction model, hypergraph flow prediction model, multi-task collaboration module and scheduling optimization algorithm, giving the twin the ability to analyze and extrapolate; Interactive application layer: Provides a 3D visualization cockpit, supporting full-domain situational monitoring, early warning information display, and multi-scenario simulation.

5. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: In step S5, the specific steps for the dual-task joint training and verification are as follows: A shared underlying spatiotemporal feature extraction backbone network is used to extract periodic trend features in the time dimension and correlation and interaction features in the spatial dimension, respectively. Set up two parallel task heads to output the hub throughput prediction value and the hub and scenic area OD flow direction prediction matrix, respectively; Using a homoscedastic uncertainty-weighted loss function, the loss weights for the two tasks are automatically learned. The weight calculation formula is as follows: In the formula For throughput prediction loss, To predict the loss of the flow, These are two artificially determined observation noise standard deviations, which are dynamically updated and adjusted during the training process.

6. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: In step S4, the construction steps of the dynamic hypergraph neural network and the meta-learning framework are as follows: S41. Abstract transportation hubs, road nodes, and scenic spots into graph nodes to construct a spatiotemporal hypergraph structure; where hyperedges connect physically adjacent nodes and dynamically aggregate non-neighboring nodes with similar functions and related flow patterns; S42. Dynamically adjust the superedge weights and topology based on real-time road conditions, scenic spot popularity, and passenger flow tides to adapt to the second-level dynamic changes in the road network and passenger flow. S43. Initialize the new scenario model parameters based on the model parameters of historical similar scenarios, and quickly fine-tune them with a small number of samples to complete the passenger flow prediction for newly opened scenic spots, new hubs, or extreme scenarios during holidays.

7. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: In step S6, the expression for the multi-objective optimization utility function is: In the formula - Let be the weighting coefficient, satisfying , For passenger experience utility, specifically, it is the negative weighted sum of hub waiting time, transit time, and scenic spot queuing time; For the benefit of business partners, specifically, it is a positive weighted sum of the load factor of connecting transport capacity, flight punctuality rate, and scenic spot consumption conversion rate; For employee efficiency, specifically, it is a positive weighted sum of the average number of customers served per person and the degree of service load balance; For the sake of social security and regional synergy, when the tourist flow exceeds the carrying capacity of the scenic area, the penalty value is negative, while the overall traffic efficiency of the road network is positive. For investment value utility, it is specifically a positive weighted sum of passenger flow growth rate and resource utilization rate.

8. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 7, characterized in that: The hierarchical deep reinforcement learning framework includes: The high-level meta controller, based on long-term passenger flow forecast results, performs global resource allocation and subsystem objective decomposition, setting optimization sub-objectives and constraints for the hub end, road network end, and scenic area end respectively; The low-level execution controller, which includes hub scheduling agents, road network management agents, and scenic area diversion agents, generates specific scheduling actions based on short-cycle real-time prediction data under the constraints of high-level objectives.

9. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: In step S8, the specific steps for error quantification calculation to reverse-calibrate the throughput prediction model, passenger flow distribution prediction model, and scheduling optimization algorithm are as follows: The system collects operational data on actual throughput of hubs, actual vehicle speeds on road sections, and actual park entry volumes within a fixed time window. It calculates prediction deviations and uses Mean Absolute Percentage Error (MAPE) to quantify prediction accuracy. The quantification expression is as follows: In the formula For predicted values, These are actual observed values. This represents the number of data samples.

10. The tourism city transportation hub-scenic area linkage scheduling method based on spatiotemporal big data according to claim 1, characterized in that: Before issuing the scheduling instruction in step S8, a scheduling strategy verification step is also included, specifically: The initial scheduling strategy is input into the integrated spatiotemporal digital twin system of hubs, road networks and scenic spots. Through the built-in traffic flow cellular automata model and human flow social force model, the system simulates and extrapolates the operation status of passenger flow and vehicle flow in the entire area after the scheduling strategy is executed, and calculates the multi-objective utility value in the simulation scenario. If the utility value is lower than the preset threshold, the strategy will be regenerated by returning to the multi-objective collaborative scheduling optimization module. If the utility value meets the preset threshold requirement, the strategy will be determined as the optimal scheduling strategy and executed.

Images