Scenic spot dynamic resource scheduling and personalized path recommendation method based on deep reinforcement learning

Abstract

The application discloses a scenic spot dynamic resource scheduling and personalized path recommendation method based on deep reinforcement learning, and belongs to the cross technical field of intelligent tourism and deep reinforcement learning. The method collects space-air-ground multi-source data and calculates the space attraction interaction strength between scenic spots; a heterogeneous topological graph of the scenic spot is constructed, a system state vector is generated through graph convolution coding; a deep deterministic policy gradient network is constructed, a multi-objective cooperative reward is calculated based on Nash bargaining; an executable scheduling instruction is generated by using an artificial potential field; a personalized recommendation path is generated based on information entropy; resource competition allocation is realized through double auction; the scheduling instruction and the recommendation path are cooperatively output; congestion entropy change is evaluated based on a dissipative structure; the expected loss test is used to predict the tourist satisfaction degree and feedback to the policy network to form a closed-loop optimization. The application realizes real-time adaptive decision of resource scheduling and personalization and equalization of path recommendation, and improves the resource utilization efficiency and the tourist satisfaction degree.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of smart tourism and deep reinforcement learning, specifically involving a method for dynamic resource scheduling and personalized path recommendation in scenic areas based on deep reinforcement learning. Background Technology

[0002] As the number of tourists visiting scenic areas continues to grow, the distribution of visitor flow within these areas exhibits significant spatial and temporal imbalances: popular attractions experience high visitor concentrations and frequent queues during peak hours, while some attractions suffer from sparse visitor numbers and underutilized resources. To alleviate these imbalances, scenic areas generally allocate shuttle buses and other transportation resources for visitor flow management and utilize information technology to provide tour guidance. However, visitor flow in scenic areas is dynamically influenced by various factors such as weather, holidays, and visitor preferences. Resource scheduling and route recommendations require real-time perception of the overall visitor flow situation, coordination of the interests of various parties, and continuous adaptive optimization. This places higher demands on the real-time nature, precision, and intelligence of scheduling decisions. Existing technologies have already explored several avenues for the intelligentization of scenic area resource scheduling and tour guidance.

[0003] According to patent application CN120146295A, published on June 13, 2025, a smart cultural tourism scenic area optimization management system and method based on big data is disclosed. This method acquires the overall layout map of the scenic area and marks the tour route. Based on the ancillary attractions along the route, the tour route is divided into several segments in sequence. Average visitor flow data within each segment is acquired and time-stamped. Based on the average visitor flow data at a single time point, the experience mobility index of each segment is calculated. The experience mobility index is then analyzed to output early warnings, and the required visitor flow is calculated and allocated, thereby balancing the load on each segment and optimizing the allocation of scenic area resources.

[0004] Patent application CN112417753A, published on February 26, 2021, discloses a method for joint scheduling of urban public transportation resources. This method focuses on the coordinated scheduling of bus systems and shared bicycle systems. Based on recorded data on passenger flow at different times and locations, and changes in passenger flow carried by various modes of transportation, it pre-constructs a demand flow map of passenger traffic changing over time. Then, using the current station status and the predicted future demand flow map as the system state, it employs reinforcement learning techniques to solve for the long-term optimal coordinated scheduling strategy to avoid local greedy algorithms and to promptly schedule other transportation resources to meet user travel needs when a certain transportation service is temporarily unavailable.

[0005] However, existing technologies have the following limitations:

[0006] First, the data collection dimensions are limited, lacking analysis of spatial gravitational interaction between attractions. Existing technologies rely on historical statistical records from ground-based collection methods, failing to effectively integrate communication networks and ground sensor data sources, resulting in blind spots in the perception of overall tourist flow distribution within the scenic area. Furthermore, existing technologies treat each line segment or station as a relatively independent statistical unit, without establishing a spatial gravitational interaction model between attractions. This makes it impossible to quantify the attraction and repulsion relationships between different attractions, and thus difficult to accurately depict the dynamic migration trends of tourist flow.

[0007] Secondly, the system's state representation capabilities are insufficient, making it difficult to capture the spatial correlation characteristics of the scenic area's topology. Existing technologies characterize the system state using line segment passenger flow indicators or station demand flow maps, failing to abstract the scenic area into a topological graph structure containing heterogeneous nodes such as attractions, transportation hubs, service facilities, and entrances / exits. Furthermore, they do not utilize graph convolutional neural networks to deeply encode the coupling relationships between nodes in multiple dimensions such as geographic space, tour logic, and passenger flow correlation. This results in a one-sided representation of the system state, which cannot support refined scheduling decisions.

[0008] Third, the lack of a multi-objective conflict coordination mechanism makes it difficult to balance the interests of all parties. Existing technologies use transportation efficiency or load balancing as optimization objectives in resource scheduling, without considering the game relationship of interests among multiple stakeholders such as scenic area operators, tourist groups, scenic area management, and equipment management. When the pressure of passenger flow surges, extreme optimization of a single objective often comes at the expense of the interests of other stakeholders, leading to unfair resource allocation and localized deterioration of service quality.

[0009] Fourth, resource allocation lacks market-based competitive regulation and is devoid of effect evaluation and closed-loop feedback. Existing technologies rely on centralized instructions or early warning rules for resource allocation, lacking the ability to dynamically bid and achieve balanced allocation based on market mechanisms. At the same time, existing technologies have not established quantitative evaluation methods for the orderly effects of scheduling strategies, nor have they incorporated tourist satisfaction feedback into the training process of decision-making models, making it difficult for the system to continuously evolve and adaptively optimize as the operating environment changes. Summary of the Invention

[0010] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a method for dynamic resource scheduling and personalized path recommendation in scenic areas based on deep reinforcement learning.

[0011] To achieve the above-mentioned objectives of this invention, this invention provides a method for dynamic resource scheduling and personalized path recommendation in scenic areas based on deep reinforcement learning, comprising the following steps:

[0012] S1: Collect multi-source data and align it in time and space, calculate the intensity of spatial gravitational interaction between attractions, and generate a gravitational interaction thermodynamic matrix;

[0013] S2, construct a heterogeneous topology graph of the scenic area, and use a graph convolutional network to propagate and aggregate node features to generate a system state vector;

[0014] S3: Construct a policy network and value network with deep deterministic policy gradients to learn the mapping from system state to scheduling actions;

[0015] S4 calculates the moment reward for multi-agent collaboration based on Nash bargaining game to balance the conflicting interests of multiple parties;

[0016] S5 converts scheduling actions into executable scheduling instructions that include path planning and speed control based on an artificial potential field.

[0017] S6, based on information entropy to measure path diversity, combined with tourist interest profiles to generate recommended paths that balance customer flow;

[0018] S7 uses a dual auction mechanism to solve for the market clearing price and allocates transportation capacity proportionally to achieve resource competitive allocation.

[0019] S8 calculates the service supply-demand ratio of each scenic spot and performs spatiotemporal alignment and collaborative output of scheduling instructions and recommended routes;

[0020] S9, based on dissipative structure theory, calculates the congestion entropy and entropy change of scenic areas, and evaluates the orderliness effect of scheduling strategies;

[0021] S10 predicts tourist satisfaction based on the expectation failure theory and feeds it back to the policy network to achieve closed-loop optimization.

[0022] In the preferred embodiment of the above technical solution, the multi-source data in S1 includes ground sensor data and mobile communication data, and the data is uniformly mapped to the scenic area's geographic coordinate system to construct a time-series data cube; the spatial gravitational interaction intensity is jointly determined by the real-time visitor flow density of the two scenic spots, the distance between the scenic spots, the landscape similarity, and the current average congestion index of the path, and the spatial gravitational interaction intensity of all scenic spot pairs is summarized to generate the gravitational interaction heat matrix.

[0023] In the preferred embodiment of the above technical solution, the nodes of the heterogeneous topology graph of the scenic area in S2 include scenic spot nodes, transportation hub nodes, service facility nodes, and entrance / exit nodes, and the edge attributes include physical length, current traffic speed, theoretical traffic capacity, and real-time congestion level; the node features are propagated and aggregated through a two-layer graph convolutional network, then global average pooling is performed, and concatenated with global statistical features to obtain the system state vector.

[0024] In the preferred embodiment of the above technical solution, the output layer of the policy network in S3 normalizes the action values ​​and then maps them to the actual physical value range of each action dimension through a linear transformation; during the training process, a target network soft update mechanism and an experience replay mechanism are introduced, the value network is updated with mean square time difference error, and a reference signal is provided for the gradient update of the policy network.

[0025] In the preferred embodiment of the above technical solution, four stakeholders are identified: the operator, the tourist group, the scenic area management, and the equipment management. The difference between the standardized benefits and the retention utility of each party is multiplied by a weighted average according to the bargaining weight, and the difference between the geometric mean and the adaptive baseline value is taken as the reward at the specified moment. The adaptive baseline value is obtained by the exponential moving average of historical rewards.

[0026] In the preferred embodiment of the above technical solution, the resource carrier is subjected to the combined force of the attraction of the target node and the repulsion of the obstacle node. The attraction increases with the increase of the remaining path distance and the queuing demand of the target node, and the repulsion is effective within the radius of action of the obstacle node. The direction of the resultant force is mapped to the driving direction, and the magnitude of the resultant force is mapped to the speed adjustment coefficient. A scheduling instruction set containing the target path sequence, suggested driving speed and expected arrival time window is generated by potential field integration.

[0027] In the preferred embodiment of the above technical solution, the initial probability of each attraction being included in the recommended path is obtained based on the tourist interest profile and a normalized exponential function. The recommendation objective function is constructed by weighting three factors: preference matching, congestion avoidance, and diversity encouragement. A greedy algorithm is then used to select the attraction that maximizes the increment of the objective function hop by hop, thereby generating an ordered sequence of recommended paths.

[0028] In the preferred embodiment of the above technical solution, each buyer region submits a demand function based on the current number of people in the queue and the patience coefficient, and each shuttle bus submits a supply function based on the remaining carrying capacity and the supply response coefficient. The auction liquidator solves for the market clearing price that makes the total demand equal to the total supply, and each region allocates its capacity share according to the proportion of demand at the market clearing price to the total demand.

[0029] In the preferred embodiment of the above technical solution, when the service supply-demand ratio is less than 1, it is determined that the corresponding node is in short supply, an acceleration command or a detour command is sent to the shuttle bus closest to the node, and alternative attractions or extended tour suggestions are pushed to tourists whose planned routes include the node; when the service supply-demand ratio is greater than 1.5, it is determined that the corresponding node is in oversupply, an evacuation command is sent to redundant shuttle buses, and the node is recommended to tourists; in S9, the entropy change is decomposed into the entropy generation inside the system and the negative entropy flow introduced by the external environment. When the absolute value of the negative entropy flow is greater than the entropy generation, the scheduling strategy is determined to be effective; otherwise, a strategy adjustment alarm is triggered.

[0030] In a preferred embodiment of the above technical solution, S10 includes: calculating a predicted tourist satisfaction value based on the expected failure rate between the tourist's expected experience rating and the actual experience rating, combined with the weighted average of real-time waiting time and path congestion cost; superimposing the average of the predicted satisfaction values ​​of all tourists in the park as a feedback signal onto the time reward; driving the parameter update of the strategy network; and periodically collecting real tourist satisfaction survey data to calibrate the satisfaction prediction model online.

[0031] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0032] Achieve precise perception and quantitative representation of the overall situation of the scenic area; deeply capture the spatial correlation characteristics of the scenic area's topological structure; realize real-time adaptive decision-making for resource scheduling in a continuous action space; balance and coordinate the conflicts of interest among multiple parties to achieve Pareto optimal resource allocation; generate smooth and executable physical scheduling instructions to ensure operational safety and stability; achieve the organic unity of personalized route recommendation and global passenger flow balance; realize efficient dynamic allocation of resources through a market-based bidding mechanism; and establish a closed-loop feedback optimization mechanism to ensure the system's continuous evolution and adaptive capabilities.

[0033] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0034] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0035] Figure 1 This is the overall flowchart of the present invention.

[0036] Figure 2 This is a flowchart of the multi-source data sensing and gravitational interaction process of the present invention.

[0037] Figure 3 This is a flowchart of the state encoding and policy training process of the present invention.

[0038] Figure 4 This is a flowchart of the scheduling, execution, and collaborative output of this invention.

[0039] Figure 5 This is a flowchart of the effect evaluation and closed-loop feedback of the present invention.

[0040] Figure 6 This is a graph showing the experimental data comparing passenger flow density and congestion entropy before and after the scheduling of this invention. Detailed Implementation

[0041] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0042] like Figures 1 to 6 As shown, this invention discloses a method for dynamic resource scheduling and personalized path recommendation in scenic areas based on deep reinforcement learning, comprising the following steps:

[0043] S1 collects multi-source data and aligns it in time and space, calculates the intensity of spatial gravitational interaction between scenic spots, and generates a gravitational interaction thermodynamic matrix.

[0044] Establish a data foundation for comprehensive situational awareness of the scenic area: acquire micro-environmental parameters and gate access records through ground sensor networks, and acquire tourist location trajectories and dwell time data through mobile communication base stations; align the above data sources in time and space, map them uniformly to the scenic area's geographic coordinate system, and construct a time-series data cube.

[0045] Based on this, the spatial gravitational interaction strength between any two scenic spots is calculated. Let the Euclidean distance between scenic spot i and scenic spot j be . The real-time visitor density of attraction i at time t is The real-time visitor density of attraction j at time t is The landscape similarity between scenic spot i and scenic spot j is Then the spatial gravitational interaction strength between scenic spot i and scenic spot j Calculate according to the following formula:

[0046] ;

[0047] Where G is the gravitational constant, with a value of 100, which serves to maintain dimensional consistency and scale numerical values. and These are the real-time visitor density for the two attractions, expressed in person per square meter. This reflects the synergistic amplification effect of the visitor flow scale of the two attractions; the greater the visitor flow density of an attraction, the stronger its attraction to other attractions. This is the distance decay term, reflecting the weakening effect of spatial distance on gravity. The greater the distance, the lower the tourists' willingness to move between the two attractions. This is the distance attenuation coefficient, with a value of 2. The congestion loss factor is set to 0.05. Landscape similarity is calculated by the cosine similarity of the landscape feature vectors of two attractions. The landscape feature vector is a vectorized representation of the inherent attribute features of the attractions. The similarity value ranges from [0,1], reflecting the impact of attraction type matching on visitor flow migration. Attractions with high similarity have a substitution effect, while attractions with low similarity have a complementary effect. The congestion reduction factor reflects the inhibitory effect of the congestion level of the connecting path on the gravitational force. The more congested the path, the weaker the effective gravitational force between attractions. The current average congestion index of the path connecting scenic spot i and scenic spot j is calculated from the real-time traffic status data of the path. The value range is [0,1], and the larger the value, the higher the degree of congestion.

[0048] The values ​​of the above coefficients are based on the following: the gravitational constant G is set to 100 because the passenger flow density is on the order of approximately 0.1 to 2.0 people per square meter, while the magnitude of the distance squared term is approximately... to The difference between the two is 6 to 7 orders of magnitude, and taking G as 100 will make the calculated result of the gravitational interaction strength fall within the range of square meters. to The magnitude of the scale facilitates numerical comparison and visualization of the gravitational interaction thermal matrix, and G, as a global scaling factor, does not change the relative order of gravitational strength between pairs of attractions. The distance attenuation coefficient α is set to 2, which is based on the classic conclusion in geographical gravitational models that the interaction strength decreases with the square of the distance. It can characterize the characteristic that tourists' willingness to move between two attractions decreases rapidly with distance. In practical applications, it can be fitted and calibrated in the range of 1.5 to 2.5 based on historical passenger flow migration data. The congestion loss coefficient β is set to 0.05, because the congestion index is a normalized value in the range of [0,1]. When β is set to 0.05, the congestion loss factor changes gently in the range of 0.951 to 1, so that the congestion factor plays a mild regulatory role on gravity rather than a dominant inhibitory role, avoiding excessive loss that would lead to the annihilation of gravitational information. In practical applications, it can be calibrated in the range of 0.01 to 0.1 based on historical congestion statistics.

[0049] By summing up the gravitational interaction strength between all pairs of attractions, a spatial gravitational interaction thermodynamic matrix of the scenic area is generated. Each row of the matrix represents the gravitational distribution exerted by the corresponding attraction on other attractions, and each column represents the gravitational convergence strength of the corresponding attraction on other attractions.

[0050] This formula transfers the physical gravity model to the field of human geography, describing the interaction of visitor flow attraction between any two attractions in a scenic area, and realizing a quantitative representation of the interaction of visitor flow between attractions.

[0051] S2. Construct a heterogeneous topology graph of the scenic area, and use a graph convolutional network to propagate and aggregate node features to generate a system state vector.

[0052] The physical space and logical relationships of the scenic area are abstracted into a structured representation that can be processed by deep networks. Each functional area within the scenic area is abstracted as a graph node, with node types including: scenic spot nodes, transportation hub nodes, service facility nodes, and entrance / exit nodes. Roads or paths connecting the various areas are abstracted as graph edges, with edge attributes including: physical length, current traffic speed, theoretical traffic capacity, and real-time congestion level.

[0053] Constructing a heterogeneous topology map of the scenic area Let V be the set of nodes, E be the set of edges, X be the node feature matrix, and A be the adjacency matrix. Node features It includes the following dimensions: real-time passenger flow density, average dwell time, facility service capacity, current queue length, historical passenger flow benchmark, environmental comfort index, and gravitational convergence intensity.

[0054] A two-layer graph convolutional network is used to encode the state of the scenic area's topology. The propagation rule of the l-th layer graph convolution is:

[0055] ;

[0056] Where l takes the value 0 or 1, corresponding to two layers of graph convolution; The adjacency matrix is ​​a symmetric normalized adjacency matrix. Its function is to take the weighted average of the characteristics of the neighboring nodes according to the square root of the node degree, so as to avoid the excessive dominance of high-degree nodes on low-degree nodes and ensure the fairness of information transmission. The adjacency matrix for adding self-loops, where I is the identity matrix; for The degree matrix; Let X be the node representation matrix of the l-th layer, with the initial value being the original node features X, i.e., the initial input. ; Let be the learnable parameter matrix of the l-th layer, to realize the linear transformation and dimension mapping of features; The ReLU activation function is used to introduce non-linear expressive power.

[0057] After two layers of graph convolution, the representation of each node is subjected to global average pooling and concatenated with global statistical features. These global statistical features are scenic area-level statistics obtained by statistically summarizing the features of all nodes. The total dimension of the concatenated features is 128, resulting in the scenic area system state vector. Its dimension is 128, which can fully accommodate the information capacity of the concatenated global pooling features and global statistical features, and matches the input dimension of the policy network. If the dimension is too low, state information will be lost, and if it is too high, the training burden on the network will be increased. This state vector encodes global information about the scenic area's topology, visitor flow distribution, and facility status.

[0058] The reason for using two layers in the graph convolutional network is as follows: a single-layer graph convolution can only aggregate information from directly adjacent nodes, resulting in a limited receptive field; three or more layers of graph convolution are prone to excessive smoothing due to overlapping neighborhoods, leading to convergence of node representations, and increasing the number of parameters and training costs; a two-layer structure can cover the second-order neighborhood of nodes, matching the average topological interval between functional areas within the scenic area, and taking into account both the ability to capture spatial correlation features and the distinguishability of node representations.

[0059] The formula obtains the new representation of each node by weighted aggregation of the representations of all its neighboring nodes, followed by linear transformation and nonlinear activation. This achieves smooth propagation and aggregation of node representations on the graph topology and is the core propagation rule of symmetric normalized graph convolution.

[0060] S3 constructs a policy network and a value network with deep deterministic policy gradients to learn the mapping from system state to scheduling actions.

[0061] Establish a direct mapping network from system state to resource scheduling actions, and construct a deep deterministic policy gradient network with an Actor-Critic architecture, including a policy network. With value network ,in These are the learnable parameters of the policy network. For the learnable parameters of the value network, Let be the resource scheduling action vector at time t.

[0062] The policy network adopts a four-layer fully connected structure with an input dimension of 128, hidden layer dimensions of 256, 256, and 128 respectively, and an output dimension representing the resource scheduling action space. The output layer of the policy network uses the Tanh activation function to normalize action values ​​to the [-1, 1] interval, and then maps them to the actual physical value range of each action dimension through a linear transformation. This actual physical value range is determined by the operational physical boundaries of the corresponding scheduling parameters.

[0063] The value network also employs a four-layer fully connected structure. Its input is a concatenation of the state and action vectors, and its output is a scalar value estimate. The value network's role is to evaluate the expected cumulative reward of taking a specific action in a given state, providing a baseline signal for the gradient updates of the policy network.

[0064] During training, a target network mechanism and an experience replay mechanism are introduced. The target network parameters are slowly copied from the online network using a soft update method.

[0065] ;

[0066] in, For target network parameters; For online network parameters; τ is the soft update coefficient, with a value of 0.001. The reason for this value is that τ being much less than 1 allows the target network parameters to slowly track the online network, ensuring the stability of the target value calculation; if the value of τ is too large, it will cause fluctuations in the target network and trigger training oscillations; if the value of τ is too small, the target network update will lag and slow down convergence; 0.001 is a classic value verified by a large number of experiments in the deep deterministic policy gradient algorithm, and it can be adjusted in the range of 0.0001 to 0.01 in practical applications.

[0067] Experience replay pool stores state transition samples During training, small batches of data are randomly sampled for network updates to break data correlation and improve training stability. The moment reward calculated for S4.

[0068] The gradient update formula for the policy network is:

[0069] ;

[0070] in, is the performance objective function of the policy network, which is the expected cumulative reward corresponding to the output action of the policy network; N is the number of samples in the mini-batch, and the subscript i is the index of the batch sample; Indicates the state The gradient of actions in the value network, i.e., the direction in which an action changes to bring about an increase in value, is equivalent to the guidance signal that the value network gives to the policy network. This represents the gradient of the policy network output with respect to the network parameters, i.e., how changes in the parameters affect the policy output; This means that the value gradient is passed to the policy parameters through the chain rule; exist The value is taken at that location.

[0071] Guided by the value network, the policy network adjusts its parameters to make its output actions tend towards higher value directions. Since the policy is deterministic, it avoids the high variance problem of integral sampling in stochastic policies and achieves efficient and stable policy improvement. This formula is a concrete implementation of the deterministic policy gradient theorem, describing the update direction of the policy network parameters.

[0072] The gradient update of the value network uses mean squared time-series difference error:

[0073] ;

[0074] in, , For the target value network, For the target policy network, , These are the parameters for the target policy network and the target value network, respectively. The reward for each moment in the experience replay sample. This is a discount factor, with a value of 0.99. The basis for this value is that the allocation of scenic area resources needs to take into account the long-term cumulative performance of the entire day's operation. Approaching 1 makes the policy network prioritize long-term returns and avoid short-sighted decision-making; Values ​​that are too small can cause the strategy to focus excessively on immediate rewards. Taking too large a value will increase the variance of the value estimate. In practical applications, it can be adjusted within the range of 0.9 to 0.999.

[0075] S4 calculates the moment reward for multi-agent collaboration based on Nash bargaining game to balance the conflicting interests of multiple parties.

[0076] This approach coordinates multiple conflicting optimization objectives to generate a comprehensive reward signal that guides the learning direction of the policy network. It identifies four core stakeholders in scenic area resource allocation: operators focus on minimizing operating costs, tourists focus on minimizing average waiting time, scenic area management focuses on maximizing service stability, and equipment management focuses on balancing facility utilization.

[0077] Map the above four objectives to a standardized profit function , , , After normalizing the corresponding cost-related objectives, each payoff function takes values ​​in the range [0,1], where 1 indicates the objective is optimal and 0 indicates the objective is worst. Let the retention utility of each party in the event of a breakdown in negotiations be... , , , Its value is set according to the minimum revenue level that each entity can independently guarantee when there is no coordinated scheduling. Then the Nash bargaining solution is obtained by solving the following optimization problem:

[0078] This formula is the optimization objective of the Nash bargaining solution, describing the optimal coordination scheme under multiple conflicting interests.

[0079] ;

[0080] in, The profit value of subject k, The retention utility of subject k is the minimum guarantee that subject can independently obtain when negotiations break down; ∏ represents the excess profit obtained by subject k through cooperation; ∏ is the multiplication symbol. The bargaining weight of subject k satisfies This reflects the player's power in the game and allows for differentiated treatment of the importance of different players; players with greater weight receive stronger protection in the optimization process. This product form inherently prevents the extreme compression of any player's payoff, because if one player's excess payoff approaches zero, the entire product approaches zero, regardless of how high the payoffs of other players are. This characteristic ensures the fairness of the Pareto optimal solution and avoids the extreme unfair distribution that may occur in utilitarian summation.

[0081] The bargaining weights are determined based on the following criteria: the bargaining weights of each entity reflect their relative priority in the scenic area's operational goals and can be determined according to the scenic area's operational strategy. For example, an operational strategy that prioritizes visitor experience can increase the bargaining weights of the visitor group. The specific values ​​can be determined through the analytic hierarchy process or regression analysis based on historical operational data. The retention utility is determined based on the following criteria: it represents the minimum guaranteed revenue that each entity can independently obtain when negotiations break down. It can be determined based on the historical worst operational statistics of each entity's corresponding goals. For example, the standardized revenue value under the highest operating cost scenario in historical operations can be used as the operator's retention utility.

[0082] Final Moment Reward Defined as the difference between the Nash bargaining total return and the baseline return:

[0083] ;

[0084] in, Let be the standardized return value of subject k at time t, as mentioned above. Same meaning; The adaptive baseline value is calculated by performing an exponential moving average on historical rewards, which is used to reduce the variance of the reward signal and stabilize the learning process.

[0085] The rationale for taking the fourth root of the product term is that the geometric mean can eliminate the scaling effect that causes the product result to shrink exponentially as the number of subjects increases, thus keeping the value of the comprehensive Nash bargaining return within the dimension range of [0,1], which facilitates direct comparison with the adaptive baseline value.

[0086] S5 converts scheduling actions into executable scheduling instructions that include path planning and speed control based on an artificial potential field.

[0087] The abstract action vectors output by the policy network are converted into executable scheduling instructions for resource carriers in the physical world; the scenic resource carriers are regarded as intelligent agents moving in the potential field space, and each resource carrier is subjected to the combined effects of attractive forces from target nodes and repulsive forces from obstacle nodes.

[0088] For the m-th shuttle bus, the attractive potential field exerted on its target node j Calculate using the following formula:

[0089] ;

[0090] in, The attraction gain coefficient controls the overall strength of the potential field, and its value is 0.5. The basis for this value is: By controlling the slope of the attraction potential field as distance increases, when the slope is set to 0.5, the potential field strength generated at typical road network distances in the range of hundreds of meters to kilometers is at the same level as the passenger flow demand. An excessively large value will cause the attraction to distant targets to be too strong, triggering a long-distance descent for the shuttle bus. If the value is too small, the near-end response will be slow. In practical applications, it can be adjusted within the range of 0.1 to 1.0 according to the scale of the scenic area's road network.

[0091] Let m be the remaining path distance from the current position of the shuttle bus m to the target node j. The quadratic term reflects the elastic recovery characteristic of Hooke's Law. The farther the distance, the greater the gravitational potential energy, which drives the resource carrier to move closer to the target to release potential energy.

[0092] The weighting coefficient for passenger flow demand is 0.1, which balances the relative importance of distance factors and demand factors. The value is based on the fact that the standardized value of the number of passengers in queue is in the range of 0 to 1. Taking 0.1 makes the adjustment range of passenger flow demand on attractiveness control within the order of 10%. The demand signal is used as a correction term for the distance factor rather than the dominant term, so as to avoid demand fluctuations from overwhelming the distance gradient. In practical applications, it can be adjusted within the range of 0.05 to 0.5.

[0093] The standardized value of the number of passengers queuing at target node j at time t is defined as the ratio of the number of people queuing to the upper limit of the queuing capacity of the target node. This term embeds the external demand signal into the potential field, making high-demand nodes more attractive.

[0094] The resource carrier is subjected to a virtual force in the potential field space. The magnitude of the force is equal to the negative gradient of the potential field, and the direction is towards the direction of the fastest decrease in potential energy, thereby causing the resource carrier to spontaneously move towards the target node that is in high demand and reachable. This formula describes the attractive potential field of the target node on the resource carrier, which is similar to the elastic potential energy model.

[0095] Repulsive potential field exerted by obstacle node k Calculate using the following formula:

[0096] ;

[0097] in, κ is the repulsive force gain coefficient, with a value of 100. The value is determined based on the following: κ needs to match the magnitude of the attractive potential field so that the repulsive potential generated by the obstacle node within the radius of action is sufficient to counteract the attractive potential. If the value of κ is too small, obstacle avoidance will fail. If the value of κ is too large, the shuttle bus will experience directional shaking near the obstacle. In practical applications, it can be adjusted within the range of 10 to 1000.

[0098] Let m be the distance from the shuttle bus to the obstacle node k. The radius of the repulsive force is set at 500 meters. The value is based on the following: its radius of action should not be less than the emergency braking distance at the maximum speed of the shuttle bus and a safety margin should be reserved. At the same time, it should cover the typical oncoming sight distance of the scenic road. In actual application, it can be adjusted within the range of 300 to 800 meters according to the vehicle type and road conditions.

[0099] The direction of the resultant force acting on the shuttle bus determines its direction of travel, and the magnitude of the resultant force is mapped to a speed adjustment coefficient. The action vector output by the policy network... Each dimension corresponds to a scheduling parameter such as the attraction weight adjustment amount and the global speed adjustment amount for different target nodes, and each dimension corresponds one-to-one with the dimensional division of the action space described in S3. Through potential field integration, the target path sequence, suggested driving speed, and estimated arrival time window for each shuttle bus are generated, forming a complete physical scheduling instruction set.

[0100] S6 measures path diversity based on information entropy and combines tourist interest profiles to generate recommended paths that balance customer flow.

[0101] To generate a tour route for each visitor that balances personalized preferences with overall visitor flow: First, based on visitors' historical behavior data and real-time interest tags, a visitor interest profile vector is constructed. Its dimensions are the same as the number of scenic spots, and each component... This represents the interest rating of tourist u towards attraction i.

[0102] For tourist u, the initial probability that attraction i is included in the recommended path. Calculated using the Softmax function:

[0103] ;

[0104] Wherein, T is a temperature parameter that controls the sharpness of the probability distribution, with a value of 1.0. The basis for this value is as follows: when T is 1.0, the probability distribution output by the normalized exponential function retains the original relative differences in interest ratings; when T approaches 0, the probability distribution degenerates into selecting only the attractions with the highest interest, losing path diversity; when T approaches infinity, the probability distribution degenerates into a uniform distribution, losing personalization. In practical applications, it can be adjusted within the range of 0.5 to 2.0 to regulate the concentration of the recommendation results.

[0105] To achieve route diversity, an information entropy regularization term is calculated. Let the set of attractions covered by the recommended routes be denoted as . Then the path diversity entropy Defined as:

[0106] ;

[0107] In the formula In the set First, normalize the content; the larger the diversity entropy value, the more evenly the distribution of attraction types in the recommended path is, which is more conducive to guiding tourists to non-hotspot areas.

[0108] Taking into account both path preference matching degree and diversity entropy, the objective function for path recommendation is:

[0109] ;

[0110] in, The preference matching weight is set to 1.0. To mitigate congestion, a weight of 0.5 is used. To encourage diversity, a weight of 0.3 is assigned. Let be the real-time visitor density of attraction i at time t. The first term of the objective function encourages recommendations of attractions that tourists are interested in, the second term penalizes recommendations of highly congested attractions, and the third term rewards the diversity of routes.

[0111] The weighting process involves setting the preference matching weight to 1.0 as a normalization benchmark to ensure that tourists' personalized needs are prioritized; the congestion avoidance weight to 0.5, half the preference matching weight, to ensure that congestion penalties play a diversion role without overriding personalized preferences; and the diversity encouragement weight to 0.3, a relatively small value, to make diversity an auxiliary optimization objective and avoid recommending attractions that tourists are not interested in in pursuit of diversity. The specific values ​​of the three weights can be calibrated through a grid search of tourist satisfaction feedback data, maintaining a relative relationship where the preference matching weight is greater than the congestion avoidance weight, and the congestion avoidance weight is greater than the diversity encouragement weight.

[0112] A greedy algorithm is used to solve the above optimization problem: starting from the entry node, at each step, the next hop attraction that maximizes the increment of the objective function is selected until the tourist's time budget limit is reached or all feasible attractions have been traversed, and finally an ordered sequence of recommended paths is generated.

[0113] The objective function for path recommendation is a comprehensive optimization objective for personalized path recommendation, which includes three mutually constraining terms: the first term... For preference matching, the recommendation system is encouraged to select attractions that tourists are interested in, with weights... Ensure that personalized needs are met first; the second item As a congestion avoidance measure, penalties are imposed on recommendations of highly congested tourist attractions, with weighting... Control the intensity of congestion avoidance; the third item As a diversity incentive, the richness of attraction types along the recommended route is measured by information entropy, with weights assigned accordingly. Balancing personalization and diversity. This formula enables the recommendation system to find the optimal balance among three dimensions: visitor interests, system congestion, and path diversity. It fully responds to individual needs, avoids exacerbating congestion in popular areas, and ensures a rich visitor experience.

[0114] S7 uses a dual auction mechanism to solve for the market clearing price and allocates transportation capacity proportionally to achieve resource competition allocation.

[0115] The efficient allocation of limited resources is achieved through a market-based bidding mechanism. The scheduling of scenic area transportation resources is modeled as a continuous dual auction process, with the buyers being the passenger queues in each area, the sellers being the shuttle buses, and the auctioned item being the carrying capacity of the shuttle buses for the next trip.

[0116] The demand function submitted by each buyer region i is:

[0117] ;

[0118] in, Let be the current number of people queuing in region i. Let be the patience coefficient for region i, reflecting passengers' sensitivity to waiting time. Its value is determined by the following: the patience coefficient determines the rate at which buyer demand decreases as the virtual ticket price increases. Its value should match the dimensions of the virtual ticket price, ensuring that when the virtual ticket price changes within the market clearing range, buyer demand decreases to 60% to 90% of the initial value, consistent with observed patterns of tourist queuing behavior. The patience coefficient can be calibrated through regression analysis of historical queuing numbers and waiting times in each region, with an optimal value range of 0.01 to 0.05.

[0119] p represents the virtual ticket price, measured in standardized price units; this function uses an exponential decay form to reflect the economic law that demand decreases as prices rise, with a patience coefficient... This determines the steepness of the demand curve.

[0120] The supply function submitted by each shuttle bus m is:

[0121] ;

[0122] in, The remaining carrying capacity of shuttle bus m The supply response coefficient for shuttle bus m is determined by the following: the supply response coefficient determines the saturation rate of seller supply as the virtual ticket price increases. Its value should match the magnitude of the patience coefficient, ensuring that the demand and supply curves have a unique and stable intersection point within the positive price range, guaranteeing that the market-clearing price is solvable and numerically stable. The supply response coefficient can be calibrated based on historical shuttle bus dispatch response data, with a preferred value range of 0.01 to 0.1.

[0123] This function adopts a saturated growth form to reflect the economic law that the quantity supplied by sellers increases when prices rise, and the supply response coefficient... It determines the steepness of the supply curve.

[0124] Auction liquidators seek market clearing prices This makes aggregate demand equal to aggregate supply:

[0125] ;

[0126] This equation represents the market clearing condition: In a competitive market, the price mechanism spontaneously adjusts the supply and demand relationship. When aggregate demand equals aggregate supply, the market reaches equilibrium, and the price at this point is called the clearing price. Solving the above equation numerically yields the market clearing price. .

[0127] The capacity allocated to each region is determined proportionally based on its demand at that price:

[0128] ;

[0129] This allocation result serves as a reference for shuttle bus route planning, guiding vehicles to prioritize areas with the highest demand prices, thus achieving efficient market-based allocation of transportation resources. Given a limited total supply, each region is allocated resources according to its proportion of demand at the equilibrium price to total demand; regions with higher demand receive a larger share of capacity. This demand-based allocation method ensures both efficiency and fairness in resource allocation.

[0130] S8 calculates the service supply-demand ratio for each scenic spot and performs spatiotemporal alignment and collaborative output of scheduling instructions and recommended routes.

[0131] Resource scheduling instructions and path recommendation results are spatiotemporally aligned and output collaboratively to ensure the compatibility and consistency of the two types of decisions.

[0132] A unified spatiotemporal coordinate system is established, matching the estimated arrival time series of shuttle buses in the scheduling instructions with the estimated arrival time series of tourists in the route recommendations. For each scenic spot node, its service supply-demand ratio is calculated:

[0133] ;

[0134] in, Let be the service supply-demand ratio of node i at time t; Let m be the transport capacity that the shuttle bus can deliver to node i at time t. The demand generated by tourist u arriving at node i at time t; molecule Let be the total transport capacity of all shuttle buses that can reach node i at time t, representing the supply-side capacity; denominator: Let be the total demand generated by all tourists arriving at node i at time t, representing the pressure on the demand side.

[0135] when When node i is in a state of supply shortage, the collaborative adjustment mechanism is triggered: on the one hand, an acceleration command or a detour command is sent to the shuttle bus closest to node i; on the other hand, alternative attractions or extended tour suggestions are pushed to tourists who are planning routes that include node i.

[0136] when When node i is in a state of oversupply, a reverse adjustment mechanism is triggered: an evacuation command is sent to redundant shuttle buses to guide them to other demand nodes; at the same time, the node is appropriately recommended to tourists to improve resource utilization. Here, 1.5 is a preset redundancy threshold that takes into account capacity adjustment margin.

[0137] The basis for setting the redundancy threshold to 1.5 is as follows: setting a 50% margin based on the supply and demand balance point 1 can avoid frequent triggering of the reverse adjustment mechanism when the service supply and demand ratio fluctuates near the balance point, forming an adjustment lag range and preventing scheduling command jitter; this threshold can be adjusted in the range of 1.2 to 2.0 according to the time cost of shuttle vehicle capacity allocation. The higher the allocation time cost, the larger the threshold should be.

[0138] Through the aforementioned two-way collaborative mechanism, scheduling instructions and recommended routes are dynamically coordinated under a unified optimization framework, avoiding conflicts and internal friction between the two types of decisions; the concept of supply and demand equilibrium in economics is transferred to the field of scenic area operation, and the two-way collaboration between resource scheduling and route recommendation is achieved by quantifying the supply and demand gap of each node; the service supply and demand ratio of scenic spot nodes is defined to trigger the core judgment indicator of the collaborative adjustment mechanism.

[0139] S9 calculates the congestion entropy and entropy change of scenic areas based on dissipative structure theory, and evaluates the orderliness effect of scheduling strategies.

[0140] Assessing the orderliness of scenic area operation from a systems science perspective provides macroscopic state feedback for closed-loop optimization. The scenic area is viewed as an open, dissipative structural system that maintains its orderly operation through continuous exchange of matter, energy, and information with the external environment.

[0141] Define scenic area congestion entropy A measure of the degree to which passenger flow density in each area deviates from equilibrium:

[0142] ;

[0143] in, Let be the average visitor density of the scenic area at time t. The congestion entropy measures the uniformity of visitor flow distribution across different areas of the scenic area, representing a generalization of relative entropy in information theory to geographic space; when the visitor density is equal in all areas, each item... When the value is zero, the congestion entropy takes the minimum value of 0. When passenger flow is highly concentrated in a few areas, the positive term in high-density areas and the negative term in low-density areas are both large, and the congestion entropy increases significantly.

[0144] Calculate the rate of change of congestion entropy over time, i.e., entropy change:

[0145] ;

[0146] Entropy change is decomposed into two parts: entropy generation and negative entropy flow.

[0147] ;

[0148] in, The entropy generated within the system is always non-negative, representing the irreversible trend of the system spontaneously moving towards disorder. This stems from the fact that tourists' autonomous movement behavior always tends towards a few popular attractions. Here, the subscript "in" indicates the internal system. The negative entropy flow introduced into the system from the external environment represents the external order input imposed by the scheduling system. Intelligent scheduling guides the even distribution of passenger flow. Here, the subscript *ex* represents the external environment. The entropy change decomposition treats the scenic area as an open dissipative structure system, whose degree of order evolution is jointly determined by internal entropy generation and external negative entropy flow. When the absolute value of the negative entropy flow is greater than the internal entropy generation, the total entropy change of the system is less than zero, and the scenic area tends to operate in a more orderly state. This is precisely the core objective pursued by the scheduling system of this invention.

[0149] when When the total entropy change of the system is less than zero, the scenic area tends to operate in a more orderly state, indicating that the current scheduling strategy is effective; conversely, a strategy adjustment alarm needs to be triggered, and the emergency response mechanism needs to be activated. Since entropy generation is always non-negative, it can be judged by the sign of the total entropy change: a total entropy change less than zero indicates that the absolute value of the negative entropy flow is greater than the entropy generation, and the current scheduling strategy is effective.

[0150] S10 predicts tourist satisfaction based on the expectation failure theory and feeds it back to the policy network to achieve closed-loop optimization.

[0151] A predictive model for tourist satisfaction is established, and the satisfaction feedback signal is incorporated into a reinforcement learning framework to form a closed-loop optimization. According to the expectation-failure theory, tourist satisfaction depends on the difference between the actual experience and prior expectations.

[0152] Let tourist u's expected experience rating for attraction i be . The actual experience rating is The expected failure rate is:

[0153] ;

[0154] The expected failure rate reflects the difference between the actual experience and the previous expectation; a positive value indicates that the experience exceeds the expectation, and a negative value indicates that the experience falls short of the expectation.

[0155] Overall satisfaction forecast of tourist u Calculated using the following formula:

[0156] ;

[0157] in, The rating represents tourists' prior expectations of the attraction, reflecting their level of expectation based on word-of-mouth and advertising. The real-time waiting time for attraction i, in hours, reflects the impact of service efficiency on satisfaction. Let be the route congestion cost to reach attraction i, and let be a normalized dimensionless value that reflects the impact of the travel experience on satisfaction. A collection of attractions covered by the recommended route The number of attractions in the area; , , , The weighting coefficients for each factor are 0.3, 0.4, 0.2, and 0.1, respectively, based on the following criteria: The weighted averages are normalized. The weight corresponding to the degree of failure is set to a maximum of 0.4, reflecting the core conclusion of the expected failure theory that the relative difference between actual experience and expectation is the dominant factor in satisfaction. The weight corresponding to the expected score is set to a second maximum of 0.3, reflecting the anchoring effect of previous expectations on satisfaction. The weight corresponding to waiting time is set to 0.2, and the weight corresponding to route congestion cost is set to 0.1, serving as auxiliary correction terms for service efficiency and travel experience. Each weight can be calibrated through multiple regression analysis of tourist satisfaction questionnaire data. After calibration, the relative size relationship should be maintained as follows: the weight of the degree of failure is the largest, followed by the weight of the expected score, then the weight of the waiting time, and finally the weight of the route congestion cost.

[0158] A tourist satisfaction prediction model is constructed based on the expectation-disapproval theory. Tourist satisfaction is jointly determined by four dimensions: expectation level, degree of disappointment, waiting time, and travel cost. The core insight of expectation-disapproval theory is that satisfaction depends not only on the quality of the objective experience but also on the relative difference between the experience and expectations. Even if the objective experience is good, satisfaction may still be low if expectations are not met; conversely, even if the objective experience is average, satisfaction may still be high if expectations are exceeded. This model engineers this psychological theory, providing quantifiable feedback signals for the closed-loop optimization of the scheduling system.

[0159] The predicted satisfaction value is fed back as an additional reward signal to the deep deterministic policy gradient network of S3, forming an extended reward function:

[0160] ;

[0161] in, The satisfaction feedback weight is set at 0.5, and the basis for this value is as follows: By superimposing the satisfaction feedback item and the time reward on the same order of magnitude, we can avoid the satisfaction signal drowning out the learning signal of the main reward, and ensure that the satisfaction feedback has a substantial impact on the policy update. Taking excessively large values ​​will lead to strategies that cater too much to short-term satisfaction at the expense of operational efficiency. If the value is too small, the closed-loop feedback will fail. In practical applications, it can be adjusted within the range of 0.1 to 1.0.

[0162] This represents the total number of visitors currently in the park. Through this closed-loop mechanism, the resource allocation strategy not only pursues the optimization of operational efficiency but also directly responds to visitor satisfaction feedback, achieving a dual improvement in system efficiency and user experience.

[0163] Regularly collect real tourist satisfaction survey data, calibrate the satisfaction prediction model online, correct systematic biases between expected and actual scores, and ensure the long-term accuracy of the prediction model.

[0164] To make the technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0165] Example: Small to medium-sized natural scenic area. Taking a small to medium-sized natural scenic area as the application object, the scenic area has a total area of ​​approximately 15 square kilometers, includes 5 functional sub-areas, is equipped with 3 electric shuttle buses, and receives an average of about 5,000 visitors per day; the 5 sub-areas are: entrance service area, lakeside viewing area, mountain hiking area, waterfall sightseeing area, and cultural exhibition area; the areas are connected by a ring road with a total length of approximately 8 kilometers.

[0166] S1 collects multi-source data and aligns it in time and space, calculates the intensity of spatial gravitational interaction between scenic spots, and generates a gravitational interaction thermodynamic matrix.

[0167] By acquiring uplink and downlink communication data from mobile communication data receiving terminals within the scenic area, and ground sensor networks distributed at the entrances and exits of five sub-areas (i.e., the number of passengers passing through the turnstiles), the real-time passenger flow density, queuing length, and landscape feature vectors for the five areas are obtained as follows:

[0168] Entrance service area: per person per square meter people;

[0169] Lakeside Viewing Area: per person per square meter people;

[0170] Mountain hiking area: per person per square meter people;

[0171] Waterfall Scenic Area: per person per square meter people;

[0172] Cultural Exhibition Area: per person per square meter people;

[0173] The above to This represents the original number of people queuing in each area.

[0174] The Euclidean distance matrix between each region is as follows: rice, rice, rice, rice, rice, rice, rice, rice, rice, rice.

[0175] The cosine similarity of landscape feature vectors is: , , , , , , , , , .

[0176] The congestion index of the connecting path is: , , , , , , , , , .

[0177] The spatial gravitational interaction strength between the entrance service area and the lakeside viewing area was calculated using the formula for gravitational interaction strength. :

[0178] ;

[0179] Similarly, the gravitational interaction strength between all pairs of regions is calculated, generating a 5×5 spatial gravitational interaction thermodynamic matrix. The diagonal elements of this matrix are zero, and the off-diagonal elements are calculated as above. The matrix shows that the gravitational interaction between the waterfall viewing area and the lakeside viewing area is the strongest because they have the highest landscape similarity and both have high visitor density; the gravitational interaction between the cultural exhibition area and the mountain hiking area is the weakest because they are far apart and have low landscape similarity.

[0180] S2. Construct a heterogeneous topology graph of the scenic area, and use a graph convolutional network to propagate and aggregate node features to generate a system state vector.

[0181] The five sub-regions are abstracted into five graph nodes, with node characteristics... It includes 7 dimensions. Taking the entrance service area as an example, its node characteristics are: Where 0.8 represents real-time passenger flow density, 45 represents average dwell time in minutes, 500 represents facility service capacity, 120 represents current queue length, 0.75 represents the historical passenger flow benchmark for the same period, and 0.85 represents the environmental comfort index. This represents the sum of the gravitational forces acting on that node.

[0182] Construct a 5×5 adjacency matrix A, where This indicates that region i and region j are directly adjacent; otherwise, it is 0. In this embodiment, the scenic area roads are connected in a loop, therefore... , , , , The remaining off-diagonal elements are zero.

[0183] After adding a self-loop, Calculate the degree matrix. This is a diagonal matrix, where each diagonal element is the degree of the corresponding node plus one. Then, the normalized Laplacian matrix is ​​calculated. .

[0184] Assuming an initial node feature matrix The parameter matrix of the first layer graph convolution is a 5×7 matrix. It is a 7×64 matrix. After matrix multiplication and activation function operations, we obtain... It is a 5×64 matrix. The second layer of graph convolution maps the 64 dimensions to 32 dimensions, resulting in... It is a 5×32 matrix.

[0185] right Global average pooling is performed to average the 32-dimensional representations of the five nodes into a 1×32 global feature vector. This vector is then concatenated with the 96-dimensional global statistical features to obtain a 128-dimensional system state vector. .

[0186] S3 constructs a policy network and a value network with deep deterministic policy gradients to learn the mapping from system state to scheduling actions.

[0187] Building a policy network The input dimension is 128, with 256 neurons in the first hidden layer, 256 neurons in the second hidden layer, and 128 neurons in the third hidden layer. The output layer is the resource scheduling action space with a dimension of 15. The 15 output dimensions correspond to the five scheduling parameters for each of the three shuttle buses: target area number adjustment, speed adjustment, and stop duration adjustment. The target area number adjustment corresponds to the adjustment of the attraction weight in S5, the speed adjustment corresponds to the generation of the speed adjustment coefficient, and the stop duration adjustment is used to generate the expected arrival time window.

[0188] Building a value network The input is a concatenation of a 128-dimensional state vector and a 15-dimensional action vector, totaling 143 dimensions. After passing through three hidden layers, the output is a scalar value estimate.

[0189] During training, set the soft update coefficient. Discount factor The experience replay pool has a capacity of 100,000. State transition samples are collected from the environment at each step. Store the data in the replay pool, and randomly sample 64 records from it for each training session to update the network.

[0190] Example of gradient calculation for policy network: Assume a batch contains one record, state The corresponding strategy output is The partial derivatives of the value network for each action in this state are: Then the policy gradient is the direction of the dot product of the two.

[0191] S4 calculates the moment reward for multi-agent collaboration based on Nash bargaining game to balance the conflicting interests of multiple parties.

[0192] Suppose that at a certain moment, the normalized returns of the four objectives are as follows: operating costs Average waiting time Service stability Facility utilization rate The retention utility of each party is... , , , Bargaining weight is , , , .

[0193] The overall bargaining benefit of Nash is:

[0194] ;

[0195] Taking the fourth root yields a geometric mean of 0.8059. Let the adaptive baseline value be... Then reward at any time This positive reward indicates that the current scheduling strategy is performing better than the historical average, and the policy network will update its parameters in this direction.

[0196] S5 converts scheduling actions into executable scheduling instructions that include path planning and speed control based on an artificial potential field.

[0197] Assume that shuttle bus No. 1 is currently located on the road between the entrance service area and the lakeside viewing area, and is a short distance from the lakeside viewing area. There are still 800 meters left, that is , Meters. Current target node. The maximum queue capacity is 200 people. After standardization, it is 0.4.

[0198] The attractive potential field of the lakeside viewing area for shuttle bus No. 1 was calculated using the formula:

[0199] ;

[0200] Assume there is a temporary congestion point k on the road ahead, 200 meters from the shuttle bus. What is the radius of the congestion? Meters. The repulsive potential field exerted at this congestion point is:

[0201] ;

[0202] The attractive force, derived from the negative gradient of the potential field, is much greater than the repulsive force. The resultant force points towards the lakeside viewing area, and shuttle bus No. 1 will continue to travel towards the lakeside viewing area. The speed adjustment amount output by the strategy network is 0.3, and after mapping, the recommended driving speed is 15 kilometers per hour, where the upper limit of the adjustment amount corresponds to the maximum operating speed limit of 50 kilometers per hour.

[0203] S6 measures path diversity based on information entropy and combines tourist interest profiles to generate recommended paths that balance customer flow.

[0204] Suppose that the interest profile vector of a tourist u is These correspond to the interest levels of the five regions. Temperature parameter Calculate the initial probability of each region being included in the path:

[0205] ;

[0206] ;

[0207] ;

[0208] ;

[0209] .

[0210] Assuming the recommended route includes three areas: the lakeside viewing area, the waterfall sightseeing area, and the cultural exhibition area, normalizing the probabilities within this set yields 0.4076, 0.3688, and 0.2237. The path diversity entropy is:

[0211] ;

[0212] In the path recommendation objective function, the preference matching term is Congestion avoidance items are Diversity encouragement items are The objective function value is obtained by summing the three weighted terms. The overall score for this path is positive, confirming the effectiveness of the recommendation.

[0213] S7 uses a dual auction mechanism to solve for the market clearing price and allocates transportation capacity proportionally to achieve resource competition allocation.

[0214] Assume the queue numbers for the lakeside viewing area, waterfall sightseeing area, and mountain hiking area are 80, 150, and 30 respectively, with patience coefficients of 0.02, 0.015, and 0.025 respectively. The remaining capacities of the three shuttle buses are 20, 15, and 25 respectively, with a supply response coefficient of 0.05 for each.

[0215] Establish demand and supply equations, and solve for the market clearing price through numerical iteration. Assuming that the result is obtained through iterative calculation .

[0216] The demand for the lakeside viewing area is as follows: ;

[0217] The needs of the waterfall tourist area are: ;

[0218] The needs of the mountain hiking area are: ;

[0219] Total demand is approximately 59.2, and total supply is... Total demand equals total supply, satisfying the market clearing condition. The proportionally allocated capacity share for each region guides shuttle buses to prioritize access to the waterfall area.

[0220] S8 calculates the service supply-demand ratio for each scenic spot and performs spatiotemporal alignment and collaborative output of scheduling instructions and recommended routes.

[0221] At 10:30 AM, the service supply-demand ratio at the lakeside viewing area was:

[0222] ;

[0223] In the formula, the molecule This represents the total transport capacity that 3 shuttle buses can provide to the lakeside scenic area; consistent with the remaining transport capacities of 20, 15, and 25 for the 3 shuttle buses in S7; denominator This is the sum of the current queuing demand of 80 people at this node and the arrival demand calculated based on the recommended probability, where 0.263 is the initial probability that the lakeside viewing area is included in the recommended path in the S6 example, and 500 is the calculation base of the total number of visitors currently in the park.

[0224] because The system determined that the lakeside viewing area was in short supply. It triggered a coordinated adjustment mechanism: sending an acceleration command to shuttle bus number 2, the closest to the lakeside viewing area, requesting it to increase its speed from the current 12 km / h to 18 km / h; simultaneously, it suggested a mountain hiking area as an alternative attraction to 10 tourists planning routes that included the lakeside viewing area.

[0225] S9 calculates the congestion entropy and entropy change of scenic areas based on dissipative structure theory, and evaluates the orderliness effect of scheduling strategies.

[0226] The passenger flow densities in the five areas were 0.8, 1.5, 0.6, 2.0, and 0.4 respectively, with an average density of... .

[0227] The congestion entropy of the scenic area is:

[0228]

[0229] Assume that after 10 minutes, the passenger flow density in each area changes to 0.9, 1.3, 0.7, 1.8, and 0.5, with the average density becoming 1.04. At this point, the congestion entropy becomes 0.504. The entropy change is approximated by the entropy difference between two adjacent moments; the entropy change value is... An entropy change of less than 0 indicates that the system tends to become more ordered under external scheduling intervention.

[0230] S10 predicts tourist satisfaction based on the expectation failure theory and feeds it back to the policy network to achieve closed-loop optimization.

[0231] Suppose a tourist's expected rating for the lakeside viewing area is 4.2, and their actual experience rating is 3.8. Then the expectation-failure rate is: The real-time waiting time is 15 minutes, which is equivalent to 0.25 hours, and the path congestion cost is 0.3.

[0232] The tourist's predicted satisfaction level is:

[0233] ;

[0234] Taking the average satisfaction level of all visitors in the park, assuming it to be 0.85, the satisfaction feedback term in the extended reward function is: This feedback signal, along with the Nash bargaining reward, drives the parameter updates of the policy network.

[0235] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for dynamic resource scheduling and personalized path recommendation in scenic areas based on deep reinforcement learning, characterized in that, Includes the following steps: S1: Collect multi-source data and align it in time and space, calculate the intensity of spatial gravitational interaction between attractions, and generate a gravitational interaction thermodynamic matrix; S2, construct a heterogeneous topology graph of the scenic area, and use a graph convolutional network to propagate and aggregate node features to generate a system state vector; S3: Construct a policy network and value network with deep deterministic policy gradients to learn the mapping from system state to scheduling actions; S4 calculates the moment reward for multi-agent collaboration based on Nash bargaining game to balance the conflicting interests of multiple parties; S5 converts scheduling actions into executable scheduling instructions that include path planning and speed control based on an artificial potential field. S6, based on information entropy to measure path diversity, combined with tourist interest profiles to generate recommended paths that balance customer flow; S7 uses a dual auction mechanism to solve for the market clearing price and allocates transportation capacity proportionally to achieve resource competitive allocation. S8 calculates the service supply-demand ratio of each scenic spot and performs spatiotemporal alignment and collaborative output of scheduling instructions and recommended routes; S9, based on dissipative structure theory, calculates the congestion entropy and entropy change of scenic areas, and evaluates the orderliness effect of scheduling strategies; S10 predicts tourist satisfaction based on the expectation failure theory and feeds it back to the policy network to achieve closed-loop optimization.

2. The method according to claim 1, characterized in that, The S1 multi-source data from air, space, and ground includes ground sensor data and mobile communication data. The data is uniformly mapped to the scenic area's geographic coordinate system to construct a time-series data cube. The spatial gravitational interaction intensity is determined by the real-time visitor density of the two scenic spots, the distance between the scenic spots, the landscape similarity, and the current average congestion index of the path. The spatial gravitational interaction intensity of all scenic spot pairs is summarized to generate the gravitational interaction heat matrix.

3. The method according to claim 1, characterized in that, The nodes of the heterogeneous topology graph of the scenic area in S2 include scenic spot nodes, transportation hub nodes, service facility nodes, and entrance / exit nodes. The edge attributes include physical length, current traffic speed, theoretical traffic capacity, and real-time congestion level. The node features are propagated and aggregated through a two-layer graph convolutional network, then global average pooling is performed, and concatenated with global statistical features to obtain the system state vector.

4. The method according to claim 1, characterized in that, The output layer of the policy network in S3 normalizes the action values ​​and then maps them to the actual physical value range of each action dimension through a linear transformation. During training, a target network soft update mechanism and an experience replay mechanism are introduced. The value network is updated with mean square time difference error and provides a reference signal for the gradient update of the policy network.

5. The method according to claim 1, characterized in that, In step S4, four stakeholders are identified: the operator, the tourist group, the scenic area management, and the equipment management. The difference between the standardized revenue and the retention utility of each party is multiplied by a weighted average of the bargaining power, and the difference between the geometric mean and the adaptive baseline value is taken as the reward at the specified moment. The adaptive baseline value is obtained by the exponential moving average of historical rewards.

6. The method according to claim 1, characterized in that, S5 includes: the resource carrier is subjected to the combined force of the attraction of the target node and the repulsion of the obstacle node. The attraction increases with the increase of the remaining path distance and the queuing demand of the target node. The repulsion is effective within the radius of action of the obstacle node. The direction of the resultant force is mapped to the driving direction, and the magnitude of the resultant force is mapped to the speed adjustment coefficient. A scheduling instruction set containing the target path sequence, the suggested driving speed and the expected arrival time window is generated by potential field integration.

7. The method according to claim 1, characterized in that, S6 includes: obtaining the initial probability of each attraction being included in the recommended path based on the tourist interest profile and a normalized exponential function; constructing a recommendation objective function with three weighted factors: preference matching, congestion avoidance, and diversity encouragement; using a greedy algorithm to select the attraction that maximizes the increment of the objective function hop by hop; and generating an ordered path recommendation sequence.

8. The method according to claim 1, characterized in that, The S7 includes: each buyer region submits a demand function based on the current number of people in the queue and the patience coefficient; each shuttle bus submits a supply function based on the remaining carrying capacity and the supply response coefficient; the auction liquidator solves for the market clearing price that makes the total demand equal to the total supply; and each region allocates its capacity share according to the proportion of demand at the market clearing price to the total demand.

9. The method according to claim 1, characterized in that, In step S8, when the service supply-demand ratio is less than 1, it is determined that the corresponding node is in short supply. An acceleration command or detour command is sent to the shuttle bus closest to the node, and alternative attractions or extended tour suggestions are pushed to tourists whose planned routes include the node. When the service supply-demand ratio is greater than 1.5, it is determined that the corresponding node is in oversupply. An evacuation command is sent to redundant shuttle buses, and the node is recommended to tourists. In step S9, the entropy change is decomposed into entropy generation within the system and negative entropy flow introduced by the external environment. When the absolute value of the negative entropy flow is greater than the entropy generation, the scheduling strategy is determined to be effective; otherwise, a strategy adjustment alarm is triggered.

10. The method according to claim 1, characterized in that, S10 includes: calculating the predicted tourist satisfaction value based on the expected failure rate between the tourist's expected experience rating and the actual experience rating, combined with the weighted average of real-time waiting time and path congestion cost; superimposing the average of the predicted satisfaction values ​​of all tourists in the park as a feedback signal onto the time reward; driving the parameter update of the strategy network; and periodically collecting real tourist satisfaction survey data to calibrate the satisfaction prediction model online.

Images