A multi-period dynamic site selection method and system for mobile public service facilities within urban areas

By constructing regional feature maps and using multi-agent reinforcement learning algorithms, the generalization ability and training efficiency of multi-period dynamic site selection schemes for mobile public service facilities were solved, achieving efficient and adaptive multi-period site selection decisions and improving service coverage and the comprehensive capabilities of the urban service system.

CN123117550APending Publication Date: 2026-11-17APPLIED TECH COLLEGE OF SOOCHOW UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611366965.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for multi-period dynamic site selection schemes for mobile public service facilities suffer from poor generalization ability, low training efficiency, and inability to fully utilize the structural characteristics of urban road networks, resulting in an inability to effectively cope with changes in pedestrian flow distribution in different urban areas and at different times.

Method used

We employ a method based on regional feature maps and multi-agent reinforcement learning. By constructing an encoder-decoder structure consisting of regional feature maps, graph attention networks, and graph pointer networks, and combining it with a multi-agent proximal policy optimization algorithm, we build a multi-period dynamic location model to enable agents to make adaptive location decisions within urban areas.

Benefits of technology

It significantly improves the training efficiency and convergence speed of the model, increases the coverage of pedestrian traffic in multi-cycle site selection, has good dynamic environment adaptability and generalization ability, and can maximize the total number of residential users served while ensuring the power supply constraints of the facilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117550A_ABST
    Figure CN123117550A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-period dynamic site selection method and system for mobile public service facilities within urban areas, relating to the fields of urban public service facility site selection planning and artificial intelligence optimization technology. The method includes: acquiring spatial attribute data of the target urban area and configuration data of mobile public service facilities; constructing a regional feature map based on sub-regional division information and feature information; treating each mobile public service facility as an agent, constructing and training a multi-period dynamic site selection model based on the regional feature map, configuration data, and a multi-agent reinforcement learning algorithm based on graph neural networks to obtain a policy network; each agent independently executes site selection decisions according to the policy network, outputting a sequence of site selection locations for each unit time period. This invention solves the problems of poor generalization ability, low training efficiency, and inability to fully utilize road network structure features in existing site selection schemes, thereby improving training efficiency and site selection effectiveness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of urban public service facility site selection planning and artificial intelligence optimization technology, and in particular to a multi-period dynamic site selection method and system for mobile public service facilities within urban areas. Background Technology

[0002] In urban areas, the site selection, planning, and deployment of various mobile public service facilities, such as advertising facilities, mobile health service stations, and express delivery service points, are crucial for serving residents, building smart communities, and improving the urban service system. However, limited by factors such as the scale and cost of public service facilities, the primary objective of site selection and deployment of mobile service facilities is to cover the largest possible number of residents in order to provide sufficient services and effectively utilize the facilities. This optimization problem can be simplified into a set coverage problem with NP-hard complexity.

[0003] Due to the differences in urban functions, pedestrian traffic varies across different areas over time. Mobile public service facilities leverage this characteristic by periodically relocating and adjusting their service locations to ensure they remain in areas with high pedestrian traffic, thereby maximizing the number of users they serve. However, the energy (e.g., electricity) required for the movement of these facilities limits their operational range. The multi-period location problem for energy-constrained mobile public service facilities cannot be simply analogized to multiple consecutive single-period location planning problems. Instead, it requires rational path planning and location scheduling of the mobile service facilities while ensuring their power is not depleted.

[0004] Currently, there is considerable research on single-cycle fixed facility site selection. Since this problem is NP-hard, exact algorithms are not feasible when the number of facilities reaches a certain scale; the main solutions are approximate and heuristic algorithms. In recent years, algorithms based on reinforcement learning and combined with Transformer, Graph Neural Network, and other technologies have become increasingly popular. Compared to approximate and heuristic methods, while ordinary reinforcement learning algorithms can achieve better site selection results, they suffer from long training cycles, unstable training processes, and poor versatility.

[0005] Existing research has proposed several solutions to the dynamic site selection problem for multi-period mobile service facilities. Some experts have proposed a multi-step dynamic site selection planning problem for mobile public service facilities, employing a two-step heuristic algorithm for adaptive scheduling of facilities while considering energy constraints (all facilities must return to their unique origin after service completion). Other experts have evaluated the spatiotemporal distribution of vehicle charging demand across multiple periods and scenarios, proposing a charging station site selection strategy based on evolutionary algorithms. Still others have solved the dynamic site selection problem using deep reinforcement learning and gated recurrent units, but their facility site selection is a multi-period, progressively refined deployment process; facilities selected and deployed within a certain time period remain stationary and persist, which is fundamentally different from multi-step dynamic site selection. Furthermore, some experts have used multi-agent deep reinforcement learning methods to solve the multi-period dynamic attack or defense problem of defense weaponry; however, the requirements for attack path planning and ammunition constraints differ fundamentally from the dynamic planning of mobile service facilities in real urban road networks.

[0006] In summary, the existing technologies have the following shortcomings: (1) Traditional approximation, heuristic and genetic algorithm model designs generally rely on expert knowledge, have poor generalization ability, and cannot adapt to the multi-cycle dynamic location problem of large-scale mobile service facilities when environmental conditions change; (2) Ordinary reinforcement learning algorithms have problems such as long training cycle, unstable training process, and inability to make full use of urban road network structure (graph structure) and road network features. Summary of the Invention

[0007] To address this, embodiments of the present invention provide a multi-period dynamic location method and system for mobile public service facilities within urban areas, which solves the problems of poor generalization ability, low training efficiency, and inability to fully utilize urban road network structural features in existing multi-period dynamic location schemes for mobile public service facilities.

[0008] To address the aforementioned technical problems, embodiments of the present invention provide a multi-period dynamic site selection method for mobile public service facilities within urban areas, comprising: Step S1: Obtain spatial attribute data of the target city area and configuration data of mobile public service facilities within the target city area; the spatial attribute data includes sub-region division information of the target city area and characteristic information of each sub-region; the configuration data includes the number of mobile public service facilities, the initial power of each mobile public service facility, the service capacity of each mobile public service facility, the initial location of each mobile public service facility, and the number of unit time periods included in the total service cycle. Step S2: Construct a region feature map based on the sub-region division information and the feature information; the region feature map uses each sub-region as a node, the adjacency relationship between each sub-region as an edge, and the feature information of each sub-region as node features; Step S3: Treat each of the mobile public service facilities as an agent, and construct a multi-period dynamic location model based on the regional feature map, the configuration data, and a multi-agent reinforcement learning algorithm based on graph neural networks; Step S4: Train the multi-period dynamic addressing model to obtain the trained policy network; Step S5: In the prediction phase, each agent independently performs a location decision based on the trained policy network, the current remaining battery power, and the current sub-region, and moves to and deploys in the target sub-region in the next unit time period. Step S6: Output the location sequence of each of the mobile public service facilities in each of the unit time periods, as a multi-period dynamic location scheme for the target urban area.

[0009] Preferably, in step S1, the sub-region division information is used to divide the target city region into... Independent and disjoint subregions The feature information includes the road network structure, point-of-interest data, public transportation facility data, and spatiotemporal travel trajectory data of residents in each sub-region; the configuration data includes the total service cycle. Classified as Equal time units need to be processed. Step-by-step dynamic location selection; the set of mobile public service facilities is represented as The corresponding service capacity limit within a unit time period is The corresponding initial set of electricity is ,in This refers to the total number of mobile public service facilities; Each service facility is located within the sub-region, represented as a set. , and much smaller All sub-regions in The flow of people within a unit time period is represented as a set. .

[0010] Preferably, in step S2, the region feature map is constructed as an undirected graph structure. ,in This represents a set of nodes, where each graph node corresponds to a sub-region of a city. The adjacency matrix is ​​the set of adjacency relationships for all sub-regions. It is a set of node features; the region feature map is used to perform graph coding representation of the network features of the entire region in reinforcement learning.

[0011] Preferably, in step S3, the multi-period dynamic site selection problem of mobile public service facilities is modeled as a partially observable Markov game mathematical model, using a six-tuple... Indicates; among which, It is the set of global state spaces of all intelligent agents. Represents the set of action spaces for all agents. This represents the set of observation spaces of all intelligent agents in the global space. It is a global reward function shared by all agents. Represents the state transition probability. It is a discount factor; all agents have the same optimization goal and belong to a completely cooperative relationship.

[0012] Preferably, the multi-period dynamic location model is based on a centralized training and decentralized execution architecture, where all agents share a policy network and a value network. The policy network is used to generate action policies for each agent based on their independent observations, and the value network is used to evaluate the joint actions of all agents to guide the parameter updates of the policy network. During the training phase, each agent performs centralized training using global state and independent observations.

[0013] Preferably, the policy network adopts an encoder-decoder structure based on graph attention network and graph pointer network; the encoder takes the observation set of each agent and the regional feature map as input, encodes the input using the graph attention network, and outputs the latent features of the corresponding nodes of each agent after graph attention encoding; the decoder takes the output of the graph attention network as input, and uses the graph pointer network to generate the probability distribution of all mobile public service facilities to be located in each sub-region; the core of the graph pointer network is a gated recurrent unit, which is used to make sequential decisions based on the latent features of the current unit time period and the output of the previous unit time period.

[0014] Preferably, the graph attention network is used for region feature maps. nodes , No. The latent features of a layer are represented as follows: , in, Represents nodes The set of adjacent nodes, Representing nodes in a graph attention network With neighboring nodes The local attention weight parameter is used to characterize the strength of feature association between adjacent nodes; This represents the global weight parameters in the graph attention network, used for linear transformation of all node features; Represents a node In the The hidden features of the layer; The sigmoid non-linear activation function is used to compress the input values ​​to between 0 and 1; the sum of the latent features of all layers is the encoder's output. ,in The total number of layers in the graph attention network.

[0015] Preferably, the value network adopts an encoder-decoder structure based on graph attention networks and fully connected networks; the encoder takes the global state and the region feature map as input and encodes them using the graph attention network; the decoder uses a fully connected network with two hidden layers to decode the output of the encoder and outputs the fitted state value. Preferably, within a unit time period Within the system, the reward function shared by all agents is designed as follows: , in, Indicates a unit of time period Instant reward value within; Indicates the first Sub-regions; Represents the set of all subregions; As an indicator variable, if the subregion Within a unit of time period If at least one mobile public service facility exists within the premises... ,otherwise ; Subregion Within a unit of time period The minimum sum of pedestrian traffic within the sub-area and the service capacity of all mobile public service facilities within that sub-area, i.e. ; This represents the collection of all mobile public service facilities; Indicates mobile public service facilities Within a unit of time period The amount of electricity consumed during internal movement; The weighting parameter for the power consumption penalty is used to normalize the impact of power consumption on the reward. Indicates a unit of time period The number of intelligent agents that cannot return to the location of the service facility; The weight parameters for the returned penalty term are used to normalize the impact of the number of unreturned facilities on the reward; the global reward function is... That is, in all The cumulative sum of instant rewards within each unit time period.

[0016] This invention also provides a multi-period dynamic site selection system for mobile public service facilities within urban areas, comprising: The data acquisition module is used to acquire spatial attribute data of the target city area and configuration data of mobile public service facilities within the target city area; the spatial attribute data includes sub-region division information of the target city area and characteristic information of each sub-region; the configuration data includes the number of mobile public service facilities, the initial power of each mobile public service facility, the service capacity of each mobile public service facility, the initial location of each mobile public service facility, and the number of unit time periods included in the total service cycle. The graph construction module is used to construct a region feature graph based on the sub-region division information and the feature information; the region feature graph uses each sub-region as a node, the adjacency relationship between each sub-region as an edge, and the feature information of each sub-region as node features; The model building module is used to treat each of the mobile public service facilities as an agent and build a multi-period dynamic location model based on the regional feature map, the configuration data and a multi-agent reinforcement learning algorithm based on graph neural networks. The model training module is used to train the multi-period dynamic addressing model to obtain the trained policy network. The deployment and execution module is used to, during the prediction phase, each of the agents independently executes a location decision based on the trained policy network, the current remaining power, and the current sub-region, and moves to and deploys in the target sub-region in the next unit time period. The scheme output module is used to output the location sequence of each of the mobile public service facilities in each of the unit time periods, as a multi-period dynamic site selection scheme for the target urban area.

[0017] As can be seen from the above technical solutions, this invention application has the following beneficial effects: First, it significantly improves the training efficiency and convergence speed of the model: By introducing Regional Feature Map (RCG), this invention encodes the urban road network structure information into a graph model, enabling the policy network to fully explore the spatial association features and sequence dependencies between sub-regions. Compared with ordinary reinforcement learning algorithms, it effectively shortens the training cycle and improves the stability of the training process.

[0018] Second, it significantly improves the coverage of pedestrian traffic in multi-cycle site selection: This invention constructs an encoder-decoder structure based on graph attention network (GAT) and graph pointer network (GPN) to perform fine-grained encoding and decoding of observation state and regional feature map, enabling the agent to make better sequential decisions based on historical state information, thereby maximizing the total number of residents served while ensuring facility power constraints, and significantly improving site selection effect.

[0019] Third, it has good adaptability to dynamic environments and generalization ability: By combining the multi-agent proximal policy optimization (MAPPO) with a graph model, the trained policy network can adaptively cope with dynamic environmental conditions such as changes in pedestrian flow distribution in different urban areas and at different times. It does not require model redesign based on expert knowledge and has strong generalization ability and universality. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Referring to the drawings will make the features and advantages of the present invention clearer. The drawings are illustrative and should not be construed as limiting the present invention in any way. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a multi-period dynamic site selection method for mobile public service facilities within an urban area, provided by an embodiment of the present invention; Figure 2 This is a flowchart of the GMAPPO training process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the policy network Actor structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the value network Critic structure in an embodiment of the present invention; Figure 5 This is a block diagram of a multi-period dynamic site selection system for mobile public service facilities within an urban area, provided by an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1: To address the problems of poor generalization ability, low training efficiency, and inability to fully utilize urban road network structural features in existing multi-period dynamic site selection schemes for mobile public service facilities, this invention proposes a multi-period dynamic site selection method for mobile public service facilities within urban areas, such as... Figure 1 As shown, the method includes the following steps: Step S1: Obtain spatial attribute data of the target city area and configuration data of mobile public service facilities within the target city area.

[0023] Specifically, the spatial attribute data includes sub-region division information of the target city area and feature information of each sub-region. For example... Figure 2 As shown, the target city area Divided into Independent and disjoint subregions The feature information of each sub-region includes data such as road network structure, important points of interest (POI), public transportation facilities, and spatiotemporal travel trajectories of residents in the road network.

[0024] The configuration data includes the number of mobile public service facilities, the initial power of each mobile public service facility, the service capacity of each mobile public service facility, the initial location of each mobile public service facility, and the number of time periods per unit within the total service cycle. Specifically, the total service cycle... Classified as A number of equal unit time periods (sub-cycles) are required, that is, to perform The process involves dynamic location selection; multiple dynamic deployment cycles are required. A set of mobile public service facilities is represented as The corresponding service capacity limit (maximum number of users that can be served) within a unit time period is: The corresponding initial set of electricity is ; Each service facility location (charging station) is located within a sub-region and is represented as a set. , and much smaller All sub-regions in The flow of people within a unit time period is represented as a set. Initially, all mobile public service facilities are located in the service facility storage area (charging station).

[0025] Step S2: Construct a Regional Characteristic Graph (RCG) based on the sub-region division information and feature information.

[0026] The region feature map uses each of the sub-regions as nodes, the adjacency relationships between the sub-regions as edges, and the feature information of each sub-region as node features. Adjacency relationships and characteristics of independent, non-overlapping sub-regions This forms an undirected graph (network) structure, formally represented as: ,in It represents the set of nodes in an undirected graph, where each graph node represents a sub-region of a city. The adjacency matrix is ​​the set of adjacency relationships for all sub-regions; This is the feature set of the nodes. (The graph...) This is the region feature map (RCG).

[0027] The role of region feature maps is to introduce graph models into multi-agent reinforcement learning, enabling subsequent graph attention networks (GAT) and graph pointer networks (GPN) to encode the network features of region division and the environmental state of the agents in a fine-grained manner, thereby improving the overall expressive and generalization capabilities of the model.

[0028] Step S3: Treat each mobile public service facility as an agent and construct a multi-period dynamic location model based on regional feature maps, configuration data, and a multi-agent reinforcement learning algorithm based on graph neural networks.

[0029] In this step, the multi-period dynamic location problem of mobile public service facilities within the urban area is first modeled as a partially observable Markov game (POMG) mathematical model, that is, a mathematical model of multiple agents making sequential decisions under incomplete information and environmental uncertainty, and then modeled using a six-tuple. To represent. Among them, It is the set of global state spaces of all intelligent agents. Represents the set of action spaces for all agents. This represents the set of observation spaces of all intelligent agents in the global space. It is a global reward function shared by all agents. Represents the state transition probability. It is a discount factor. All agents have the same optimization goal, making it a multi-agent reinforcement learning task with complete cooperation.

[0030] Then, based on the above POMG model, a multi-period dynamic location model GMAPPO (Graph-improved MAPPO) is constructed, which is based on regional feature maps (RCG) and multi-agent proximal policy optimization (MAPPO). Figure 2 As shown, the GMAPO framework is based on the Actor-Critic structure and is mainly divided into three parts: the training environment, the policy network Actor, and the value network Critic.

[0031] The GMAPPO model is based on a Centralized Training and Decentralized Execution (CTDE) architecture, where all agents share a single policy network and a value network. During training, each agent utilizes the global state for centralized training, thus fully considering the observed states and individual behaviors of other agents to learn a more accurate policy. The trained policy network (Actor) can then be used for final multi-period dynamic location prediction.

[0032] Step S4: Train the multi-period dynamic addressing model to obtain the trained policy network.

[0033] The training objective of the GMAPPO model is to find all mobile service facilities before they return to their storage location (charging station) and without running out of energy. A dynamic deployment (mobilization) scheme is proposed for a continuous unit of time period to serve as many residential users as possible (i.e., to obtain the maximum cumulative reward value). The design of the reward function will be described in detail below.

[0034] Based on time period (unit of time period) For example, the training process of the GMAPPO model is as follows: Each agent obtains the current environmental state. ( The observations are then merged with the observations of all agents to generate independent observations, which are then passed to the policy network (Actor). The policy network (Actor) generates the action policy (location decision) for each agent based on the observations of all agents, and combines these actions into a joint action. The action is passed on to the environment; the environment generates a reward based on these actions. The shared value network Critic receives rewards from the environment. Given the states and actions of all agents, evaluate the merits of each agent's actions, and then update the parameters of the policy network (Actor) based on the evaluation values. This process is executed continuously in a loop. times (i.e., total service period) This completes one round of training for the policy network Actor.

[0035] Specifically, the policy network Actor is represented as ,in It is a time period The set of independent observations by all agents within the system. , It is the set of parameters for the policy network. For example... Figure 3 As shown, in order to fully utilize and process the region feature map The policy network Actor is designed as an encoder-decoder structure based on the graph attention network GAT and the graph pointer network GPN, so as to encode and decode the observation state and regional feature map, so as to achieve more fine-grained urban regional feature extraction and latent feature representation, and improve the overall performance and generalization ability of the policy network.

[0036] Specifically, in terms of time period For example, during the training phase, the encoder sets the observations of all agents. and regional feature map As input, a graph attention network module is then used to encode the input. For region feature maps... nodes (i.e., intelligent agent) ), No. The latent features of a layer are represented as follows: , in, Represents nodes The set of adjacent nodes; Representing nodes in a graph attention network With neighboring nodes The local attention weight parameter is used to characterize the strength of feature association between adjacent nodes; This represents the global weight parameters in the graph attention network, used for linear transformation of all node features; Represents a node In the The hidden features of the layer; The sigmoid non-linear activation function is used to compress the input values ​​to the range of 0 to 1. The sum of the latent features of all layers is the output of the encoder (graph attention network). ,in The total number of layers in the graph attention network.

[0037] The graph attention network consists of two parts: a decoder and an address probability output (Softmax). The core of the decoder is a gated recurrent unit (GRU), which uses the output of the graph attention network. and its previous time period Output As input, generate the output of the current step. Then, the graph pointer network passes through an attention mechanism layer and a softmax function to the decoder's output. As a reference, and using the output of the gated recurrent unit as a query, the final attention score is calculated, thereby generating the location probabilities of all sub-regions. The entire process of the graph pointer network is calculated as follows: , , , in, , and This represents the weight parameters of the corresponding module. It is a neural network activation function that compresses input values ​​to a range between -1 and 1. It is a constant. This represents the element-wise addition operation of a vector. for The location of all agents within the time period.

[0038] For intelligent agents The objective (loss) function of the policy network Actor is expressed as: , in, To calculate the mathematical expectation. Represents intelligent agents The ratio of the current action strategy to the old action strategy. , It is a clipping function. It's a clipping hyperparameter. These are weight parameters. The agent λ is calculated using the Generalized Advantage Estimation (GAE) method. i The estimate of the dominance function, , and Representing the global state space respectively and The values ​​of the two value networks Critic are below. These are weight parameters. It is the parameter set of the value network Critic. It is the policy entropy function, specifically expressed as: , Finally, the parameters of the policy network are updated using gradient ascent: , in, It is the set of parameters updated by the policy network Actor. It is the set of parameters of the policy network Actor before the update. It is the weight parameter (learning rate). This represents the gradient of the solution function.

[0039] The value network Critic is represented as ,in It is a set of parameters. For example... Figure 4 As shown, the value network is designed as an encoder-decoder structure based on graph attention network and fully connected network (FCN).

[0040] By time period For example, during the training phase, the encoder of the value network Critic uses the global state... and regional feature map As input, these inputs are encoded using a graph attention network; the decoder uses a fully connected network with two hidden layers to decode the encoder's output and finally outputs the fitted state value.

[0041] Accordingly, the objective (loss) function of the value network Critic is expressed as:

[0042] in, Representing the global state of the value network Critic Value estimation, express The true value of the value is determined by the loss function. The optimization objective is to minimize the gap between the estimated value and the actual reward, thereby guiding the updating of the policy network (Actor) so that the addressing policy network of each agent is updated in the direction of optimizing the global value.

[0043] Finally, the parameters of the value network Critic are updated using gradient descent. : , in, It is the set of parameters of the value network Critic before the update, and λ is the weight parameter (learning rate).

[0044] During training, all agents share the same optimization goal: maximizing user coverage. This is a fully cooperative multi-agent reinforcement learning task; therefore, the environmental feedback provides the same reward to each agent. Within a unit of time... Within the system, the reward function shared by all agents is designed as follows: , in, Indicates a unit of time period Instant reward value within; Indicates the first Sub-regions; Represents the set of all subregions; As an indicator variable, if the subregion Within a unit of time period If at least one mobile public service facility exists within the premises... ,otherwise ; Subregion Within a unit of time period The minimum sum of pedestrian traffic within the sub-area and the service capacity of all mobile public service facilities within that sub-area, i.e. ; This represents the collection of all mobile public service facilities; Indicates mobile public service facilities Within a unit of time period The amount of electricity consumed during internal movement. This is a penalty for excessive battery consumption. The weighting parameter for the power consumption penalty is used to normalize the impact of power consumption on the reward. Indicates a unit of time period The number of intelligent agents that cannot return to the location of the service facility; The weight parameters for the returned penalty term are used to normalize the impact of the number of unreturned facilities on the reward. The global reward function is... That is, in all The cumulative sum of instant rewards within each unit time period.

[0045] Step S5: During the prediction phase, each agent independently executes a location decision based on the trained policy network, the current remaining battery power, and the current sub-region, moving to and deploying in the target sub-region in the next unit time period.

[0046] During the prediction phase, each agent performs location decisions in a distributed manner using an independently deployed policy network and its own current state, enabling efficient execution without the need for mutual communication. For each consecutive unit of time... Based on the Actor policy network and the current status of all mobile public service facilities (i.e., remaining power and their respective sub-regions), calculate the unit time period. Collection of mobile public service facilities Site selection action .

[0047] Step S6: Output the location sequence of each mobile public service facility in each time period as a multi-period dynamic location scheme for the target urban area.

[0048] Specifically, output the total service cycle. Collection of mobile public service facilities Dynamic addressing sequence .exist After each time sub-cycle ends, all mobile public service facilities must be returned to their storage location (charging station).

[0049] Example 2: This invention provides a multi-period dynamic site selection system for mobile public service facilities within urban areas. This system is used to implement the multi-period dynamic site selection method for mobile public service facilities within urban areas described in Embodiment 1 above. Figure 5 As shown, it specifically includes: The data acquisition module 100 is used to acquire spatial attribute data of the target city area and configuration data of mobile public service facilities within the target city area. The spatial attribute data includes sub-region division information of the target city area and characteristic information of each sub-region. The configuration data includes the number of mobile public service facilities, the initial power of each mobile public service facility, the service capacity of each mobile public service facility, the initial location of each mobile public service facility, and the number of unit time periods included in the total service cycle. The graph construction module 200 is used to construct a regional feature graph based on sub-region division information and feature information. The regional feature graph uses each sub-region as a node, the adjacency relationship between each sub-region as an edge, and the feature information of each sub-region as node features. The model building module 300 is used to treat each mobile public service facility as an agent and build a multi-period dynamic location model based on regional feature maps, configuration data and a multi-agent reinforcement learning algorithm based on graph neural networks. The model training module 400 is used to train the multi-period dynamic addressing model to obtain the trained policy network. The deployment execution module 500 is used to enable each agent to independently execute location decisions based on the trained policy network, the current remaining power, and the current sub-region during the prediction phase, and move to and deploy in the target sub-region in the next unit time period. The solution output module 600 is used to output the location sequence of each mobile public service facility in each unit time period, as a multi-period dynamic site selection scheme for the target urban area.

[0050] This embodiment provides a multi-period dynamic location system for mobile public service facilities within urban areas, used to implement the aforementioned multi-period dynamic location method for mobile public service facilities within urban areas. Therefore, the specific implementation of the multi-period dynamic location system for mobile public service facilities within urban areas can be found in the previous section on the implementation of the multi-period dynamic location method for mobile public service facilities within urban areas. To avoid redundancy, it will not be repeated here.

[0051] The technical solution provided by the above embodiments of the present invention treats mobile public service facilities as agents and applies graph models and multi-agent reinforcement learning to the multi-period dynamic location problem of mobile service facilities in urban areas. Considering the energy consumption of service facilities moving within the urban road network, a multi-period dynamic location model, GMAPPO, based on regional feature graphs (RCG) and multi-agent proximal strategies to optimize MAPPO, is proposed. The GMAPPO model has strong adaptability to dynamic environments, can better mine environmental sequence data information, and learn the agent sequence structure features and dependencies between adjacent time periods. For multi-period location prediction tasks, it not only has good training efficiency (fast convergence speed) and location effect (total population coverage), but also good generalization ability. The implementation of this solution can bring convenience to the daily work and life of urban residents. By moving and serving between sub-regions with different population flows within different time periods, service facilities can significantly reduce the overall cost of service facilities, while increasing the overall coverage of residents within the region. This improves the city's service system, enhances the city's comprehensive service capabilities, alleviates urban traffic and environmental pressures, and promotes social equity.

[0052] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0053] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0054] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A multi-period dynamic site selection method for mobile public service facilities within an urban area, characterized in that, include: Step S1: Obtain spatial attribute data of the target city area and configuration data of mobile public service facilities within the target city area; the spatial attribute data includes sub-region division information of the target city area and characteristic information of each sub-region; the configuration data includes the number of mobile public service facilities, the initial power of each mobile public service facility, the service capacity of each mobile public service facility, the initial location of each mobile public service facility, and the number of unit time periods included in the total service cycle. Step S2: Construct a region feature map based on the sub-region division information and the feature information; the region feature map uses each sub-region as a node, the adjacency relationship between each sub-region as an edge, and the feature information of each sub-region as node features; Step S3: Treat each of the mobile public service facilities as an agent, and construct a multi-period dynamic location model based on the regional feature map, the configuration data, and a multi-agent reinforcement learning algorithm based on graph neural networks; Step S4: Train the multi-period dynamic addressing model to obtain the trained policy network; Step S5: In the prediction phase, each agent independently performs a location decision based on the trained policy network, the current remaining battery power, and the current sub-region, and moves to and deploys in the target sub-region in the next unit time period. Step S6: Output the location sequence of each of the mobile public service facilities in each of the unit time periods, as a multi-period dynamic location scheme for the target urban area.

2. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 1, characterized in that, In step S1, the sub-region division information is used to divide the target city region into... Independent and non-overlapping subregions The feature information includes the road network structure, point-of-interest data, public transportation facility data, and spatiotemporal travel trajectory data of residents in each sub-region; the configuration data includes the total service cycle. Classified as Equal time units need to be processed. Step-by-step dynamic location selection; the set of mobile public service facilities is represented as The corresponding service capacity limit within a unit time period is The corresponding initial set of electricity is ,in This refers to the total number of mobile public service facilities; Each service facility is located within the sub-region, represented as a set. , and much smaller All sub-regions in The flow of people within a unit time period is represented as a set. .

3. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 1, characterized in that, In step S2, the region feature map is constructed as an undirected graph structure. ,in This represents a set of nodes, where each graph node corresponds to a sub-region of a city. The adjacency matrix is ​​the set of adjacency relationships for all sub-regions. For node feature set; The region feature map is used to perform graph encoding representation of the network features of the entire region in reinforcement learning.

4. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 1, characterized in that, In step 3, the multi-period dynamic site selection problem of mobile public service facilities is modeled as a partially observable Markov game mathematical model, using a six-tuple... Indicates; among which, It is the set of global state spaces of all intelligent agents. Represents the set of action spaces for all agents. This represents the set of observation spaces of all intelligent agents in the global space. It is a global reward function shared by all agents. Represents the state transition probability. It is a discount factor; all agents have the same optimization goal and belong to a completely cooperative relationship.

5. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 4, characterized in that, The multi-period dynamic location model is based on a centralized training and decentralized execution architecture, where all agents share a policy network and a value network. The policy network is used to generate action policies for each agent based on their independent observations, and the value network is used to evaluate the joint actions of all agents to guide the parameter updates of the policy network. During the training phase, each agent performs centralized training using global state and independent observations.

6. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 5, characterized in that, The policy network adopts an encoder-decoder structure based on graph attention network and graph pointer network. The encoder takes the observation set of each agent and the regional feature map as input, encodes the input using the graph attention network, and outputs the latent features of the corresponding nodes of each agent after graph attention encoding. The decoder takes the output of the graph attention network as input and uses the graph pointer network to generate the probability distribution of all mobile public service facilities to be located in each sub-region. The core of the graph pointer network is a gated recurrent unit, which is used to make sequential decisions based on the latent features of the current unit time period and the output of the previous unit time period.

7. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 6, characterized in that, The graph attention network is for region feature maps nodes , No. The latent features of a layer are represented as follows: , in, Represents nodes The set of adjacent nodes, Representing nodes in a graph attention network With neighboring nodes The local attention weight parameter is used to characterize the strength of feature association between adjacent nodes; This represents the global weight parameters in the graph attention network, used for linear transformation of all node features; Represents a node In the The hidden features of the layer; The sigmoid non-linear activation function is used to compress the input values ​​to between 0 and 1; the sum of the latent features of all layers is the encoder's output. ,in The total number of layers in the graph attention network.

8. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 5, characterized in that, The value network adopts an encoder-decoder structure based on graph attention network and fully connected network; the encoder takes the global state and the region feature map as input and encodes them using the graph attention network; The decoder uses a fully connected network with two hidden layers to decode the encoder's output, outputting the fitted state value.

9. The multi-period dynamic site selection method for mobile public service facilities within urban areas according to claim 4, characterized in that, Within a unit of time period Within the system, the reward function shared by all agents is designed as follows: , in, Indicates a unit of time period Instant reward value within; Indicates the first Sub-regions; Represents the set of all subregions; As an indicator variable, if the subregion Within a unit of time period If at least one mobile public service facility exists within the premises... ,otherwise ; Subregion Within a unit of time period The minimum sum of pedestrian traffic within the sub-area and the service capacity of all mobile public service facilities within that sub-area, i.e. ; This represents the collection of all mobile public service facilities; Indicates mobile public service facilities Within a unit of time period The amount of electricity consumed during internal movement; The weighting parameter for the power consumption penalty is used to normalize the impact of power consumption on the reward. Indicates a unit of time period The number of intelligent agents that cannot return to the location of the service facility; The weight parameters for the returned penalty term are used to normalize the impact of the number of unreturned facilities on the reward; the global reward function is... That is, in all The cumulative sum of instant rewards within each unit time period.

10. A multi-period dynamic site selection system for mobile public service facilities within an urban area, characterized in that, include: The data acquisition module is used to acquire spatial attribute data of the target city area and configuration data of mobile public service facilities within the target city area; the spatial attribute data includes sub-region division information of the target city area and characteristic information of each sub-region; the configuration data includes the number of mobile public service facilities, the initial power of each mobile public service facility, the service capacity of each mobile public service facility, the initial location of each mobile public service facility, and the number of unit time periods included in the total service cycle. The graph construction module is used to construct a region feature graph based on the sub-region division information and the feature information; the region feature graph uses each sub-region as a node, the adjacency relationship between each sub-region as an edge, and the feature information of each sub-region as node features; The model building module is used to treat each of the mobile public service facilities as an agent and build a multi-period dynamic location model based on the regional feature map, the configuration data and a multi-agent reinforcement learning algorithm based on graph neural networks. The model training module is used to train the multi-period dynamic addressing model to obtain the trained policy network. The deployment and execution module is used to, during the prediction phase, each of the agents independently executes a location decision based on the trained policy network, the current remaining power, and the current sub-region, and moves to and deploys in the target sub-region in the next unit time period. The scheme output module is used to output the location sequence of each of the mobile public service facilities in each of the unit time periods, as a multi-period dynamic site selection scheme for the target urban area.