Multi-agent nuclear accident dynamic emergency evacuation method and storage medium

Abstract

The application discloses a multi-agent nuclear accident dynamic emergency evacuation method and a storage medium, and belongs to the technical field of nuclear accident emergency evacuation. In view of the problem that existing evacuation algorithms do not combine real city fine-grained traffic conditions in the nuclear proliferation scene, the application proposes a grid slicing and dynamic road condition mapping mechanism based on a map API. By obtaining the boundary coordinates of the target area to generate grid slicing, a rectangular query algorithm with frequency control is used to obtain branch-level real-time traffic data and time-series processing; the traffic data is superimposed with a nuclear radiation Gaussian plume space-time diffusion model to construct a radiation and traffic dual risk joint state matrix; based on the matrix, the MAPPO multi-agent reinforcement learning algorithm is used to train the agent through a multi-dimensional reward function to obtain an optimal collaborative evacuation plan that avoids congestion and high radiation areas. The application fills the technical gap in the real fine-grained road network access and multi-agent collaborative scheduling of nuclear emergency evacuation.

Description

Technical Field

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

[0002] Nuclear accident emergency evacuation is a crucial step in ensuring public safety and mitigating disaster losses. In emergencies such as nuclear leaks, evacuees face the dual high risks of the dynamic spatiotemporal spread of radiation plumes and traffic congestion caused by large-scale simultaneous evacuations. Therefore, developing scientific and efficient evacuation route planning methods is a core task of emergency decision-making systems.

[0003] Currently, most existing nuclear accident emergency evacuation route planning methods are based on traditional static road networks or closed environment models, which have significant limitations in practical applications. First, existing methods severely deviate from real-world urban traffic conditions when modeling the environment. Most algorithms calculate road traffic impedance only based on the superimposed density generated by the evacuated population itself, i.e., endogenous congestion; while ignoring the massive and complex background traffic flow within the urban road network during a sudden nuclear accident, i.e., exogenous congestion. This planning approach, detached from real dynamic road conditions, can easily lead to the system outputting a "theoretically optimal path" that, in reality, traverses arterial roads paralyzed by rush hour traffic, road construction, or sudden traffic accidents, resulting in large-scale delays for evacuees and significantly increasing the risk of prolonged, fatal exposure to highly radioactive contamination.

[0004] On the other hand, in order to overcome the shortcomings of being detached from actual road conditions, it is urgent to introduce real urban traffic data into the evacuation model. However, when actually obtaining large-scale urban road condition data, traditional methods usually directly request data for the entire administrative region from map service providers. This coarse approach is prone to triggering API call restrictions, leading to data acquisition interruptions. At the same time, the directly obtained data has a coarse granularity, often only covering main roads and unable to accurately reach the "capillary-level" branch road network on which urban evacuation is highly dependent.

[0005] Therefore, how to construct a stable, high-frequency, and fine-grained mechanism for acquiring real-time traffic conditions in real administrative regions, and how to integrate this time-stamped external real traffic flow data with multi-agent evacuation models in a time-series manner to solve the technical disconnect between the existing nuclear emergency evacuation system and the "real complex urban business implementation" process, is a technical bottleneck that urgently needs to be overcome in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic emergency evacuation method for multi-agent nuclear accidents. Addressing the weakness of existing evacuation algorithms that lack a realistic road network context, this invention designs a dynamic grid slicing algorithm based on boundary resolution and step-size traversal, combined with an anti-frequency control mechanism to stably collect real-time traffic data at the branch road level. By fusing timestamped traffic DataFrames with dynamic radiation fields, the state space and reward mechanism of the MAPPO algorithm are reconstructed, thereby achieving high survival rates and generating congestion-prevention collaborative evacuation strategies in complex real-world urban environments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent dynamic emergency evacuation method for nuclear accidents involves executing the following steps via computer equipment. S1. Call the map API to obtain the polygon coordinates of the boundary of the target administrative region, and generate a set of grid slices covering the entire region based on the extreme values ​​of latitude and longitude and the preset latitude and longitude step size, which serves as a unified spatial unit for road condition collection and environmental modeling. S2. Based on the set of grid slices generated by S1, a rectangular query algorithm with retry fault tolerance and random frequency control is used to obtain real-time traffic situation data at the branch level at high frequency through the map API, and encapsulate it into a time-series DataFrame sequence with timestamps. S3. Dynamically overlay the time-series traffic data obtained in S2 with the spatiotemporal diffusion model of nuclear radiation Gaussian plume to construct a joint environmental state matrix that integrates radiation risk and traffic risk, providing an observation basis for intelligent agents; S4. Using the joint environment state matrix constructed in S3 as input, the MAPPO multi-agent reinforcement learning algorithm is adopted, combined with a multi-dimensional comprehensive reward function to complete the agent training and form a collaborative evacuation decision model. S5. Using the decision-making model trained in S4, combined with the real-time updated joint environmental status, output the optimal collaborative evacuation plan to avoid high-radiation pollution areas and real road network congestion.

[0008] 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.

[0009] 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.

[0010] In summary, this invention discloses a multi-agent dynamic emergency evacuation method for nuclear accidents, belonging to the field of nuclear accident emergency evacuation. This method addresses the pain point of existing evacuation algorithms in high-risk nuclear proliferation scenarios, which are detached from the fine-grained traffic conditions of real cities. It innovatively proposes a grid slicing and dynamic traffic condition mapping mechanism based on the Gaode Map developer platform API. First, the polygon coordinates of the target administrative region boundary are obtained, and a set of grid slices covering the entire region is generated using a preset latitude and longitude step size. Further, a rectangular query algorithm with retry and random frequency control mechanisms is used to frequently retrieve real-time traffic situation data at the branch road level and encapsulate it into a timestamped DataFrame sequence. Then, this time-series real traffic data is superimposed with a nuclear radiation Gaussian plume spatiotemporal diffusion model to construct a joint environmental state matrix reflecting the dual risks of physical radiation and social traffic. Finally, based on this matrix, the MAPPO multi-agent reinforcement learning algorithm is driven, using a multi-dimensional comprehensive reward function to train the agents and output the optimal collaborative evacuation plan to avoid real-world traffic congestion and high-radiation pollution areas. This invention fills the technological gap in nuclear emergency evacuation regarding fine-grained road network data access and multi-agent collaborative scheduling in real, complex urban environments.

[0011] Specifically, the advantages of this invention are as follows: 1) Targeted Breakthrough: Addressing the unique high-risk scenario of nuclear accident emergency evacuation, characterized by the dual superposition of "dynamic diffusion of radiation plumes" and "complex traffic conditions in real urban environments," a dynamic evacuation environment model integrating real geographic information and traffic conditions was constructed, filling the technological gap in traditional evacuation algorithms that are detached from the real road network context. Innovatively, it utilizes the map developer platform API to extract polygonal boundaries of administrative regions and generates a set of grid tiles using latitude and longitude steps. Combined with setting a high-granularity road level parameter of level=4 or higher to extract "capillary-level" branch road conditions, it achieves high-precision acquisition of real microscopic physical spatial data under nuclear emergency scenarios.

[0012] 2) Collaborative innovation in method coupling: The MAPPO multi-agent reinforcement learning algorithm is deeply coupled with external asynchronous real traffic data streams. The high-frequency concurrency limitation of API is solved through a random delay backoff mechanism. The time-series engineering of DataFrame with timestamps is introduced to achieve physical alignment between multi-source grid slice data and reinforcement learning time steps. At the same time, the structured real road network congestion index and vehicle speed are embedded into the state tensor. Combined with the real traffic impedance reward function, the agent can implicitly extract and fit the coupling law between the "spatiotemporal evolution boundary of the radiation field" and the "time-varying tidal characteristics of road network traffic". This is more scientific and robust than traditional pathfinding algorithms that only rely on endogenous congestion assumptions.

[0013] 3) Improved Practicality of Evacuation Plans: This method provides an optimal evacuation strategy output based on "global multi-agent collaborative risk avoidance + adaptive detour based on real-world traffic conditions," overcoming the limitations of previous evacuation algorithms that were prone to failure during peak hours or sudden traffic accidents. This approach enables efficient scheduling of large-scale populations to avoid nuclear radiation while simultaneously anticipating and bypassing real-world urban congestion bottlenecks. It significantly improves the survival rate and feasibility of the strategy under sudden nuclear proliferation events in ultra-high-density, complex urban areas, providing a scientifically rigorous and precise decision-making tool for urban nuclear emergency command, intelligent traffic management, and public safety. Attached Figure Description

[0014] Figure 1 This is a flowchart of the present invention; Figure 2 This refers to the real-time traffic conditions of a certain road retrieved via API in an embodiment of the present invention. Figure 3 This is a trend chart of the road fit rate and success rate simulated by the MAPPO algorithm in an embodiment of the present invention; Figure 4 This is a reward graph for the MAPPO algorithm in an embodiment of the present invention. Detailed Implementation

[0015] 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.

[0016] like Figure 1 As shown in the figure, the multi-agent nuclear accident dynamic emergency evacuation method described in this embodiment executes the following steps through computer equipment: S1. Call the map API to obtain the polygon coordinates of the boundary of the target administrative region, and generate a set of grid slices covering the entire region based on the extreme values ​​of latitude and longitude and the preset latitude and longitude step size, which serves as a unified spatial unit for road condition collection and environmental modeling. S2. Based on the set of grid slices generated by S1, a rectangular query algorithm with retry fault tolerance and random frequency control is used to obtain real-time traffic situation data at the branch level at high frequency through the map API, and encapsulate it into a time-series DataFrame sequence with timestamps. S3. Dynamically overlay the time-series traffic data obtained in S2 with the spatiotemporal diffusion model of nuclear radiation Gaussian plume to construct a joint environmental state matrix that integrates radiation risk and traffic risk, providing an observation basis for intelligent agents; S4. Using the joint environment state matrix constructed in S3 as input, the MAPPO multi-agent reinforcement learning algorithm is adopted, combined with a multi-dimensional comprehensive reward function to complete the agent training and form a collaborative evacuation decision model. S5. Using the decision-making model trained in S4, combined with the real-time updated joint environmental status, output the optimal collaborative evacuation plan to avoid high-radiation pollution areas and real road network congestion.

[0017] The following is a detailed explanation: 1. Fine-grained real-time traffic data acquisition based on API slicing and frequency control mechanisms To address the complexity of real-world city road networks, such as those in urban areas, and the limitations of map API interfaces, a grid-tile-based branch-level traffic data retrieval module was designed. 1.1 Administrative Region Boundary Analysis and Step-Size Traversal Slicing Call the map developer platform API, such as the Amap API, and pass in the keyword of the target administrative region to obtain the polygonal outline boundary of that administrative region. Parse the outline string to extract the set of polygonal coordinate points representing the administrative region's boundaries; then calculate the extreme latitude and longitude values ​​of that boundary. and Set a fine-grained tile step size (e.g., 0.02°, approximately 2km), and use a double loop to traverse the latitude and longitude intervals to generate a set of rectangular grid tiles covering the target administrative region. .

[0018] 1.2 Branch-level traffic condition acquisition based on rectangular query and retry fault tolerance Iterate through the Grid collection of grid tiles, for each rectangular tile... The system invokes the traffic situation rectangle query API. To ensure fine-grained data, the query level is set to level=4 or higher, accurate to the branch road network. To cope with network fluctuations, a fault-tolerant retry mechanism is introduced during the traversal query process: a maximum number of retries is set for each slice, such as 3 times; if the request fails, it enters a short sleep and then retryes. Furthermore, to cope with API interface frequency control limitations, a random backoff mechanism is introduced: a random delay variable, such as 0.5~1.5 seconds, is added between adjacent slice requests to ensure efficient and stable retrieval of the full road network data. All road traffic information returned from all slices is then fully aggregated.

[0019] 1.3 Time-Series DataFrame Reconstruction and Persistence Furthermore, the aggregated traffic data for the entire road network is transformed into a structured DataFrame format. A time-series marker column is added to the dataset based on the current system time. Data collection time. The final data will be the actual cross-sectional data of urban road network congestion at this moment, and saved as a standardized file with a timestamp suffix, such as CSV format, to provide a standard time-series observation environment for subsequent algorithms.

[0020] 2. Spatiotemporal joint environment modeling that couples real-time traffic and dynamic radiation Furthermore, the real-world traffic flows obtained above are integrated with the laws governing nuclear diffusion in natural physics.

[0021] 2.1 Spatiotemporal Evolution of Dynamic Nuclear Diffusion A Gaussian plume diffusion mathematical model is introduced to simulate the evolution of leaked materials from a nuclear accident over time, influenced by wind direction, wind speed, and spatiotemporal movement. Let the grid coordinates be [value missing] at time t. The radiation concentration at that location is .

[0022] 2.2 Construction of Real Traffic-Radiation Joint Feature State Matrix By fusing real-time traffic DataFrame data with timestamps and dynamic radiation fields, a spatiotemporal joint environment matrix at time t is constructed. Each grid node's eigenvector in the matrix contains: ; In the formula, and These are the actual road congestion status and average traffic speed extracted from the DataFrame; This represents the micro-density caused by the dispersal of the crowd. This matrix is ​​updated continuously with the system's time step.

[0023] 3. MAPPO Multi-Agent Path Planning Based on Real-World Road Condition Perception In the aforementioned highly realistic spatiotemporal environment, the MAPPO algorithm is used to achieve the global optimal cooperative evacuation of multiple agents.

[0024] 3.1 Observation space design integrating API traffic conditions Suppose the system contains N agents. At time t, the local observation state of the i-th agent is... Extracted from the environment matrix Its observation field of view breaks through traditional limitations, including real road condition slices of the candidate road network ahead: ; In the formula, Let be the two-dimensional coordinates of the grid node where the i-th agent is located at time t; For an intelligent agent within its own local field of vision Within, a slice of the dynamic spatiotemporal joint environment matrix is ​​perceived; This refers to the distance or relative spatial orientation of the current location from various shelters. This is the actual congestion index vector of the road segment ahead within the agent's field of vision.

[0025] 3.2 Design of Reward Function under Traffic-Radiation Dual Constraints Design a multidimensional comprehensive reward function To resolve the conflict between "radiation avoidance" and "real traffic congestion avoidance": ; In the formula, , which are preset weighting coefficients; Positive rewards for guiding people to the safe zone; As punishment for exposure to radiation fields; To address collision and crowding penalties among multiple agents, a realistic traffic impedance penalty is introduced. When an agent attempts to drive into a traffic area marked as "severely congested" by the API and its actual speed... In extremely low-speed sections, the system applies a high penalty value based on the data in the DataFrame, forcing it to abandon the theoretical shortest path and choose a real, unobstructed detour route.

[0026] 3.3 MAPPO Algorithm Co-training Network Update Actor Network: Each agent has its own independent Actor Network. The probability distribution of output actions is based on observations that include real road conditions.

[0027] Critic Network: A centralized Critic Network receives global status data including the entire region's API traffic situation. Regarding the current state The global value function is evaluated.

[0028] By combining the truncated PPO objective function with parameter updates, the agent learns to predict the coupling pattern between the radial drift trajectory and the real urban traffic tidal characteristics.

[0029] 4. Generation of evacuation decision-making plans and output of dynamic contingency plans Furthermore, the well-trained Actor model is integrated into the urban emergency command system. During the actual inference phase, a timed task continuously retrieves the latest traffic DataFrame, triggering the Actor network to perform forward inference on the current environment. The final output is a dynamic set of optimal evacuation trajectories for the global multi-agent network, along with traffic flow predictions and shelter reception scheduling tables for each time period, providing decision support for the emergency command center.

[0030] like Figure 2 This is a real-time traffic data table for a certain urban area. It records the real-time traffic data for a specific road segment at 16:21:13 on December 29, 2025. The table includes information such as road name, traffic status, driving direction, road angle, average speed, road segment code, road coordinate string, and collection time. The status code 1 represents smooth traffic and 2 represents slow traffic / congestion. It can be used as a basis for route planning decisions.

[0031] Figure 3 This is a trend chart of the road fit rate and success rate simulated by the MAPPO algorithm in an embodiment of the present invention; Figure 3 The yellow curve represents the agent's average road fit rate. Initially, the value of the yellow curve is very low, indicating that the agent initially explores recklessly and does not fully follow the road. This is extremely unreasonable in actual vehicle driving. Therefore, under the setting of the reward environment, the agent quickly learns the rule of exploring along the road, and the road fit rate stabilizes at 100%.

[0032] The green curve represents the success rate of the agent reaching the endpoint. The green curve in the figure generally tends to converge, eventually converging at nearly 100%. During the upward process, there are several points where the success rate of the agent drops sharply. This is because the entropy of the agent's exploration settings is relatively high. In the early stages, the agent changes its strategy too quickly, and the agent chooses a less effective strategy, causing the success rate to drop sharply. Later, the agent self-corrects during the learning process, which solves this problem well.

[0033] Figure 4 It represents the trend of the agent's total reward value. During training, the agent will evolve in the direction of maximizing the reward value.

[0034] In the early stages (0-5000 rounds): the reward increases rapidly from near 0, indicating that the agent is learning quickly and the strategy is converging from random exploration to effective behavior.

[0035] Mid-term (5000-15000 rounds): Rewards remain at a high level of 15000-20000, but fluctuate greatly. It is a stage of balance between exploration and utilization, and occasional strategic fluctuations may occur.

[0036] Later stages (rounds 15,000-35,000): Rewards remain stable in the range of 10,000-15,000.

[0037] 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.

[0038] 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.

[0039] 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.

[0040] 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)).

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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.

[0042] 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.

[0043] 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 dynamic emergency evacuation method for multi-agent nuclear accidents, characterized in that, Includes the following steps, S1. Call the map API to obtain the polygon coordinates of the boundary of the target administrative region, and generate a set of grid slices covering the entire region based on the extreme values ​​of latitude and longitude and the preset latitude and longitude step size, which serves as a unified spatial unit for road condition collection and environmental modeling. S2. Based on the set of grid slices generated by S1, a rectangular query algorithm with retry fault tolerance and random frequency control is used to obtain real-time traffic situation data at the branch level at high frequency through the map API, and encapsulate it into a time-series DataFrame sequence with timestamps. S3. Dynamically overlay the time-series traffic data obtained in S2 with the spatiotemporal diffusion model of nuclear radiation Gaussian plume to construct a joint environmental state matrix that integrates radiation risk and traffic risk, providing an observation basis for intelligent agents; S4. Using the joint environment state matrix constructed in S3 as input, the MAPPO multi-agent reinforcement learning algorithm is adopted, combined with a multi-dimensional comprehensive reward function to complete the agent training and form a collaborative evacuation decision model. S5. Using the decision-making model trained in S4, combined with the real-time updated joint environmental status, output the optimal collaborative evacuation plan to avoid high-radiation pollution areas and real road network congestion.

2. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 1, characterized in that: S1 includes, The map developer platform API is invoked, passing in the keyword of the target administrative region to obtain the polygonal outline boundary of that administrative region. By parsing the outline string, the set of polygonal coordinate points of the administrative region boundary is extracted; the latitude and longitude extreme values ​​of the boundary are then calculated. and Set the fine-grained tiling step size (step), and use a double loop to traverse the latitude and longitude intervals to generate a set of rectangular grid tiles covering the target administrative region. .

3. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 2, characterized in that: S2 employs a rectangular query algorithm with retry fault tolerance and random frequency control to frequently acquire real-time traffic situation data at the branch road level through the map API, specifically including: Iterate through the Grid collection of grid tiles, for each rectangular tile... Call the traffic situation rectangle query API; Set the query level to level=4 or higher, accurate to the branch road network; A fault-tolerant retry mechanism is introduced during the traversal query process: a maximum number of retries is set for each slice, and if the request fails, it will enter a short sleep and then retry. Introduce a random backoff mechanism: simultaneously add a random delay variable between adjacent slice requests; The traffic information for roads returned from all slices is fully summarized.

4. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 3, characterized in that: In S2, it is encapsulated as a time-stamped sequence of DataFrames, including, The aggregated traffic data for the entire road network is converted into a structured DataFrame format, and a time-series marker column is added to the dataset based on the current system time. The data collection time is used to generate real-time traffic congestion cross-section data of the urban road network at that moment, and saves it as a standardized file with a timestamp suffix, providing a standard time-series observation environment for subsequent algorithms.

5. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 4, characterized in that: S3 specifically includes integrating the real social activity traffic flow obtained above with the laws of natural physical nuclear diffusion; Dynamic nuclear diffusion spatiotemporal evolution: A Gaussian plume diffusion mathematical model is introduced to simulate the evolution of leaked materials from a nuclear accident over time and with varying wind direction, wind speed, and spatiotemporal movement; Let time t be the grid coordinates... The radiation concentration at that location is ; Construction of Real Traffic-Radiation Joint Feature State Matrix: Fusing real-time traffic DataFrame data with timestamps with dynamic radiation fields to construct a spatiotemporal joint environment matrix at time t. Each grid node feature vector in the matrix contains: ； In the formula, and These are the actual road congestion status and average traffic speed extracted from the DataFrame; The matrix is ​​updated in a rolling fashion to account for the micro-density caused by the evacuation of the crowd itself, along with the system time step.

6. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 5, characterized in that: S4 specifically includes the design of an observation space that integrates API road conditions, and the steps are as follows. Suppose the system contains N agents. At time t, what is the local observation state of the i-th agent? Extracted from the environment matrix Its observation field of view breaks through traditional limitations, including real road condition slices of the candidate road network ahead: ； In the formula, Let be the two-dimensional coordinates of the grid node where the i-th agent is located at time t; For an intelligent agent within its own local field of vision Within, a slice of the dynamic spatiotemporal joint environment matrix is ​​perceived; This refers to the distance or relative spatial orientation of the current location from various shelters. This is the actual congestion index vector of the road segment ahead within the agent's field of vision.

7. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 6, characterized in that: S4 also includes the design of a reward function with traffic-radiation dual constraints, the steps of which are as follows. Design a multidimensional comprehensive reward function To resolve the conflict between "radiation avoidance" and "real traffic congestion avoidance": ； In the formula, , which are preset weighting coefficients; Positive rewards for guiding people to the safe zone; As punishment for exposure to radiation fields; Implement collision and crowding penalties for multiple agents; introduce realistic traffic impedance penalties. When an agent attempts to enter a traffic jam marked as "severely congested" by the API and its actual speed... In extremely low-speed sections, the system applies a high penalty value based on the data in the DataFrame, forcing it to abandon the theoretical shortest path and choose a real, unobstructed detour route.

8. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 7, characterized in that: S4 also includes the 3MAPPO algorithm for collaborative training and network updates, with the following steps. Actor Network: Each agent has its own independent Actor Network. Based on the probability distribution of observed actions that include real road conditions; Critic Network: A centralized Critic Network receives global status data including the entire region's API traffic situation. Regarding the current state The global value function is used for evaluation; By combining the truncated PPO objective function with parameter updates, the agent learns to predict the coupling pattern between the radial drift trajectory and the real urban traffic tidal characteristics.

9. The multi-agent nuclear accident dynamic emergency evacuation method according to claim 8, characterized in that: Specifically, S5 involves using a timed task to continuously retrieve the latest traffic DataFrame, triggering the Actor network to perform forward reasoning on the current environment, and finally outputting a dynamic global multi-agent optimal evacuation trajectory set. It also generates road network traffic flow prediction and shelter reception scheduling tables for each time period, providing decision support for the emergency command center.

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