Multi-agent emergency evacuation method and storage medium in dynamic radiation scene of nuclear accident

Abstract

The application discloses a multi-agent emergency evacuation method under a dynamic radiation scene of a nuclear accident and a storage medium, and the method comprises the following steps: firstly, a route network topology is constructed by using a set of longitude and latitude fine-grained grids, radiation diffusion is simulated by combining a Gaussian plume model, and environment modeling is completed by fusing a dynamic space-time interaction environment matrix; secondly, an agent state / action space and a multi-dimensional comprehensive reward function are designed, an Actor-Critic collaborative network of the MAPPO algorithm is built, and multi-agent training is completed; then, a globally optimal evacuation trajectory is generated; finally, a compliant evacuation strategy is output by quantitatively evaluating and outputting global evacuation total time, per capita radiation dose and utilization rate indexes of an evacuation terminal. The application realizes coupling modeling of a dynamic radiation field and a complex route network, solves scheduling conflicts of radiation avoidance and multi-source and multi-sink cooperation, fills a technical blank of a traditional method in a high-dynamic risk scene of a nuclear emergency, improves practicality and reliability of an evacuation scheme, and provides strong support for nuclear emergency command.

Description

Technical Field

[0001] This invention relates to the field of nuclear accident emergency evacuation technology, specifically to a multi-agent emergency evacuation method and storage medium under dynamic radiation scenarios of nuclear accidents. Background Technology

[0002] Nuclear accident emergency evacuation is a crucial step in ensuring public safety and mitigating potential disaster losses. Existing research on emergency evacuation optimization largely focuses on static path planning for routine emergencies such as floods and earthquakes, including shortest distance analysis and road capacity assessment. However, current evacuation planning systems still have shortcomings when dealing with the highly unique and dynamic high-risk scenario of nuclear accidents. The unique challenges of nuclear accident emergency evacuation include: ① Evacuation risks stem not only from simple traffic delays but also from the dynamic superposition effect of "spatiotemporal diffusion of radioactive materials + continuous exposure of the population"; ② The evacuation physical space is complex, involving multi-starting point to multi-endpoint coordination under a vast road network topology, and each evacuation endpoint has strict capacity limits, which can easily lead to local congestion; ③ The environmental conditions change drastically over time, such as the real-time drift of the nuclear radiation plume and dynamic congestion of the road network, which traditional planning methods struggle to capture.

[0003] Traditional evacuation planning methods, such as graph-based static pathfinding algorithms and traditional microscopic simulations, have significant limitations in multi-agent emergency evacuation scenarios during nuclear accidents. Graph-based static pathfinding algorithms, such as the A* algorithm, rely solely on fixed road rights, setting the "cost" of traversing a particular road as a fixed, unchanging value. This fails to address the dynamic changes in the radiation field over time, potentially leading to planned paths traversing highly contaminated areas. Traditional microscopic behavioral models, such as social force models, suffer from severe "curse of dimensionality" when facing complex road network topology constraints, large-scale multi-agent obstacle avoidance, and endpoint capacity games. Neither method can effectively balance the contradiction between "dynamic radiation avoidance" and "multi-source, multi-sink coordinated scheduling," resulting in evacuation schemes that are computationally inefficient and unreliable in complex and ever-changing nuclear emergency scenarios. Summary of the Invention

[0004] The present invention proposes a multi-agent emergency evacuation method and storage medium for dynamic radiation scenarios in nuclear accidents, which can at least solve one of the technical problems in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent emergency evacuation method for dynamic radiation scenarios in nuclear accidents, comprising the following steps executed via computer equipment. S1: Construct a dynamic evacuation environment model with coupled radial road network to provide an accurate global environmental benchmark for subsequent multi-agent decision-making and learning. Specifically, the road network topology is built through a 0.001° latitude and longitude fine-grained grid, and a two-dimensional Gaussian plume model is used to simulate the spatiotemporal diffusion of radiation. Then, the road network, radiation field and crowd congestion characteristics are integrated to generate a dynamic spatiotemporal interactive environment matrix that is updated over time. S2: Design a MAPPO multi-agent reinforcement learning system adapted to nuclear emergency scenarios. Based on the dynamic evacuation environment model built in S1, set the local / global state space and discrete action space of the agent, design a multi-dimensional comprehensive reward function that takes into account the requirements of radiation avoidance and capacity constraints, build an Actor-Critic dual-core multi-source multi-sink global collaborative network, and clarify the core rules of agent decision-making and interaction. S3: Conduct global collaborative training of multiple agents and achieve network convergence. Using the dynamic environment of S1 as the interaction carrier and the learning system of S2 as the core framework, calculate the agent's action advantage function through generalized advantage estimation. Update the Actor and Critic network parameters by using the truncated objective function and the mean square error minimization strategy respectively. After multiple rounds of iterative training, the agent learns the optimal decision-making strategy in the nuclear emergency scenario. S4: Generate the globally optimal evacuation strategy based on the converged model, inject the real-time parameters of the nuclear accident and the initial state of the crowd into the environment matrix of S1 to complete the initialization, realize multi-agent distributed collaborative pathfinding based on the Actor network trained in S3, complete the adaptive scheduling of the refuge destination capacity with the help of the global game mechanism of the algorithm, and record the trajectory to generate a set of globally multi-agent spatiotemporally optimal evacuation trajectories. S5: Quantitatively evaluate and output compliant nuclear emergency evacuation decision-making schemes. For the evacuation trajectory set generated by S4, construct an evaluation index system including the total global evacuation time, the average cumulative radiation dose per person, and the utilization rate of the refuge endpoint, and determine the safety threshold and collaborative balance. If the requirements are met, package and output the visualized decision-making scheme; if not, return to S1 to adjust the modeling parameters and re-execute the subsequent full-process planning.

[0006] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0007] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0008] As can be seen from the above technical solution, the multi-agent emergency evacuation method under the dynamic radiation scenario of nuclear accidents of the present invention aims to address the contradiction between the characteristics of the dynamic spatiotemporal diffusion of nuclear accident radiation plumes and the constraints of the massive road network topology. It constructs an intelligent evacuation path planning model that can cope with high-dimensional complex environments, has global collaborative scheduling capabilities, and meets the capacity limit of the endpoint. This enables efficient and safe evacuation of large-scale populations from multiple starting points to multiple endpoints, filling the technical gap of existing graph theory-based static pathfinding algorithms and traditional microscopic simulations in dynamic high-risk scenarios of nuclear emergencies.

[0009] This invention addresses the high-risk scenarios of nuclear accidents, characterized by the overlapping of dynamic spatiotemporal diffusion of radiation plumes and continuous exposure of populations, as well as the constraints of massive road networks and endpoint capacity limitations. It innovatively constructs a road network topology map using a fine-grained latitude and longitude grid. This embodiment employs a 0.001 latitude and longitude grid, achieving coupled and collaborative modeling of dynamic radiation field evolution and complex physical spatial road networks. By deeply coupling the MAPPO multi-agent reinforcement learning algorithm with the dynamic radiation environment, and combining it with a multi-dimensional comprehensive reward function to resolve scheduling conflicts between radiation avoidance and multi-source / multi-sink coordination, while introducing fine-grained grid state encoding to handle high-dimensional complex environmental features, the algorithm utilizes global collaboration and game theory mechanisms to achieve adaptive capacity at evacuation endpoints, thereby outputting the optimal evacuation strategy. This invention fills the technical gap in applying traditional static pathfinding algorithms and microscopic simulations to high-dynamic-risk scenarios in nuclear emergencies, improving the practicality and reliability of evacuation plans, and providing strong support for nuclear accident emergency command, public safety protection, and efficient scheduling of large-scale populations.

[0010] Compared with the prior art, the advantages of the present invention are as follows: 1. Targeted Breakthrough: Addressing the unique high-risk scenario of nuclear accidents, where the dynamic spatiotemporal diffusion of radiation plumes overlaps with continuous exposure of the population, a dedicated dynamic emergency evacuation model is constructed, filling the technological gap of traditional static pathfinding algorithms in highly dynamic and risky nuclear emergency scenarios. Innovatively, a massive road network topology map is constructed using a 0.001 latitude and longitude fine-grained grid, achieving coupled and collaborative modeling of dynamic radiation field evolution and complex physical spatial road networks.

[0011] 2. Synergistic Innovation of Algorithm and Multiple Constraint Coupling: The MAPPO reinforcement learning algorithm is deeply coupled with the dynamic radiation and diffusion environment. By designing a multi-dimensional comprehensive reward function, the scheduling conflict between "dynamic radiation avoidance" and "multi-source and multi-sink coordination" is solved. The fine-grained grid state coding of 0.001 latitude and longitude is introduced to enhance the agent's adaptability to complex road network topology constraints and drastic spatiotemporal evolution data. The MAPPO algorithm comprehensively handles multi-agent game and refuge endpoint capacity constraints in high-dimensional continuous space, which is more efficient and reliable than traditional static pathfinding and microscopic simulation.

[0012] 3. Improved practicality and reliability of evacuation strategies: It provides the optimal evacuation strategy output of "global multi-agent collaboration + endpoint capacity adaptation", realizing efficient scheduling of large-scale populations from multiple starting points to multiple endpoints, and providing a decision support tool that is both scientific and rigorous and has a strong global scheduling capability for nuclear accident emergency command and public life safety protection. Attached Figure Description

[0013] Figure 1 This is a flowchart of the present invention; Figure 2 This is a road topology map of a certain area according to an embodiment of the present invention; Figure 3 This is a path diagram for the last five rounds in an embodiment of the present invention; Figure 4 The road fit rate (yellow, to verify whether the road is being followed) and success rate (green) are shown in the embodiments of the present invention. Detailed Implementation

[0014] 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 some embodiments of the present invention, but not all embodiments.

[0015] like Figure 1 As shown in this embodiment, the multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents includes the following steps: S1: Construct a dynamic evacuation environment model with coupled radial road network to provide an accurate global environmental benchmark for subsequent multi-agent decision-making and learning. Specifically, this involves setting up a fine-grained grid of latitude and longitude to build the road network topology, simulating the spatiotemporal diffusion of radiation using a two-dimensional Gaussian plume model, and then integrating the road network, radiation field, and crowd congestion characteristics to generate a dynamic spatiotemporal interactive environment matrix that is updated over time.

[0016] S2: Design a MAPPO multi-agent reinforcement learning system adapted to nuclear emergency scenarios. Based on the environment model built in S1, set the local / global state space and discrete action space of the agent, design a multi-dimensional comprehensive reward function that takes into account the requirements of radiation avoidance and capacity constraints, build an Actor-Critic dual-core multi-source multi-sink global collaborative network, and clarify the core rules of agent decision-making and interaction.

[0017] S3: Conduct global collaborative training of multiple agents and achieve network convergence. Using the dynamic environment of S1 as the interaction carrier and the learning system of S2 as the core framework, calculate the agent's action advantage function through generalized advantage estimation. Update the Actor and Critic network parameters by using the truncated objective function and the mean square error minimization strategy respectively. After multiple rounds of iterative training, the agent learns the optimal decision-making strategy in the nuclear emergency scenario.

[0018] S4: Generate the globally optimal evacuation strategy based on the converged model, inject the real-time parameters of the nuclear accident and the initial state of the crowd into the environment matrix of S1 to complete the initialization, realize multi-agent distributed collaborative pathfinding based on the Actor network trained in S3, complete the adaptive scheduling of the refuge destination capacity with the help of the global game mechanism of the algorithm, and record the trajectory to generate a set of globally multi-agent spatiotemporally optimal evacuation trajectories.

[0019] S5: Quantitatively evaluate and output compliant nuclear emergency evacuation decision-making schemes. For the evacuation trajectory set generated by S4, construct an evaluation index system including the total global evacuation time, the average cumulative radiation dose per person, and the utilization rate of the refuge endpoint, and determine the safety threshold and collaborative balance. If the requirements are met, package and output the visualized decision-making scheme; if not, return to S1 to adjust the modeling parameters and re-execute the subsequent full-process planning.

[0020] The following is a detailed explanation: 1. Evacuation Environment Modeling In response to the highly dynamic risk characteristics of nuclear accidents, this method first constructs an evacuation environment model that deeply couples physical space with dynamic radiation, providing a precise interactive foundation for multi-agent reinforcement learning.

[0021] 1.1 Construction of fine-grained road network topology with latitude and longitude grids at a resolution of 0.001° To achieve high-precision microscopic evacuation simulation, a fine-grained grid with a resolution of approximately 111 meters and a latitude / longitude of 0.001 was introduced to discretize the real urban road network. The latitude / longitude boundaries of the study area are set as follows: Divide it into The two-dimensional grid matrix G: ; In the formula, This represents the maximum latitude of the study area under the condition that North latitude is positive and South latitude is negative. The minimum latitude of the study area, This indicates the total number of grid rows in the latitudinal direction; This represents the maximum longitude of the study area under the condition that east longitude is positive and west longitude is negative. The minimum longitude of the study area, This indicates the total number of grid columns in the longitude direction.

[0022] Furthermore, set the grid node coordinates as In the formula Indicates that it is located at the th Column, No. The coordinates of the grid nodes in the row. Define the connectivity matrix between grid nodes. If there are passable roads between adjacent grid nodes, then Otherwise, it is 0. This forms the basic road network topology. and in each grid The above includes road capacity attributes. This indicates the road capacity of the area represented by the grid, such as the maximum number of vehicles or lanes.

[0023] 1.2 Dynamic Spatiotemporal Diffusion Analysis of Radiation A two-dimensional Gaussian plume mathematical model is used to simulate the spatiotemporal diffusion process of nuclear radioactive materials with wind direction, wind speed, and time. Let the number of grid nodes be [value missing]. The radiation concentration at that location is : ; In the formula, The rate of radioactive source release; The coordinates of the mesh nodes of the leakage source; Average wind speed; The diffusion parameter is a two-dimensional parameter that is dynamically updated over time t.

[0024] Furthermore, the first The cumulative radiation dose of an agent during continuous exposure during evacuation. The calculation is a discrete summation over the time step: ; In the formula, For the first The coordinates of the grid node where the agent is located at time t. For time step.

[0025] 1.3 Dynamic Spatiotemporal Interactive Environment Matrix By integrating static road network topology with dynamic radiation field, a dynamic spatiotemporal interactive environment matrix is ​​constructed. Each element in the matrix contains an eigenvector. : ; In the formula, This represents the agent density within the grid node at time t, used to reflect the dynamic congestion situation of the crowd in real time. This matrix is ​​updated continuously over time step t, serving as a global environmental benchmark for agent decision-making.

[0026] MAPPO Multi-Agent Reinforcement Learning Design 2.1 Setting the state and action space of the agent (1) State space: Suppose the system includes The first intelligent agent. For the first... An intelligent agent, its local observation state Includes information on its own location, the radiation distribution within its surrounding field of vision, road congestion levels, and the target evacuation destination: ; In the formula, At time t, the first The two-dimensional coordinates of the grid node where each agent is located; Slice the environment matrix within its field of view grid; This refers to its distance from each refuge destination; The current remaining capacity of each refuge destination. Global state. For use by a centralized Critic network.

[0027] (2) Action space: Based on a latitude and longitude grid with a resolution of 0.001°, a discrete action space is defined. . No. An agent performs an action at time t. And transfer to the adjacent connected grid.

[0028] 2.2 Design of the Comprehensive Reward Function Furthermore, to resolve the contradiction between "dynamic radiation avoidance" and "endpoint capacity limitation," a multi-dimensional comprehensive reward function is designed. Guide the agent to learn the optimal strategy: ; In the formula, , which are the preset weighting coefficients: Introductory Rewards A positive reward is given for approaching the finish line, and a large positive reward is given for reaching the finish line.

[0029] Radiation avoidance penalty : ; In the formula, The preset proportionality coefficient, This is a negative reward that is proportional to the radiation concentration of the current grid.

[0030] Capacity game penalty : ; In the formula, A large negative reward is preset; if the reached refuge destination has reached its capacity limit, a severe penalty is imposed (simulating a failed run), forcing it to replan its route: Congestion / Collision Penalty If moved to the grid ,in This represents the population density of the grid at time t. The maximum road capacity is used to impose congestion penalties when the population density exceeds the road capacity.

[0031] 2.3 Multi-source and multi-sink global collaborative network model The model contains two core neural networks: Actor Network Each intelligent agent The agent outputs the probability distribution of actions based on its own local observations. To accelerate convergence, multiple agents share the Actor network parameters. .

[0032] Critic Network During the training phase, the Critic network receives the global state. (Including the positions of all agents and the global radiation field), the global value function evaluates the current state, with parameters as follows: .

[0033] Multi-start-to-multi-endpoint collaborative training based on the MAPPO algorithm Dominance function calculation: Calculate the agent's advantage function using generalized advantage estimation (GAE) Quantify the quality of the current action compared to the average level: ; Where T represents the terminal time step; TD error ; This is a preset discount factor; These are preset GAE smoothing parameters used to balance variance and bias.

[0034] Actor network parameter updates (coupled radiation avoidance and capacity limitation): The step size is updated by using the PPO truncation objective function constraint strategy to ensure training stability in a highly dynamic nuclear radiation environment. ; In the formula, K represents the number of agents, and the probability ratio is... , The preset truncation hyperparameters, The policy entropy is used to encourage exploration of complex road networks; the larger the value, the higher the likelihood of the agent exploring in the early stages. This is the entropy coefficient.

[0035] Critic network parameter updates: Minimize the mean squared error between the predicted value and the actual return: ; in To accumulate discount rewards, the network gradually converges through tens of thousands of training episodes, and the agent learns to automatically find the path with the lowest radiation and remaining capacity at the destination in a dynamic radiation field.

[0036] Generate the globally optimal evacuation strategy During the actual inference (execution) phase, each agent retains only the trained Actor network. .

[0037] Initial state injection: Input the population size at each starting point and real-time parameters of the nuclear accident to initialize the environment matrix. .

[0038] Distributed cooperative pathfinding: all agents input local observations in parallel The next action is directly output through forward propagation of the Actor network. .

[0039] Dynamic game theory and adaptation: In the process of moving towards multiple destinations, agents implicitly possess the ability to coordinate and schedule globally. When a destination is about to reach its capacity limit, the agent will adaptively detour to other safe and capable destinations based on the strategies learned during training, thereby avoiding local deadlocks and secondary congestion.

[0040] Generation strategy: Record all agents from The time-series location coordinates from the arrival at the safe endpoint are used to generate the set of spatiotemporally optimal evacuation trajectories for global multi-agent systems.

[0041] Nuclear accident emergency evacuation decision-making and output The generated global evacuation strategy is quantitatively evaluated and determined. A set of evaluation indicators is constructed. : Total evacuation time The time it takes for the last agent to arrive at the safe haven.

[0042] Average cumulative radiation dose : ; like or an individual If the threshold specified by nuclear safety regulations is exceeded, an early warning will be triggered, and the process will return to step 1 to adjust the parameters and replan.

[0043] Refugee terminal utilization rate : To examine the balance of multi-source and multi-sink synergy.

[0044] ; In the formula, This represents the capacity of the j-th shelter that has been used, usually referring to the actual number of refugees accommodated. This represents the maximum capacity of the j-th refuge site.

[0045] Furthermore, strategies that meet the safety threshold criteria will be packaged and output. The system will automatically generate a visualized dynamic evacuation route map, a road network traffic flow forecast table for each time period, and a shelter personnel reception and dispatch table, providing the emergency command center with a decision-making solution that is globally optimal and highly reliable.

[0046] like Figure 2 The image shows a fine-grained road topology at 0.001 latitude and longitude in a certain area (not all roads are displayed). Only highways, urban expressways, national highways, provincial highways, county roads, township roads, and residential roads are shown. If needed, road topologies for different areas or with different levels of granularity can be obtained. Specifically... Figure 2 This is a road topology map of a certain district. A certain nuclear power plant is located in a certain district, and the nuclear power plant is mainly located in coastal cities. The road network of a certain district is representative and universal. This invention is based on the road topology map of a certain district. It can also be said that the road topology map of a certain district is a constraint environment. The fine-grained road topology of 0.001° latitude and longitude represents a high degree of refinement. In vehicle route planning, this is a high degree of refinement.

[0047] The following is the design of the reward function that has already been added: The road reward function is specifically mapped to each different road grid. Higher-level roads are given higher rewards, such as 2 points for highways and 0.8 points for residential roads, to train drivers to adopt strategies that conform to human driving habits (prioritizing main roads and avoiding back alleys).

[0048] Zero tolerance for derailment penalty: if the vehicle goes out of bounds or into off-road areas, a huge negative reward of -300.0 is given, and the training round ends. Successfully reaching the finish line grants a huge positive reward of 1000; if the finish line has reached its capacity limit, a reward of -20 is given. The reward system calculates the latitude and longitude distances from the finish line for both the previous and current actions. If the distance decreases, a reward of 25 times the decrease in latitude and longitude is given; if the distance increases, a reward of -3 is given. Figure 3 and Figure 4 The training results are generated by pairing 10 agents, starting from 5 starting points and reaching 4 ending points, with the ending points limited to 1, 2, 3, and 4.

[0049] from Figure 3The path printing of the last five rounds of the multi-agent system shows that the agents have learned to allocate the globally optimal path. From the path printing of each round, each destination reaches the capacity limit exactly, indicating that the utilization rate of the refuge point has reached the maximum, which is of great significance in nuclear emergency evacuation.

[0050] Figure 4 The yellow curve represents the average road fit rate of 10 agents. Initially, the yellow curve did not reach 100%, indicating that the agents did not explore the road exactly when they first started, which is unreasonable in actual vehicle driving. Therefore, under the setting of the reward environment, the agents quickly learned the rule of exploring the road, and the road fit rate stabilized at 100%.

[0051] The green curve represents the success rate of 10 agents reaching the endpoint. The green curve in the figure generally tends to converge, eventually converging at nearly 100%. During the upward process, there is a point where the success rate of the agents drops sharply because the entropy of the agent's exploration settings is relatively high. In the early stage, the agents changed strategies too quickly, and the agents chose a less effective strategy, which caused the success rate to drop sharply. Later, the agents corrected themselves during the learning process and solved this problem well.

[0052] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0053] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0054] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the multi-agent emergency evacuation methods under dynamic radiation scenarios of nuclear accidents in the above embodiments.

[0055] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0056] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0058] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent emergency evacuation method under dynamic radiation scenarios of a nuclear accident, characterized in that, Includes the following steps, S1: Construct a dynamic evacuation environment model with radial road network coupling. Specifically, the road network topology is built by setting a fine-grained grid of latitude and longitude, and a two-dimensional Gaussian plume model is used to simulate the spatiotemporal diffusion of radiation. Then, the road network, radiation field and crowd congestion characteristics are integrated to generate a dynamic spatiotemporal interactive environment matrix that is updated over time. S2: Design a MAPPO multi-agent reinforcement learning system adapted to nuclear emergency scenarios. Based on the dynamic evacuation environment model built in S1, set the local / global state space and discrete action space of the agent, design a multi-dimensional comprehensive reward function that takes into account the requirements of radiation avoidance and capacity constraints, build an Actor-Critic dual-core multi-source multi-sink global collaborative network, and clarify the core rules of agent decision-making and interaction. S3: Conduct global collaborative training of multiple agents and achieve network convergence. Using the dynamic environment of S1 as the interaction carrier and the learning system of S2 as the core framework, calculate the agent's action advantage function through generalized advantage estimation. Update the Actor and Critic network parameters by using the truncated objective function and the mean square error minimization strategy respectively. After multiple rounds of iterative training, the agent learns the optimal decision-making strategy in the nuclear emergency scenario. S4: Generate the globally optimal evacuation strategy based on the converged model, inject the real-time parameters of the nuclear accident and the initial state of the crowd into the environment matrix of S1 to complete the initialization, realize multi-agent distributed collaborative pathfinding based on the Actor network trained in S3, complete the adaptive scheduling of the refuge destination capacity with the help of the global game mechanism of the algorithm, and record the trajectory to generate a set of globally multi-agent spatiotemporally optimal evacuation trajectories. S5: Quantitatively evaluate and output compliant nuclear emergency evacuation decision-making plans. For the evacuation trajectory set generated by S4, construct an evaluation index system including total global evacuation time, cumulative radiation dose per person, and utilization rate of refuge endpoints, and determine safety thresholds and collaborative balance. If the requirements are met, a visual decision-making solution will be packaged and output; otherwise, return to S1 to adjust the modeling parameters and re-execute the subsequent full-process planning.

2. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 1, characterized in that: S1's construction of the road network topology by setting a fine-grained grid of latitude and longitude specifically includes the following: A fine-grained grid with latitude and longitude 'a' is introduced to discretize the real urban road network; the latitude and longitude boundaries of the study area are set as follows: Divide it into The two-dimensional grid matrix G: ; In the formula, This represents the maximum latitude of the study area under the condition that North latitude is positive and South latitude is negative. The minimum latitude of the study area, This indicates the total number of grid rows in the latitudinal direction; This represents the maximum longitude of the study area under the condition that east longitude is positive and west longitude is negative. The minimum longitude of the study area, This indicates the total number of grid columns along the longitude direction; Set the grid node coordinate set to In the formula Indicates that it is located at the th Column, No. The coordinates of the grid nodes in the row define the connectivity matrix between grids. If there are passable roads between adjacent grid nodes, then Otherwise, it is 0; thus, a basic road network topology diagram is constructed. and in each grid The above includes road capacity attributes. , which represents the road capacity of the area represented by the grid.

3. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 2, characterized in that: The steps in simulating the spatiotemporal diffusion of radiation using the two-dimensional Gaussian plume model in S1 include: A two-dimensional Gaussian plume mathematical model is used to simulate the spatiotemporal diffusion process of nuclear radioactive materials with wind direction, wind speed, and time; assuming... At any given moment, the grid node The radiation concentration at that location is : ； In the formula, The rate of radioactive source release; The coordinates of the mesh nodes of the leakage source; Average wind speed; For two-dimensional diffusion parameters, as of time Dynamically updated; The first The cumulative radiation dose of an agent during continuous exposure during evacuation. The calculation is a discrete summation over the time step: ； In the formula, For the first An intelligent agent in The grid node coordinates at any given time. For time step.

4. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 3, characterized in that: S1 further integrates road network, radiation field, and crowd congestion characteristics to generate a dynamic spatiotemporal interactive environment matrix that updates dynamically over time, including: By integrating static road network topology with dynamic radiation field, a dynamic spatiotemporal interactive environment matrix is ​​constructed. Each element in the matrix contains an eigenvector. : ； In the formula, express The density of agents within this grid node at any given time is used to reflect the dynamic congestion situation of the crowd in real time; this matrix changes with time steps. It is continuously updated as a global environmental benchmark for agent decision-making.

5. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 4, characterized in that: S2 defines the agent's local / global state space and discrete action space, including, State space: Assume the system contains The first intelligent agent; for the first An intelligent agent, its local observation state Includes information on its own location, the radiation distribution within its surrounding field of vision, road congestion levels, and the target evacuation destination: ； In the formula, for At this moment, the first The two-dimensional coordinates of the grid node where each agent is located; Slice the environment matrix within its field of view grid; This refers to its distance from each refuge destination; The current remaining capacity of each refuge destination; Global state For use by a centralized Critic network; Action space: based on resolution The latitude and longitude grid defines the discrete action space. , No. An intelligent agent in Perform actions at all times And transfer to the adjacent connected grid.

6. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 5, characterized in that: S2's design incorporates a multi-dimensional comprehensive reward function that considers radiation avoidance and capacity constraints, including... Design a multidimensional comprehensive reward function Guide the agent to learn the optimal strategy: ； In the formula, , which are the preset weighting coefficients: Incentive rewards: Positive rewards are given for approaching the finish line, and a large positive reward is given for reaching the finish line; To evade punishment for radiation: ； In the formula, The preset proportionality coefficient, This is a negative reward that is proportional to the radiation concentration of the current grid. Penalties for capacity games: ； In the formula, As a pre-set, large negative reward, if the reached refuge destination has reached its capacity limit, a severe penalty is imposed, forcing it to replan its route: Congestion / collision penalty: If the target grid is moved to... ,in This represents the population density of the grid at time t. The maximum road capacity is used to impose congestion penalties when the population density exceeds the road capacity.

7. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 6, characterized in that: S2 constructs an Actor-Critic dual-core multi-source multi-sink global collaborative network, clarifying the core rules for agent decision-making and interaction, and includes two core neural networks: Actor Network Each intelligent agent The action probability distribution is output based on the agent's own local observations; to accelerate convergence, multiple agents share the Actor network parameters. ; Critic Network During the training phase, the Critic network receives the global state. It includes the locations of all agents and the global radiation field, and evaluates the global value function of the current state with parameters as follows: .

8. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 7, characterized in that: S3 employs multi-start-point to multi-endpoint collaborative training based on the MAPPO algorithm. include, Dominance function calculation: Calculating the agent's advantage function using generalized advantage estimation Quantify the quality of the current action compared to the average level: ； Where T represents the terminal time step; TD error ; This is a preset discount factor; These are preset GAE smoothing parameters used to balance variance and bias; Actor network parameter updates, i.e., coupling radiation avoidance and capacity limitation: The step size is updated by using the PPO truncation objective function constraint strategy to ensure training stability in a highly dynamic nuclear radiation environment. ； In the formula, K represents the number of agents, and the probability ratio is... , The preset truncation hyperparameters, The policy entropy is used to encourage exploration of complex road networks; the larger the value, the higher the likelihood of the agent exploring in the early stages. It is the entropy coefficient; Critic network parameter updates: Minimize the mean squared error between the predicted value and the actual return: ； in This is a cumulative discount reward.

9. The multi-agent emergency evacuation method under dynamic radiation scenarios of nuclear accidents according to claim 8, characterized in that: S5 includes the determination and output of nuclear accident emergency evacuation decision-making schemes, with the following steps: The generated global evacuation strategy is quantitatively evaluated and determined, and a set of evaluation indicators is constructed. : Total evacuation time The time it takes for the last agent to arrive at the safe haven; Average cumulative radiation dose : ； like or an individual If the threshold specified by nuclear safety regulations is exceeded, an early warning is triggered, and the system returns to S1 to adjust parameters and reprogram. Refugee terminal utilization rate : To test the balance of multi-source and multi-sink synergy; ； In the formula, This represents the capacity of the j-th shelter that has been used; This represents the maximum capacity of the j-th refuge site.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 9.

Images