Dynamic evacuation simulation method for subway station flood disaster based on multi-dimensional coupling

Abstract

The application relates to a kind of subway station flood disaster dynamic evacuation simulation methods based on multidimensional coupling. Its space discretization is completed by color semantics recognition, and the full-field real-time water level is obtained by shallow water wave diffusion calculation. Based on the population statistics parameters, heterogeneous agents with differentiated movement ability and social relationship are generated. Water depth resistance, density friction and kinship synergy gain are integrated into multi-field coupled instantaneous speed estimation. Through facility failure determination and congestion driven Bernoulli game mechanism, multi-level dynamic re-planning of the path is realized. Finally, statistical indicators and heat maps are output by motion residual accumulation and survival state determination. At the same time, aiming at the action obstacle group, auxiliary dependent pairing, rigid joint speed fusion and facility exclusive dependence labeling are introduced. The physical support constraint and single point failure risk of special groups are preposed and made explicit, so that the simulation can truly depict the behavior degradation and path obstruction characteristics of this group in flood disaster.

Description

Technical Field

[0001] This invention belongs to the field of computer simulation technology, specifically relating to a simulation method for dynamic evacuation of subway stations during floods based on multidimensional coupling. Background Technology

[0002] The core challenge in simulating flood evacuation in subway stations lies in the strong spatiotemporal coupling between hydrodynamic field evolution and crowd behavior decisions. On the one hand, continuously rising water levels significantly weaken pedestrians' ability to wade through water through nonlinear blocking effects, potentially causing critical facilities to fail and blocking existing escape routes. On the other hand, panic among the crowd due to perceived water level threats and passageway congestion triggers spontaneous path-based games and diversion behaviors, thereby altering local density distribution and inversely affecting the effective passage cross-section of the water flow spread path. Existing evacuation simulation methods mostly treat hydrodynamic simulation and crowd movement modeling as independent modules, lacking a stepwise long-term real-time interaction mechanism between the water level field, density field, and individual decision field within a unified cellular automata framework. This results in simulation results that fail to reflect the true dynamic characteristics of the cascade feedback chain of rising water—road blockage—people blocking—rerouting in a disaster environment. In wading environments, the challenge is not a gradual slowdown, but a drastic, nonlinear collapse of propulsion capacity. Furthermore, there is an inseparable, rigid kinematic coupling constraint between the individual and their companions, as well as an exclusive path dependence on scarce vertical access routes such as accessible elevators. Existing methods fail to incorporate this rigid physical binding and single-point failure path dependence into the coupled calculations, resulting in simulation systems providing optimistic predictions of the escape rates for certain groups that are significantly detached from physical reality. This conceals deep-seated regional congestion bottlenecks and deadly risk blind spots.

[0003] Therefore, an optimized simulation scheme for dynamic evacuation during subway floods is desired. Summary of the Invention

[0004] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a simulation method for dynamic evacuation during subway station flooding based on multidimensional coupling, comprising: Step 1: Based on the color semantic recognition rules of the subway station geometric model, the subway station space is discretized into cellular meshes and the regional attributes are labeled. The shallow water wave diffusion calculation under dynamic topological constraints is performed on the external water injection flux to obtain the cellular simulation environment and global static distance potential field that carry the real-time water level of each mesh. Step 2: Based on demographic parameters, differentiated assignment of baseline speed and risk tolerance and social relationship mapping are performed on the heterogeneous agent set containing multiple age groups to obtain a set of family units with bidirectional kinship index links. Step 3: Based on the real-time water level and local population density of each grid, perform multi-field coupled rate estimation for each individual in the heterogeneous intelligent agent set using nonlinear water depth hindrance reduction and crowding friction correction, and combine the member spacing of the family unit set to trigger dual-mode switching gain to obtain the instantaneous movement rate of each intelligent agent. Step 4: Based on the comparison results between the real-time water level of each grid and the critical water level for facility damage, the failure status of key facilities is determined, and multi-level dynamic game decision-making is carried out on the heterogeneous intelligent agent set according to the local congestion density to obtain the updated path target of each intelligent agent. Step 5: Based on the instantaneous movement rate and simulation step size, perform discrete coordinate jump driven by the cumulative motion residual amount and survival status determination triggered by water level exceeding limit for each agent. Then, serialize the coordinate sequence and update path target of each agent into a structured time-series data frame to obtain the batch statistical indicators and spatiotemporal trajectory heat map of the evacuation simulation.

[0005] Compared with existing technologies, this invention proposes a dynamic evacuation simulation method for subway station flood disasters based on multi-dimensional coupling. First, within a unified cellular automata framework, the grid discretization of the subway station space and shallow water wave diffusion calculations are performed to establish a progressively long, dynamically updated water level field environment. Then, a heterogeneous set of intelligent agents is generated, incorporating differentiated motion parameters of multiple age groups and family relationships. For individuals with mobility impairments, auxiliary dependency pairing, rigid association velocity fusion, and facility exclusive path dependency labeling mechanisms are introduced. Based on this, multi-field coupling rate estimation is performed on the behavioral gains triggered by water depth hindrance, crowd density friction, and family member spacing. Simultaneously, individuals are driven to perform multi-level dynamic game-based path decisions based on facility failure states and local congestion density. Finally, through discrete coordinate jumps based on accumulated motion residuals and survival status determination when water levels exceed limits, batch statistical indicators and spatiotemporal trajectory heatmaps of the evacuation process are output, thereby achieving integrated dynamic simulation and deduction of the risks of road closures due to rising water, crowd diversion, and the stranding of special groups. Attached Figure Description

[0006] Figure 1 This is a flowchart of a dynamic evacuation simulation method for subway station flood disasters based on multidimensional coupling, according to an embodiment of this application. Figure 2 This is a data flow diagram of a dynamic evacuation simulation method for subway station flood disasters based on multidimensional coupling, according to an embodiment of this application. Figure 3 This is a simplified schematic diagram of the multidimensional coupled cellular environment layered architecture of the dynamic evacuation simulation method for subway station flood disasters based on multidimensional coupling according to an embodiment of this application. Figure 4The flowchart illustrates the process of using demographic parameters to differentiate the baseline velocity and risk tolerance of a heterogeneous set of intelligent agents containing multiple age groups and map their social relationships to obtain a set of family units with bidirectional kinship index links, based on demographic parameters in the dynamic evacuation simulation method for subway station flood disasters based on multidimensional coupling according to embodiments of this application. Figure 5 This is a flowchart illustrating the process of using a multidimensional coupling-based dynamic evacuation simulation method for subway station flood disasters, according to an embodiment of this application, to perform bidirectional kinship index linking and pairing between vulnerable individuals and adult individuals in a heterogeneous intelligent agent set to obtain a family unit set; Figure 6 Based on the comparison results between the real-time water level of each grid and the critical water level for facility damage in the dynamic evacuation simulation method for subway flood disaster based on multidimensional coupling according to the embodiments of this application, the failure state of key facilities is determined, and multi-level dynamic game decision is made on the heterogeneous intelligent agent set according to the local congestion density to obtain the flowchart of the updated path target of each intelligent agent. Figure 7 To determine the survival status of each agent based on the instantaneous movement rate and simulation step size in the dynamic evacuation simulation method for subway flood disasters based on multidimensional coupling according to the embodiments of this application, the method performs discrete coordinate jump driven by the accumulation of motion residuals and the water level exceeding the limit. The method also serializes the coordinate sequence and update path target of each agent into a structured time-series data frame to obtain the flowchart of batch statistical indicators and spatiotemporal trajectory heat map of evacuation simulation. Figure 8 This is a visualized heat map of pedestrian evacuation trajectories based on a multidimensional coupling-based dynamic evacuation simulation method for subway station flood disasters, according to an embodiment of this application. Detailed Implementation

[0007] 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, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0008] Figure 1 This is a flowchart of a dynamic evacuation simulation method for subway station flood disasters based on multidimensional coupling, according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating the dynamic evacuation simulation method for subway station flood disasters based on multidimensional coupling, according to an embodiment of this application. Figure 1 and Figure 2As shown, the simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to an embodiment of this application includes the following steps: S1, based on the color semantic recognition rules of the subway station geometric model, the subway station space is discretized into cellular grids and regional attribute labeled, and shallow water wave diffusion calculation under dynamic topological constraints is performed on the external water injection flux to obtain the cellular simulation environment and global static distance potential field carrying the real-time water level of each grid; S2, based on demographic parameters, the heterogeneous intelligent agent set containing multiple age groups is differentially assigned benchmark speed and risk tolerance and social relationship mapping is performed to obtain a family unit set with bidirectional kinship index links; S3, based on the real-time water level of each grid and local population density, nonlinear simulation is performed on each individual in the heterogeneous intelligent agent set. Multi-field coupled rate estimation with water depth resistance reduction and congestion friction correction, combined with bimodal switching gain triggered by member spacing of the family unit set, is used to obtain the instantaneous movement rate of each agent; S4, based on the comparison results of the real-time water level of each grid and the critical water level for facility damage, failure status judgment is performed on key facilities, and multi-level dynamic game decision is performed on the heterogeneous agent set according to the local congestion density to obtain the updated path target of each agent; S5, based on the instantaneous movement rate and simulation step size, discrete coordinate jump driven by the accumulation of motion residuals and survival status judgment triggered by water level exceeding limit are performed on each agent, and the coordinate sequence and updated path target of each agent are serialized into structured time-series data frames to obtain batch statistical indicators and spatiotemporal trajectory heatmaps of evacuation simulation.

[0009] Specifically, in step S1, based on the color semantic recognition rules of the subway station geometric model, the subway station space is discretized into cellular meshes and its regional attributes are labeled. Then, a shallow water wave diffusion calculation under dynamic topological constraints is performed on the external water injection flux to obtain the cellular simulation environment and global static distance potential field that bear the real-time water level of each mesh. It should be noted that, as an underground enclosed space, the subway station contains complex structural elements such as platform levels, concourse levels, turnstiles, staircases, and multiple entrances and exits. Directly performing joint calculations of water flow spread and crowd movement on a continuous geometric space would face problems such as complex boundary condition handling and excessive computational overhead, and would make it difficult to achieve gradual and long-term synchronous updates of the water level field and crowd positions. Therefore, the technical solution of this application first discretizes the subway station space into cellular meshes and labels its regional attributes based on the color semantic recognition rules of the subway station geometric model, and then performs a shallow water wave diffusion calculation under dynamic topological constraints on the external water injection flux to obtain the cellular simulation environment and global static distance potential field that bear the real-time water level of each mesh. Through the above processing, the continuous space can be transformed into a structured discrete grid representation, enabling subsequent hydrodynamic evolution and agent pathfinding to complete progressively longer real-time interactions within a unified cellular framework.

[0010] More specifically, in a concrete example of this application, the pixel drawing data of the subway station geometric model is first spatially discretized. Taking a simplified 2D CAD drawing of a subway station as the object, the entire pixel drawing is divided into equal-spaced sections along the horizontal and vertical directions according to a preset physical size ratio. Each division unit corresponds to a 1-meter by 1-meter square area in the actual physical space, thus mapping the continuous 2D plane into a discretized grid matrix consisting of several rows and columns. After completing the grid division, the pixel color values ​​of the areas covered by each grid in the discretized grid matrix are extracted and matched one by one with preset color semantic intervals. In the drawing specifications of this paper, different color intervals are assigned clear physical semantics. For example, black pixel intervals correspond to impassable obstacles such as walls or columns, white pixel intervals correspond to open passage areas of the concourse or platform, green pixel intervals correspond to exit locations, yellow pixel intervals correspond to turnstile facilities areas, and blue pixel intervals correspond to staircases or escalator areas. Each grid is labeled with the corresponding area attribute category according to the semantic interval into which its pixel color value falls. The attribute labeling results of all grids are summarized to form an area attribute label set.

[0011] After obtaining the discretized grid matrix and the set of region attribute labels, a global static distance potential field is constructed. All grids marked as exits are selected from the region attribute label set, their distance potential values ​​are initialized to zero, and they are added to the traversal queue as starting nodes for breadth-first search. Subsequently, the traversal of the discretized grid matrix is ​​performed layer by layer, with each layer incremented by a unit of Manhattan distance weight on the potential value of the current grid relative to its parent node, while skipping grids marked as walls or other impassable structures. At the end of the traversal, each traversable grid in the discretized grid matrix is ​​assigned a scalar potential value, representing the shortest topological distance from that grid along a traversable path to the nearest exit. This forms a global static distance potential field covering the entire field, providing gradient direction references for subsequent navigation and pathfinding by the agent.

[0012] While constructing the static field, the hydrodynamic field is initialized and its gradual long-term evolution is calculated. External water flux is injected into a pre-defined flood source grid in the discretized grid matrix. This source grid corresponds to physical locations where rainwater backflow may actually occur, such as subway station entrances and exits. Within each simulation step, for each grid in the discretized grid matrix, its topological connectivity with its four adjacent von Neumann neighborhood grids is determined based on the region attribute label set. If there are insurmountable obstacles such as walls in the neighborhood direction, the dynamic topological connectivity factor in that direction is set to zero to block water flow; otherwise, it is set to one to allow water exchange. Based on the local flux balance mechanism, the water level update of each grid follows the following calculation rules: in, The water level will be updated in the next time step. The water level at the current moment. The external water injection flux is the injection component of this grid and is greater than zero only at the source grid. The fluid diffusion coefficient is used to characterize the physical correction of the ground material for the shallow water wave propagation rate. This is the set of four-way von Neumann neighborhoods of the grid. It is the dynamic topological connectivity factor. The water level of the neighboring grid at the current moment is represented by the above formula. Parallel water level updates are performed on all grids within each simulation step. The water flow naturally generates flow around and splits at the boundary of the obstacle. Finally, the cellular simulation environment carrying the real-time water level of each grid is output, providing progressively long synchronously updated hydrodynamic field data for the subsequent rate reduction and survival status determination of the agent.

[0013] Figure 3 This is a simplified schematic diagram of a multidimensional coupled cellular environment layered architecture according to an embodiment of this application. Figure 3 As shown, the cellular simulation environment consists of four functional layers stacked vertically from bottom to top. The bottom layer is the real physics layer, which carries the spatial layout information of the physical structures of the subway station, such as walls, turnstiles, stairs, and safety exits. Above this is the hydrodynamic evolution layer, which carries the real-time water level data of each grid and performs shallow water wave diffusion calculations based on the local flux balance mechanism within each simulation step. The arrows in the figure indicate the direction of water flow diffusion in this layer. The next layer is the Agent action layer (i.e., the intelligent agent action layer), which carries the spatial position and motion state of each agent in the heterogeneous agent set. The top layer is the potential energy field navigation layer, which stores the global static distance potential energy field generated based on breadth-first search, providing the agents with gradient navigation directions pointing to the nearest exit. Within each simulation step, each agent simultaneously senses the water level information of the lower hydrodynamic evolution layer and the gradient direction of the upper potential energy field navigation layer, and performs discrete coordinate jumps after completing multi-field coupling rate calculations and path decisions.

[0014] Specifically, in step S2, based on demographic parameters, a heterogeneous set of intelligent agents comprising multiple age groups is differentiated in terms of baseline speed and risk tolerance, and mapped to social relationships to obtain a set of family units with bidirectional kinship index links. It should be noted that, given the significant individual differences in age composition, physical fitness level, and water tolerance among passengers stranded at subway stations during floods, and the fact that children and the elderly have much lower walking speed and resistance to water flow than young adults, and that in real evacuation scenarios there are social behavioral constraints such as mutual waiting and regrouping between vulnerable individuals and their adult relatives, uniformly modeling all individuals using homogenized parameters would prevent simulation results from reflecting the differentiation in mobility among different groups under water-resistant conditions and the impact of kinship collaboration on evacuation efficiency. Therefore, the technical solution of this application further differentiates the baseline speed and risk tolerance of the heterogeneous set of intelligent agents comprising multiple age groups based on demographic parameters, and maps to social relationships to obtain a set of family units with bidirectional kinship index links. Through the above processing, it is possible to construct various types of intelligent agents with different mobility and risk tolerance thresholds in the simulation environment. The kinship index relationship of the family unit provides the member spacing trigger basis for the bimodal switching gain calculation in the subsequent steps, enabling the evacuation simulation to capture real behavioral characteristics such as the accelerated reunion of guardians when relatives are separated and the risk tolerance of the ward being regulated by social relationships.

[0015] Figure 4 This document describes a flowchart illustrating the process of applying demographic parameters to differentiate the baseline velocity and risk tolerance of a heterogeneous set of intelligent agents encompassing multiple age groups, and mapping their social relationships to obtain a set of family units with bidirectional kinship index links, based on demographic parameters in a multidimensional coupled dynamic evacuation simulation method for subway station floods according to embodiments of this application. Figure 4 As shown, step S2 includes: S21, randomly sampling and assigning group labels to the total number of simulated individuals in the demographic parameters according to a preset age group ratio to generate an attribute-distributed agent set; S22, based on a Gaussian perturbation superposition model, assigning normal perturbation values ​​of baseline velocity and risk tolerance to each individual in the attribute-distributed agent set according to the statistical expectation value corresponding to its group label to obtain a heterogeneous agent set; S23, pairing vulnerable individuals and adult individuals in the heterogeneous agent set with a kinship bidirectional index link to obtain a family unit set.

[0016] In step S21, the total number of simulated individuals in the demographic parameters is randomly sampled and assigned group labels according to a preset age group ratio to generate an attribute-distributed agent set. It should be noted that since subway station passengers during peak hours include children, teenagers, adults, and the elderly, each group exhibits differences in movement ability and risk tolerance parameters in subsequent simulations. Therefore, it is necessary to first clarify the age category to which each agent belongs before assigning it physical attributes that match the group characteristics. Based on this, the technical solution of this application further randomly samples and assigns group labels to the total number of simulated individuals in the demographic parameters according to a preset age group ratio to generate an attribute-distributed agent set. Through the above processing, a clear group affiliation basis can be provided for the subsequent differentiated assignment of baseline speed and risk tolerance.

[0017] More specifically, in a concrete example of this application, the total number of simulated individuals and the preset distribution ratios of each age group are first read from demographic parameters obtained based on historical security surveillance video analysis or ticket gate card swiping data. For example, adults account for 50%, the elderly for 20%, children for 15%, and teenagers for 15%. Then, random sampling is performed on the total number of simulated individuals based on a multinomial distribution model. That is, for each individual to be generated, a random sample is independently performed using the above ratios as the sampling probability for each category, and the sampling result is used as the age group label for that individual. After all individuals have been labeled, the actual number of individuals in each age category statistically approaches the target number specified by the preset ratio. All individuals carrying age group labels are encapsulated to form an attribute-distributed agent set. Each individual in this set has a clear group affiliation identifier but has not yet been assigned specific motion parameters, serving as the input object for the next sub-step of differential attribute assignment.

[0018] In step S22, based on the Gaussian perturbation superposition model, each individual in the attribute-distributed agent set is assigned a normal perturbation value of baseline speed and risk tolerance according to the statistical expected value corresponding to its group label, to obtain a heterogeneous agent set. It should be noted that since different individuals within the same age group are not entirely consistent in terms of physical condition, physical status, and tolerance to water flow impact, assigning the same speed and tolerance parameters to all individuals under the same group label would result in a large number of individuals with completely homogeneous motion characteristics in the simulation, failing to reflect the individual differences within a real population. Therefore, the technical solution of this application further assigns a normal perturbation value of baseline speed and risk tolerance to each individual in the attribute-distributed agent set according to the statistical expected value corresponding to its group label, based on the Gaussian perturbation superposition model, to obtain a heterogeneous agent set. Through the above processing, individuals within the same group can exhibit reasonable parameter dispersion while maintaining overall statistical consistency, providing physically realistic individualized input for subsequent multi-field coupling rate calculation and survival status determination.

[0019] More specifically, in a concrete example of this application, each individual in the attribute distribution agent set is traversed, its age group label assigned in the previous sub-step is read, and the pre-defined statistical expected speed and expected risk tolerance of that group are retrieved accordingly. For the assignment of the baseline speed, the expected speed of the group is used as the mean, and the corresponding speed standard deviation is used as the dispersion amplitude, according to... Perform calculations, where For the first The average expected speed of the age group to which each agent belongs The standard deviation of the velocity of this group. This refers to an independent random perturbation factor that follows a standard normal distribution. For example, if the expected speed of an adult population is set at 1.2 meters per second with a standard deviation of 0.2 meters per second, then the perturbed baseline speed of individuals in this population will be normally distributed around 1.2 meters per second. Conversely, if the expected speed of an elderly population is set at 0.7 meters per second with a standard deviation of 0.15 meters per second, their individual speeds will be distributed around the lower mean. The same normal perturbation mechanism is used for assigning risk tolerance, following... Perform calculations, where The average risk tolerance depth for this group. The corresponding standard deviation is... This is another independent standard normal random perturbation factor, which characterizes the critical water depth threshold that triggers a trapped state for an individual while wading. After all individuals have completed the perturbation assignment of baseline speed and risk tolerance, they are updated and iterated into a heterogeneous agent set. In this set, each agent has its own independent motion capability parameters and risk tolerance threshold while retaining its group affiliation label.

[0020] In step S23, a bidirectional kinship index link is used to pair vulnerable individuals and adult individuals in the heterogeneous intelligent agent set to obtain a family unit set. It should be noted that, given the social behavioral constraints in real subway station evacuation scenarios where vulnerable individuals such as children and the elderly often maintain close proximity and wait with their adult relatives when facing floodwaters, adult guardians will actively accelerate reunification when they are separated due to water flow obstruction or crowding. The psychological resilience of the ward also changes due to the presence of relatives. This collaborative behavior among relatives has a direct impact on local evacuation efficiency and individual survival status. Based on this, the technical solution of this application further uses a bidirectional kinship index link to pair vulnerable individuals and adult individuals in the heterogeneous intelligent agent set to obtain a family unit set. Through the above processing, a clear guardian-ward pairing relationship can be provided for the instantaneous rate calculation of the bimodal switching gain triggered by member spacing in subsequent steps, enabling the simulation process to capture the accelerated reunification behavior when relatives are separated and the dynamic characteristics of the ward's risk tolerance being regulated by social relationships.

[0021] Figure 5 This is a flowchart illustrating the process of using a multidimensional coupling-based dynamic evacuation simulation method for subway station flood disasters, according to an embodiment of this application, to perform bidirectional kinship index linking and pairing between vulnerable individuals and adult individuals in a heterogeneous intelligent agent set to obtain a family unit set. For example... Figure 5 As shown, step S23 includes: S231, based on preset action obstacle labeling rules, identifying obstacle attributes and screening adult candidates for the heterogeneous intelligent agent set to generate an auxiliary pairing set; S232, performing optimal matching calculation and bidirectional binding on the auxiliary pairing set to obtain a joint parameter set; S233, based on the caregiver gain coefficient and risk transmission coefficient, performing joint velocity fusion and joint tolerance correction on the joint parameter set to obtain an enhanced intelligent agent set; S234, performing facility dependency labeling and risk weight estimation on the enhanced intelligent agent set to obtain a family unit set.

[0022] In step S231, based on preset mobility impairment labeling rules, the heterogeneous intelligent agent set is subjected to impairment attribute identification and adult candidate screening to generate an auxiliary pairing set. It should be noted that, given that there are inevitably some mobility-impaired passengers in real subway stations, the decline in mobility of these individuals in wading environments is not a gradual slowdown but rather faces a nonlinear collapse of propulsion force. Furthermore, they and their accompanying caregivers constitute a rigid coupling constraint of physical support and propulsion. Therefore, it is necessary to accurately identify these individuals from the heterogeneous intelligent agent set and match them with adult candidates possessing sufficient assistance capabilities. Based on this, the technical solution of this application further uses preset mobility impairment labeling rules to identify impairment attributes and adult candidates in the heterogeneous intelligent agent set to generate an auxiliary pairing set. Through the above processing, individuals with mobility impairments can be independently separated from the general vulnerable group, and the optimal caregiver can be identified by comprehensively evaluating the mobility redundancy and spatial proximity of the candidates, providing a solid pairing relationship foundation for subsequent rigid joint velocity fusion and facility dependency labeling.

[0023] More specifically, in a concrete example of this application, the entire set of heterogeneous agents is first traversed. Each individual's attribute fields are checked one by one according to preset movement impairment labeling rules. Individuals with specific labels such as wheelchair dependence, visual impairment dependence, or cane dependence are identified as disabled agents. Then, for each identified disabled agent, using its current grid coordinates as the center and a preset assistance binding radius as the search range, adult individuals not yet assigned to any family unit or assistance pairing relationship are retrieved from the heterogeneous agent set. Adult individuals meeting the spatial constraints are included in the disabled agent's companion candidate list. Since simple spatial distance cannot reflect the physical cost of rescue, a fitness function is introduced to comprehensively score each pair of disabled agents and adult candidate individuals in the candidate list. This score considers both the candidate's speed redundancy relative to the disabled person and the spatial proximity between them. The fitness score is calculated according to the following formula: in, For the intelligent agent of the obstacle With caregiver candidates The overall compatibility score between the pairs For caregiver candidates The base speed, For the intelligent agent of the obstacle Extremely low reference speed under normal conditions This represents the initial Euclidean distance when a match is determined between the two. The preset auxiliary binding radius is the extreme value. The baseline velocity ratio factor at the front end of the formula tilts the scoring towards candidates with high mobility, ensuring that the caregiver still has surplus energy to overcome fluid obstacles while sharing the impact of the water flow and dragging their companion. The exponential Gaussian decay kernel at the back end smoothly transforms the span of physical space into a weighted penalty, fitting the group behavior characteristics of people in a disaster environment who tend to help those closest to them. For each obstacle agent, the adult individual with the highest fitness score is selected from its candidate list and locked, establishing a one-to-one binding relationship between the obstacle and the caregiver. All locked pairings are summarized to form an auxiliary pairing set.

[0024] For example, in a subway station concourse near the turnstiles, there is a wheelchair user with a baseline speed of 0.15 meters per second. Within a five-meter radius of the assist pairing, three unassigned adult candidates are found with baseline speeds of 1.3 meters per second, 1.0 meters per second, and 1.1 meters per second, respectively. Their initial Euclidean distances from the wheelchair user are 2 meters, 3.5 meters, and 4.8 meters, respectively. Using the aforementioned fit formula, the candidate with a distance of 2 meters and a baseline speed of 1.3 meters per second receives the highest score due to its combination of high speed redundancy and close proximity advantage. This candidate is then selected as the wheelchair user's caregiver and added to the assist pairing set.

[0025] In step S232, optimal matching calculation and bidirectional binding are performed on the auxiliary pairing set to obtain the joint parameter set. It should be noted that, given that the candidate list of each obstacle agent in the auxiliary pairing set generated in the previous sub-step may contain multiple adult individuals, and the same adult individual may appear in the candidate lists of multiple obstacle individuals simultaneously, without global optimal matching and exclusive binding, an adult individual would be repeatedly assigned to multiple obstacle individuals, failing to provide effective physical assistance. Based on this, the technical solution of this application further performs optimal matching calculation and bidirectional binding on the obstacle individuals and adult candidates in the auxiliary pairing set to obtain the joint parameter set. Through the above processing, it can be ensured that each obstacle agent and its companion form a one-to-one exclusive rigid binding relationship, and the original motion parameters of both parties are encapsulated into the joint parameter set, providing complete attribute inputs for subsequent joint velocity fusion and joint tolerance correction.

[0026] More specifically, in a concrete example of this application, the comprehensive fitness score calculated in the previous sub-step is first read from the auxiliary pairing set between all obstacle agents and their candidate adult individuals. For each obstacle agent, its candidate list is sorted from highest to lowest score, and the adult candidate with the highest score is selected as the preferred binding partner. When there is a conflict where multiple obstacle agents' preferred binding partners point to the same adult individual, that adult individual is assigned to the obstacle agent with the highest fitness score. The remaining conflicting obstacle agents are then reverted to the next best option in the candidate list. This competitive allocation process is repeated until all obstacle agents have obtained a unique companion binding partner or the candidate list is exhausted. After the global allocation is completed, a bidirectional index link is established between each confirmed match between the obstacle agent and the adult companion. That is, the unique identifier of the bound companion is written into the obstacle's attribute record, and the unique identifier of the bound obstacle is written into the companion's attribute record, forming a rigid pairing constraint that cannot be unilaterally released. Subsequently, the original motion parameters such as the baseline speed and risk tolerance of each pair of binding relationships are extracted and encapsulated into a joint parameter set containing parameters of the obstacle and the caregiver. Each record in this set carries the complete physical attributes of both parties in a rigid pairing unit, which serves as the direct input for the next sub-step to perform joint speed fusion and joint tolerance correction.

[0027] For example, in the previous sub-step, a wheelchair user in the concourse level turnstile area has identified an adult candidate with a baseline speed of 1.3 meters per second and an initial distance of 2 meters as their preferred match. Meanwhile, another cane user in the platform level also has the same adult as their preferred match. After competitive allocation, because the wheelchair user's compatibility score with this candidate is higher than the cane user's score with it, the adult is ultimately assigned to the wheelchair user, and a bidirectional index is established. The cane user then reverts to the second-highest-scoring adult in their candidate list to complete the pairing. The baseline speed and risk tolerance parameters of each pair are extracted, encapsulated, and written into the joint parameter set.

[0028] In step S233, based on the caregiver gain coefficient and the risk transmission coefficient, the set of parameters of the joint body is fused with joint velocity and corrected with joint tolerance to obtain the enhanced agent set. It should be noted that, given the rigid pairing relationship established in the previous sub-step, the displacements of the person with the obstacle and the caregiver under physical support must remain absolutely consistent. The two are no longer independent moving individuals but constitute an indivisible joint rigid body. Furthermore, the physical limitations of the person with the obstacle constitute the kinematic weakness of the entire unit, and the physical energy of the caregiver can only be converted into controlled auxiliary increments. Therefore, the kinematic parameters of both parties need to be fused into unified equivalent parameters of the joint body. Based on this, the technical solution of this application further performs joint velocity fusion and joint tolerance correction on the set of parameters of the joint body based on the caregiver gain coefficient and the risk transmission coefficient to obtain the enhanced agent set. Through the above processing, each rigid pairing unit can participate in multi-field coupling calculations with a single equivalent velocity and a single equivalent tolerance threshold in subsequent simulations, without having to repeatedly infer the relationship between the two in subsequent steps.

[0029] More specifically, in a concrete example of this application, a joint velocity fusion calculation is first performed on each paired record in the joint parameter set. The baseline velocity of the obstacle agent and the baseline velocity of the caregiver in that record are read, and the kinematic parameters of both parties are fused into a shared joint equivalent velocity. This fusion process follows the formula below: in, To establish a rigid binding between the two parties and share the equivalent speed of the consortium, For the intelligent agent of the obstacle The base speed, For caregivers The base speed, The upper limit of physical speed that can be stimulated by assistive behavior is defined by setting the assistive gain coefficient and strictly controlling it within the range of 0 to 1. The formula uses the baseline speed of the person with disabilities as the basis and uses a hyperbolic tangent function to nonlinearly compress the speed difference between the caregiver and the person with disabilities, so that the speed gain of the combined group is locked within a reasonable physical boundary. Even if the caregiver has a much higher mobility than the person with disabilities, the equivalent speed of the combined group cannot exceed the physiological speed limit of the person with disabilities.

[0030] While completing velocity fusion, joint tolerance correction calculations are performed on the same paired records in the joint parameter set. Since the mutual support between the individuals provides a certain resistance bonus to water flow impact, the original risk tolerance parameters of each individual are read, and the wading limit boundary is reconstructed according to the following formula: in, For the equivalent risk tolerance of the consortium after binding, and These represent the original risk tolerance depth limits for both the disabled intelligent agent and the caregiver. This is a risk transmission coefficient, with a value between 0 and 0.5, used to isolate the pressure increment converted from physical support. As an indicator function, a one-way protective barrier is constructed to limit the gain to only when the caregiver is in better physical condition. After all paired records have completed joint velocity fusion and joint tolerance correction, the fused joint equivalent velocity and joint equivalent risk tolerance are written back to the attribute fields of the corresponding paired units, replacing the original independent parameters of both parties, thereby generating an enhanced agent set.

[0031] In a specific scenario, in the previous sub-step, the baseline speed of the wheelchair user in the station hall turnstile area is 0.15 m / s, and the risk tolerance is 0.12 m / s. The baseline speed of their paired caregiver is 1.3 m / s, and the risk tolerance is 0.45 m / s. The caregiver gain coefficient is 0.6, and the risk transmission coefficient is 0.3. After joint speed fusion calculation, the hyperbolic tangent function compresses the speed difference ratio between the two to a saturation range close to 1, resulting in an equivalent speed of approximately 0.24 m / s for the combined entity. This is far lower than the caregiver's independent walking speed but higher than the wheelchair user's independent mobility. After joint tolerance correction calculation, since the caregiver's risk tolerance is higher than the disabled person's, the indicator function takes effect, and the equivalent risk tolerance of the combined entity is approximately 0.22 m / s, reflecting the limited pressure gain from physical assistance. After writing back the above parameters, this paired unit is written as a complete record into the augmented agent set.

[0032] In step S234, facility dependency labeling and risk weight estimation are performed on the augmented agent set to obtain the family unit set. It should be noted that, given the exclusive path dependency of the alliance units in the augmented agent set on accessible elevators due to the loss of their ability to climb regular stairs, the escape network of such alliances will be immediately blocked if the elevator's circuit fails due to deep water. Therefore, this facility dependency disadvantage needs to be pre-quantified as an explicit risk warning parameter. Based on this, the technical solution of this application further performs facility dependency labeling and risk weight estimation on the augmented agent set to obtain the family unit set. Through the above processing, each alliance not only possesses social relationships and joint kinematic parameters but also carries a quantitative representation of its dependence on key facilities, providing a basis for identifying special group scheduling priorities and single-point failure risks for subsequent path replanning and facility failure game theory.

[0033] More specifically, in a concrete example of this application, the process first iterates through all union units carrying movement obstacle tags in the augmented agent set, reading their current grid coordinates. Then, it interfaces with the region attribute label set in the cellular simulation environment, scans all facility nodes labeled as accessible elevators, counts the total number of currently unaffected, effective accessible elevators, and calculates the Manhattan distance of each union to the nearest effective elevator node. Based on this, a facility exclusivity quantification weight is introduced, transforming the union's facility dependency disadvantage into an explicit intervention alert parameter. This weight is calculated according to the following formula: in, The facility exclusivity dependency weight represents a sensitivity index indicating the likelihood of the consortium falling into dire straits at any time. To dynamically read the total number of currently unaffected accessible elevators from environmental monitoring data, This serves as the average reference speed benchmark for the normal population in a global context. This is the equivalent velocity of the consortium calculated in the previous sub-step. The inverse decay factor at the front end of the formula means that the fewer available elevators there are, the higher the dependency weight. When all elevators fail, the weight tends to a maximum value. The velocity ratio factor at the back end takes into account the consortium's own mobility restrictions. The lower the equivalent velocity, the weaker the consortium's ability to transfer to alternative exits, and the higher its risk of single point of failure. The calculated facility exclusive dependency weight is written into the attribute field of the corresponding consortium. At the same time, regular family units without mobility impairment tags and ordinary individuals in the enhanced agent set are retained together. All units are summarized and output as a unified family unit set. The consortium units in this set simultaneously carry the kinship bidirectional index link, the consortium equivalent velocity, the consortium equivalent risk tolerance, and the facility exclusive dependency weight, and can directly flow to subsequent path replanning and facility failure game steps.

[0034] For example, the equivalent speed of the wheelchair-dependent group in the station hall's turnstile area is 0.24 meters per second. Currently, there are two accessible elevators in the entire venue, neither of which are flooded. The average reference speed for able-bodied individuals is 1.1 meters per second. Based on the formula, the facility exclusivity dependency weight of this group is approximately... If, during subsequent simulations, one of the elevators fails due to water levels exceeding the critical damage threshold, the total number of effective elevators drops to one, and this weight will increase to approximately [missing information]. This reflects a sharp increase in the risk of single-point failure. This weight value, along with the consortium's equivalent velocity and consortium's equivalent risk tolerance, is written into the attribute field of the corresponding record in the family unit set.

[0035] Specifically, in step S3, based on the real-time water level and local crowd density of each grid, a multi-field coupled rate estimation is performed on each individual in the heterogeneous agent set using nonlinear water depth stagnation reduction and congestion friction correction. This is combined with the bimodal switching gain triggered by the spacing between members of the family unit set to obtain the instantaneous movement rate of each agent. It should be noted that during the subway flood evacuation, the actual movement rate of pedestrians is simultaneously affected by the superposition of three factors: local crowding level, water depth, and social behavioral constraints among relatives. Driving agent movement solely with a static baseline speed will not reflect the true characteristics of the dynamic decay of speed with spatiotemporal conditions and local gains in a disaster environment. Therefore, the technical solution of this application further performs multi-field coupled rate estimation on each individual in the heterogeneous agent set based on the real-time water level and local crowd density of each grid, using nonlinear water depth stagnation reduction and congestion friction correction. This is combined with the bimodal switching gain triggered by the spacing between members of the family unit set to obtain the instantaneous movement rate of each agent. Through the above processing, each agent can obtain an instantaneous velocity that integrates the triple constraints of density field, water level field and social relationship field within each simulation step, providing a physically accurate displacement driving force for subsequent discrete coordinate jumps.

[0036] More specifically, in a concrete example of this application, a congestion friction correction based on local crowd density is first performed on each individual in the heterogeneous agent set. The local crowd density at each agent's current grid coordinates is read, and this density is combined with the agent's baseline velocity and the congestion friction coefficient for linear decay calculation, specifically according to the following formula: in, For the first An intelligent agent in The intermediate state rate after density correction at any given time. This is the baseline velocity of the intelligent agent. The congestion friction coefficient is used to characterize the slope of velocity decrease for every unit increase in local density. This represents the local population density of the grid where the agent resides. The minimum creeping velocity under extremely crowded conditions is used to prevent the rate from becoming a non-physical negative value. The intermediate velocities of all individuals calculated above are summed to form the density reduction rate set.

[0037] Subsequently, nonlinear water depth lag reduction is performed on each rate value in the density reduction rate set. The water level depth of each agent's current grid is extracted from the real-time water level of each grid, and this ratio is calculated with the agent's risk tolerance. This ratio is then substituted into the power-law reduction function, specifically according to the following formula: in, For the first An intelligent agent in The multi-field coupling rate after correction for both water depth and density. This represents the intermediate rate corresponding to the density-reduced deceleration rate set. For the real-time water level of each grid, the grid in which this grid is located Water level depth at any given time This represents the risk tolerance of the intelligent agent. The nonlinear water depth drag index is used to characterize the nonlinear growth curvature of the drag on the human lower limbs caused by shallow water waves. The indicator function takes a value of 1 when the water level does not exceed the tolerance limit, and a value of 0 otherwise forces the velocity to zero. The velocities of all individuals calculated above are summed to form a multi-field coupled velocity set.

[0038] After calculating the multi-field coupling rate, bimodal switching gain superposition is performed on individuals with guardianship relationships within the family unit set. The current real-time coordinates of each guardian-ward pair within the family unit set are extracted, the Euclidean dynamic distance between them is calculated, and compared with a preset safe following distance threshold. When the distance exceeds this threshold, a sprint mode gain coefficient is applied to the corresponding rate of the guardian in the multi-field coupling rate set, specifically according to the following formula: in, For the first The instantaneous movement rate ultimately output by each agent. For the corresponding rate in the set of multi-field coupling rates, The sprint mode gain coefficient is used to simulate the kinetic potential triggered when relatives face separation. for The dynamic Euclidean distance between the guardian and the ward at all times. The preset safe following distance threshold, The indicator function is set to 1 when the spacing exceeds a threshold to apply gain; otherwise, it is set to 0 to maintain the original rate. For an independent individual not belonging to any family unit, its instantaneous movement rate is directly equal to the corresponding value in the multi-field coupling rate set. After all individuals have completed the above three-level cascaded rate correction, the instantaneous movement rate of each agent is output as the displacement driving force for subsequent discrete coordinate jumps and position updates.

[0039] Specifically, in step S4, based on the comparison between the real-time water level of each grid and the critical water level for facility damage, a failure status determination is performed on key facilities. Furthermore, a multi-level dynamic game decision is made on the heterogeneous intelligent agent set based on the local congestion density to obtain the updated path target for each intelligent agent. It should be noted that, given that the continuous rise in water level during subway flood evacuation can cause key access facilities such as turnstiles, escalators, and elevators to fail due to short circuits caused by water ingress, previously usable escape routes are blocked. When stranded individuals perceive congestion ahead or the unavailability of target facilities, they will spontaneously engage in a game-like behavior of diverting their routes. This cascading feedback between facility failure and the path decision-making of the crowd constitutes the core uncertainty of the dynamic changes during the evacuation process. Based on this, the technical solution of this application further performs a failure status determination on key facilities based on the comparison between the real-time water level of each grid and the critical water level for facility damage, and performs a multi-level dynamic game decision on the heterogeneous intelligent agent set based on the local congestion density to obtain the updated path target for each intelligent agent. Through the above processing, it is possible to capture facility status changes in real time within each simulation step and drive individuals to dynamically switch between three decision states: maintaining the goal, diverting to the nearest location, and forced replanning. This enables the simulation process to reflect the linkage characteristics of path blockage and adaptive diversion of people in a disaster environment.

[0040] Figure 6 Based on the comparison results between the real-time water level of each grid and the critical water level for facility damage in the dynamic evacuation simulation method for subway station flood disasters based on multi-dimensional coupling according to the embodiments of this application, failure status determination is performed on key facilities, and multi-level dynamic game decision-making is carried out on the heterogeneous intelligent agent set according to the local congestion density to obtain a flowchart of the updated path target of each intelligent agent. Figure 6 As shown, step S4 includes: S41, comparing the maximum real-time water level of each grid within the coverage area of ​​each key facility with the critical water level for facility damage to obtain a set of key facility states; S42, based on the ratio of local congestion density to a preset congestion rejection threshold, performing a nonlinear mapping of the game-triggered probability of the environmental perception pressure of each agent in the heterogeneous agent set to obtain a set of game-triggered probabilities; S43, using the probability values ​​in the set of game-triggered probabilities as parameters, performing independent random trials on each agent, and combining the set of key facility states with the target redirection according to the three-level state machine logic to obtain the updated path target of each agent.

[0041] In step S41, the maximum real-time water level of each grid within the coverage area of ​​each critical facility is compared with the critical water level for facility damage to obtain a set of critical facility states. It should be noted that the electrical control units of critical access facilities such as turnstiles, escalators, and accessible elevators in subway stations will lose their operational capability due to short circuits caused by water ingress once the water level reaches a certain depth. Since the installation height and waterproofing level of different facilities vary, it is necessary to monitor the water level in the coverage area of ​​each facility in real-time and determine the threshold to ascertain whether it is still passable. Based on this, the technical solution of this application further compares the maximum real-time water level of each grid within the coverage area of ​​each critical facility with the critical water level for facility damage to obtain a set of critical facility states. Through the above processing, the availability status of each facility can be dynamically updated within each simulation step, providing real-time facility status information for path blocking determination and target redirection in subsequent game theory decisions.

[0042] More specifically, in a concrete example of this application, the process first traverses all grid regions in the cellular simulation environment that are marked as critical access facilities by the region attribute annotation set, obtaining the set of grid coordinates occupied by each facility in the discretized grid matrix. For each critical facility, the water level values ​​of all grids within its coverage area at the current simulation step are extracted from the real-time water level of each grid, and the maximum value is taken as the representative water level of the facility area. Subsequently, this maximum water level value is compared with the critical water level for facility damage corresponding to the facility. If the maximum water level value is lower than the critical threshold, the facility is determined to be still in normal operation and its status value is recorded as 1. If the maximum water level value reaches or exceeds the critical threshold, the facility is determined to have failed due to water ingress and its status value is forcibly set to 0. This determination process follows the following logic: when the water level of any grid within the facility's coverage area exceeds the critical value, it means that water has intruded into the critical part of the facility, and the entire facility is determined to be unusable. For example, the critical water level for damage to a certain group of turnstiles in a subway station is set at 0.15 meters. When the maximum water level in the four grids covered by the turnstile rises from 0.12 meters in the previous step to 0.17 meters in the current step, the state value of the turnstile switches from 1 to 0. After all critical facilities are judged one by one as described above, their operating status is summarized to form a critical facility state set. This set is updated synchronously in each simulation step and flows to subsequent game decision sub-steps to determine whether the facilities on the agent's current target path are still passable.

[0043] In step S42, based on the ratio of local congestion density to a preset congestion rejection threshold, a nonlinear mapping of the game-triggered probability is performed on the environmental perception pressure of each agent in the heterogeneous agent set to obtain a game-triggered probability set. It should be noted that during subway flood evacuation, whether an individual abandons their current path and seeks an alternative exit is not a deterministic behavior, but a probabilistic decision driven by their perceived local congestion level. The more severe the congestion, the higher the likelihood of an individual changing their route. Therefore, this environmental perception pressure needs to be quantified into a probability parameter that can drive subsequent random trials. Based on this, the technical solution of this application further performs a nonlinear mapping of the game-triggered probability on the environmental perception pressure of each agent in the heterogeneous agent set based on the ratio of local congestion density to a preset congestion rejection threshold to obtain a game-triggered probability set. Through the above processing, individualized trigger probability parameters can be provided for subsequent Bernoulli random trials, enabling each agent to independently generate path-switching decisions based on the congestion level of its local environment within each simulation step.

[0044] More specifically, in a concrete example of this application, the environmental perception pressure is first quantified for each individual in the heterogeneous agent set. The local congestion density at each agent's current grid coordinates is read, and its ratio is calculated with a preset congestion rejection threshold. This ratio is then weighted using a preset panic weight coefficient to obtain the environmental pressure index for each agent. After calculating the environmental pressure index, the game trigger probability is calculated based on the environmental pressure index of each agent to obtain a game trigger probability set. This calculation follows the formula: in, For the first An intelligent agent in The probability of triggering the game at any given moment. A pre-defined panic weighting coefficient is used to characterize the sensitivity of a population to decision-making shifts in a disaster environment. This represents the real-time congestion density of the local area where the intelligent agent is located. This is a preset congestion rejection threshold. Among them, This represents the environmental pressure index. When the local density is far below the threshold, the ratio approaches zero, making the probability of game triggering close to zero, and individuals tend to maintain their original paths. When the local density approaches or exceeds the threshold, the ratio rises rapidly, making the probability approach 1, and individuals' willingness to change paths increases dramatically. The normalized cutoff function at the front end ensures that the probability value is always constrained within a reasonable range of 0 to 1.

[0045] After calculating the game trigger probability, the obtained game trigger probability is used as a parameter to perform Bernoulli random trials on each agent. For each agent, its corresponding... A Bernoulli sampling is performed independently to determine the hit probability. A sampling result of 1 indicates that the individual triggers path-playing behavior within the current step size, while a sampling result of 0 indicates that the individual maintains the current path. For example, if the local congestion density of an individual's grid is 3.5 people per square meter, the congestion exclusion threshold is set to 4 people per square meter, and the panic weighting coefficient is 0.8, then the probability of triggering the game for this individual is calculated as follows: This means that the individual has a 70% probability of triggering a path switching behavior within the current step size. The game trigger probability set of all individuals and their corresponding Bernoulli sampling results are summarized to form the game trigger probability set, which flows to subsequent sub-steps to drive the target relocation decision of the multi-level state machine.

[0046] In step S43, independent random trials are performed on each agent using the probability values ​​in the game trigger probability set as parameters, and target redirection is performed according to a three-level state machine logic in conjunction with the key facility state set to obtain the updated path target for each agent. It should be noted that since the previous sub-step only completed the quantification calculation of the game trigger probability, it has not yet transformed the probability into actual path switching behavior. Furthermore, after triggering the game, each agent needs to respond in a tiered manner based on the availability status of the current target facility, rather than uniformly performing global replanning. Based on this, the technical solution of this application further performs independent random trials on each agent using the probability values ​​in the game trigger probability set as parameters, and performs target redirection according to a three-level state machine logic in conjunction with the key facility state set to obtain the updated path target for each agent. Through the above processing, each agent can dynamically switch between three levels of decision-making—maintaining the target, nearby diversion, and forced replanning—based on its local perception conditions and facility availability, outputting a deterministic navigation target for subsequent coordinate transitions.

[0047] More specifically, in a concrete example of this application, all individuals in the heterogeneous agent set are first traversed. For each individual, an independent Bernoulli random trial is performed with the probability value corresponding to the probability in the game trigger probability set as the hit probability. If the trial result is 0, then the individual has not triggered game behavior within the current simulation step, its path goal remains unchanged, and it continues to move directly along the gradient direction of the global static distance potential field towards the original nearest exit. This is the first level in the three-level state machine, namely maintaining the target state.

[0048] If the test result is 1, the individual triggers path game behavior and then queries the critical facility status set to determine whether the critical facilities on its current target path are still in normal operation. If the facility status value on the target path is 1 and the local density ahead is high but the path is still passable, then the second level of the three-level state machine, namely the nearest diversion state, is executed, switching the individual's target from the current nearest exit to an adjacent exit or passage with a lower degree of congestion within the field of view, in order to alleviate the local density concentration.

[0049] If the status value of a critical facility on the target path is set to 0, indicating that the facility has failed due to water level exceeding the critical damage threshold, the third level of the three-level state machine, namely the forced replanning state, is triggered. This marks the facility as impassable in the pathfinding topology, recalculates the shortest path distance from the individual to all other available exits based on the global static distance potential field, and designates the nearest exit with a normal facility status as the new navigation target. For example, if an individual's original target was evacuation via the gate at exit A, but that gate fails due to water level reaching 0.17 meters, the individual is forced to replan to evacuate via the stairs at exit B, and their navigation target changes from exit A to exit B. After all individuals undergo the above three-level state machine logic judgment, they output their respective deterministic navigation nodes within the current simulation step, which are then aggregated to form the updated path targets for each agent, flowing to subsequent steps for driving the direction calculation of discrete coordinate transitions.

[0050] Specifically, in step S5, based on the instantaneous movement rate and simulation step size, the survival status of each agent is determined by discrete coordinate jumps driven by the accumulation of motion residuals and by water level exceeding limits. The coordinate sequences of each agent and the updated path target are serialized into structured time-series data frames to obtain batch statistical indicators and spatiotemporal trajectory heatmaps for the evacuation simulation. It should be noted that, given that the previous steps have completed the multi-field coupling calculation of the instantaneous movement rate of each agent and the dynamic game decision of the updated path target, but the displacement of the agent in the cellular grid environment is a discrete, grid-by-grid jump rather than a continuous slide, when an individual is in a low-speed wading state, the theoretical displacement within a single simulation step size may be less than the length of a grid edge. Direct truncation would cause the individual to remain completely stationary within multiple consecutive steps, resulting in quantization errors. At the same time, the continuously rising water level may exceed the individual's risk tolerance limit at a certain moment, causing it to lose its ability to move. It is necessary to accurately process these two types of critical states and transform the spatiotemporal data of the entire process into structured outputs that can be used for decision analysis. Based on this, the technical solution of this application further uses instantaneous movement speed and simulation step size to determine the survival status of each agent through discrete coordinate jumps driven by the accumulation of motion residuals and triggered by water level exceeding limits. The coordinate sequences and update path targets of each agent are serialized into structured time-series data frames to obtain batch statistical indicators and spatiotemporal trajectory heatmaps for evacuation simulation. Through the above processing, discrete quantization errors under low-speed wading can be eliminated and trapped individuals can be marked in real time. Simultaneously, the unstructured simulation state is transformed into quantitative evaluation results and visualized heatmaps containing indicators such as total evacuation time, trapped personnel ratio, evacuation success rate, and space utilization balance, providing data support for subway station disaster prevention design and emergency plan optimization.

[0051] Figure 7To implement the dynamic evacuation simulation method for subway station flood disasters based on multidimensional coupling according to the embodiments of this application, based on instantaneous movement speed and simulation step size, the method performs discrete coordinate jump driven by the accumulation of motion residuals and survival status determination triggered by water level exceeding limits for each agent. The coordinate sequences and update path targets of each agent are serialized into structured time-series data frames to obtain a flowchart of batch statistical indicators and spatiotemporal trajectory heatmaps for evacuation simulation. For example... Figure 7 As shown, step S5 includes: S51, gradually accumulating the product of the instantaneous movement rate of each agent and the simulation step size, and performing a coordinate jump along the target direction of the update path when the accumulated amount exceeds the cell side length, so as to obtain a discrete coordinate sequence; S52, based on the comparison results of the real-time water level of each grid in the discrete coordinate sequence and its risk tolerance, performing coordinate locking and trapped state marking on the over-limit individuals to obtain the agent survival state set and the final coordinate sequence; S53, performing nonlinear color mapping rendering on the full-field historical access frequency of the final coordinate sequence, and serializing the agent survival state set and the update path target into a structured time-series data frame according to a preset protocol to obtain batch statistical indicators and spatiotemporal trajectory heatmap.

[0052] In step S51, the product of the instantaneous movement rate and the simulation step size of each agent is progressively accumulated. When the accumulated amount exceeds the cell edge length, a coordinate jump is performed along the target direction of the update path to obtain a discrete coordinate sequence. It should be noted that since the position update of the agent in the cellular automaton framework is a discrete jump per grid, when the agent is in deep water or in a high-density crowded state, its instantaneous movement rate may be extremely low after being reduced by multiple field couplings. The theoretical displacement within a single simulation step size is less than the edge length of a grid. If this displacement is directly discarded, the agent will be completely stationary within multiple consecutive steps, thus accumulating quantization errors. Based on this, the technical solution of this application further progressively accumulates the product of the instantaneous movement rate and the simulation step size of each agent. When the accumulated amount exceeds the cell edge length, a coordinate jump is performed along the target direction of the update path to obtain a discrete coordinate sequence. Through the above processing, the discrete quantization error in the low-speed state can be eliminated by the cross-step accumulation mechanism of motion residual, ensuring that even under extremely low speed conditions, the agent can still complete an effective displacement after several steps.

[0053] More specifically, in a concrete example of this application, a progressively increasing displacement accumulation and coordinate jump determination are performed on each individual in the heterogeneous agent set. First, the instantaneous movement rate of the individual within the current simulation step size, output by the previous step, is read. This rate is multiplied by the simulation step size to obtain the theoretical displacement for the current step size. This displacement is then added to the historical motion residual retained by the individual from the previous step size, specifically according to the following formula: in, For the first An intelligent agent in The cumulative residual amount of motion at any given moment This is a historical remnant retained from the previous step. This represents the instantaneous movement speed of the individual. The simulation step size is used. The accumulated residual is then compared to the physical edge length of a single cell. If the residual is greater than or equal to one cell edge length, the individual is driven to generate a one-cell coordinate jump on the discrete grid along the gradient direction indicated by the update path target. Simultaneously, the value of one cell edge length is subtracted from the residual to retain the remainder for further accumulation in subsequent steps. If the residual has not yet reached one cell edge length, the individual remains stationary within the current step size, and only its residual value is updated. For example, if an individual's instantaneous movement speed in deep water is 0.3 meters per second, and the simulation step size is 0.5 seconds, the theoretical displacement in the current step size is 0.15 meters. If its historical residual is 0.82 meters, the accumulated residual reaches 0.97 meters, still less than the 1-meter grid edge length, so the individual remains in place. After another 0.15-meter accumulation in the next step, the residual reaches 1.12 meters, triggering a coordinate jump and updating the residual to 0.12 meters. After all individuals complete the above determinations within the current step size, their respective grid coordinates are aggregated to form a discrete coordinate sequence, which serves as the location input for subsequent survival status determination and heatmap rendering.

[0054] In step S52, based on the comparison between the real-time water level and risk tolerance of each grid in the discrete coordinate sequence, coordinate locking and trapped state marking are performed on individuals exceeding the limit to obtain the set of agent survival states and the final coordinate sequence. It should be noted that, due to the continuous rise in water level during the subway station flooding, the water depth of some individuals' grids may exceed their risk tolerance limit within a certain simulation step. At this point, the individual has lost its ability to move autonomously due to the impact of the water flow. Allowing it to continue participating in coordinate jumps would cause the simulation results to deviate from physical reality. Therefore, the technical solution of this application further performs coordinate locking and trapped state marking on individuals exceeding the limit based on the comparison between the real-time water level and risk tolerance of each grid in the discrete coordinate sequence to obtain the set of agent survival states and the final coordinate sequence. Through the above processing, individuals who have lost their ability to move can be identified and frozen in real time, ensuring that they do not move in subsequent simulation steps, and providing a state basis for calculating the trapped personnel ratio in batch statistical indicators.

[0055] More specifically, in a concrete example of this application, each individual in the discrete coordinate sequence output by the previous sub-step is traversed, its current grid coordinates are read, and the water level depth at the current simulation step is extracted from the real-time water level of each grid. Simultaneously, the individual's risk tolerance parameter is read from the heterogeneous agent set, and the two are compared. If the current grid water level does not exceed the individual's risk tolerance, the individual is determined to be in a normal passage state, its survival status is marked as normal, and its coordinates remain unchanged from the updated values ​​in the discrete coordinate sequence. If the current grid water level has reached or exceeded the individual's risk tolerance, the individual is determined to have lost its autonomous movement ability due to excessive water depth, its survival status is forcibly marked as trapped, and its coordinates are locked to remain fixed at the grid position where it was trapped at the time of trapping in the current and all subsequent simulation steps, without any further displacement. For the joint unit in the family unit set, its risk tolerance is the joint equivalent risk tolerance. When the water level exceeds this equivalent value, the obstacle and the caregiver within the joint unit are simultaneously marked as trapped and their coordinates are locked synchronously. For example, if an elderly individual has a risk tolerance of 0.28 meters, and the real-time water level in their current grid rises from 0.25 meters in the previous step to 0.31 meters in the current step, exceeding their tolerance limit, the individual's survival status is marked as trapped, and their coordinates are locked to that grid. After all individuals are assessed in the above manner, their survival statuses are summarized to form an agent survival status set, and their locked and corrected coordinates are summarized to form a final coordinate sequence. Both are then used in subsequent sub-steps for time-series data serialization and statistical index calculation.

[0056] In step S53, the historical access frequency of the final coordinate sequence is rendered using nonlinear color mapping, and the agent survival state set and updated path target are serialized into structured time-series data frames according to a preset protocol to obtain batch statistical indicators and spatiotemporal trajectory heatmaps. It should be noted that since the agent survival state set, final coordinate sequence, and updated path target output from the preceding sub-steps are all instantaneous unstructured data within each simulation step, if they are not time-series encapsulated and cumulatively statistically processed according to a unified protocol, it is impossible to extract macro-level evaluation indicators such as total evacuation time and the ratio of trapped personnel from the global time dimension, nor can it identify congestion bottlenecks and high-risk stranded areas in the spatial dimension. Based on this, the technical solution of this application further performs nonlinear color mapping rendering on the historical access frequency of the final coordinate sequence and serializes the agent survival state set and updated path target into structured time-series data frames according to a preset protocol to obtain batch statistical indicators and spatiotemporal trajectory heatmaps. Through the above processing, the spatiotemporal state of the entire simulation process can be transformed into quantitative evaluation results and visualized spatial distribution maps that can be directly used for disaster prevention design and emergency plan optimization.

[0057] More specifically, in a concrete example of this application, the historical access frequency of the entire field is first accumulated and a heatmap is rendered. At the end of each simulation step, the current grid coordinates of all individuals in the final coordinate sequence are extracted, and a historical access counter is maintained for each grid in the discretized grid matrix. The number of individuals occupying that grid within the current step is accumulated into the counter. After all simulation steps have been executed, the counter value of each grid is the total frequency of that grid being trampled and accessed during the entire simulation process. Since the frequency of a few extremely congested bottleneck areas is much higher than that of other areas, if linear color mapping is used, the color differences in most areas will be difficult to distinguish. Therefore, a logarithmic enhancement algorithm is introduced to perform nonlinear color mapping rendering, specifically executed according to the following formula: in, This represents the color depth value output by mapping the grid onto the spatiotemporal trajectory heatmap, with a value range from 0 to 255. This represents the total historical access frequency of the grid. This represents the highest historical access frequency across all grids in the entire site. The contrast enhancement coefficient is used to adjust the sharpness of the color gradient. Through the nonlinear compression properties of the natural logarithmic function, the color levels between high-frequency and low-frequency regions are displayed in a balanced manner, making congestion bottlenecks and high-risk stagnation areas clearly distinguishable in the heat map, thereby generating a spatiotemporal trajectory heat map.

[0058] While rendering the heatmap, the temporal state data of the entire simulation process is structurally encapsulated. The set of agent survival states and the updated path target within each simulation step are encapsulated frame by frame into structured temporal data frames according to a preset serialization protocol and written to a buffer. Each frame contains the timestamp of that step, the coordinates of all individuals, survival status markers, and the current navigation target. After the simulation ends, all temporal data frames in the buffer are traversed. The timestamp of the last individual arriving at the exit is counted as the total evacuation time; the ratio of the number of individuals in the "trapped" state to the total number of simulated individuals is counted as the "trapped personnel ratio"; the ratio of the number of individuals successfully reaching the exit to the total number of simulated individuals is counted as the evacuation success rate; and the space utilization balance is calculated based on the actual distribution of people passing through each exit. These indicators are summarized into batch statistical indicators, which, together with the spatiotemporal trajectory heatmap, serve as the final output of the evacuation simulation.

[0059] Figure 8 This is a simplified pedestrian evacuation trajectory visualization heatmap for a subway station evacuation platform according to an embodiment of this application. For example... Figure 8As shown, this heatmap is generated by performing log-enhanced nonlinear color mapping rendering on the historical access frequency of the final coordinate sequence in step five. In the figure, the blue area corresponds to impassable obstacles such as walls or unvisited areas, the black dots correspond to pillars, the white area shows the normal evacuation trajectory of pedestrians, and the dark red highlighted area corresponds to the congestion bottleneck and the concentrated area of ​​stagnation that are frequently stepped on by the agent throughout the simulation process. The darker the color, the higher the historical access frequency. It can intuitively identify high-risk congestion nodes in the evacuation process, such as in front of the turnstile, near the stair entrance, and the exit convergence area, and provide a quantitative spatial reference for optimizing the passage width and adjusting the emergency guidance strategy in the disaster prevention design of subway stations.

[0060] In summary, a simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to an embodiment of this application is explained. First, the subway station space is discretized into cellular grids with regional attributes through color semantic recognition, and shallow water wave diffusion calculation is performed under dynamic topological constraints to obtain the real-time water level distribution across the entire site. Then, heterogeneous intelligent agents covering multiple age groups are generated based on demographic parameters, and differentiated movement capabilities are granted through Gaussian perturbation, and a social relationship mapping of family units is established. On this basis, water depth hindrance, density friction, and kinship synergy gain are fused into a multi-field coupled instantaneous velocity estimation, enabling water level field changes to be fed back to individual movement capabilities in real time. Furthermore, a multi-level dynamic replanning of paths is achieved through facility failure judgment and a Bernoulli game mechanism driven by congestion perception, solving the linkage problem between path blocking and adaptive crowd diversion in the cascading feedback chain. Finally, discrete coordinate updates are completed through a motion residual accumulation mechanism and survival status determination, and batch statistical indicators and heat maps are output. Meanwhile, for groups with mobility impairments, the system introduces assistance dependency pairing, rigid union velocity fusion and facility exclusive dependency labeling mechanisms to make the physical support constraints, motion limit compression and single-point failure risks of special groups explicit in advance, so that the simulation system can realistically depict the nonlinear behavior degradation and path obstruction characteristics of such groups in flood disasters.

[0061] Those skilled in the art will understand that the steps, measures, and schemes in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted; furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted; furthermore, the steps, measures, and schemes in the prior art that are similar to those disclosed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above-described embodiments are merely illustrative of several implementation methods of this disclosure, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent for the embodiments of this disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this disclosure, and these all fall within the protection scope of the embodiments of this disclosure. Therefore, the protection scope of the embodiments of this disclosure should be determined by the appended claims. As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and detail can be made without departing from the spirit and scope of the present invention as defined in the appended claims.

[0062] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A simulation method for dynamic evacuation during floods in subway stations based on multidimensional coupling, characterized in that, include: Step 1: Based on the color semantic recognition rules of the subway station geometric model, the subway station space is discretized into cellular meshes and the regional attributes are labeled. The shallow water wave diffusion calculation under dynamic topological constraints is performed on the external water injection flux to obtain the cellular simulation environment and global static distance potential field that carry the real-time water level of each mesh. Step 2: Based on demographic parameters, differentiated assignment of baseline speed and risk tolerance and social relationship mapping are performed on the heterogeneous agent set containing multiple age groups to obtain a set of family units with bidirectional kinship index links. Step 3: Based on the real-time water level and local population density of each grid, perform multi-field coupled rate estimation for each individual in the heterogeneous intelligent agent set using nonlinear water depth hindrance reduction and crowding friction correction, and combine the member spacing of the family unit set to trigger dual-mode switching gain to obtain the instantaneous movement rate of each intelligent agent. Step 4: Based on the comparison results between the real-time water level of each grid and the critical water level for facility damage, the failure status of key facilities is determined, and multi-level dynamic game decision-making is carried out on the heterogeneous intelligent agent set according to the local congestion density to obtain the updated path target of each intelligent agent. Step 5: Based on the instantaneous movement rate and simulation step size, perform discrete coordinate jump driven by the cumulative motion residual amount and survival status determination triggered by water level exceeding limit for each agent. Then, serialize the coordinate sequence and update path target of each agent into a structured time-series data frame to obtain the batch statistical indicators and spatiotemporal trajectory heat map of the evacuation simulation.

2. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 1, characterized in that, Step one includes: Based on the preset physical size ratio, the pixel drawing data of the subway station geometric model is divided into two-dimensional grids and matched with color semantic intervals to obtain a discretized grid matrix and a set of regional attribute annotations. Using the grid marked as the exit in the regional attribute label set as the zero potential energy starting point, the passable grids in the discretized grid matrix are traversed by layer-by-layer Manhattan distance accumulation to obtain the global static distance potential energy field. External water injection flux is injected into the preset source grid of the discretized grid matrix, and shallow water wave diffusion calculation is performed in combination with the regional attribute annotation set to obtain the cellular simulation environment that carries the real-time water level of each grid.

3. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 1, characterized in that, Step two includes: The total number of simulated people in the demographic parameters is randomly sampled and group labels are assigned according to the preset age group proportions to generate an attribute distribution agent set; Based on the Gaussian perturbation superposition model, each agent in the attribute distribution agent set is assigned a normal perturbation value of baseline speed and risk tolerance according to the statistical expectation value corresponding to its group label, so as to obtain a heterogeneous agent set. A family unit set is obtained by pairing vulnerable individuals and adult individuals in a heterogeneous set of intelligent agents using a bidirectional kinship index link.

4. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 1, characterized in that, Step three includes: Based on the crowding friction coefficient, the baseline velocity of each agent in the heterogeneous intelligent agent set is linearly attenuated and corrected according to the local crowd density of its grid to obtain the density reduction deceleration set. Based on the nonlinear water depth retardation function, the rate values ​​in the density reduction rate set are calculated by power law reduction according to the ratio of the real-time water level of each grid to the individual risk tolerance, so as to obtain the multi-field coupled rate set. Based on the comparison results of the real-time Euclidean distance between the guardian and the ward in the family unit set and the safe following distance threshold, the sprint mode gain coefficients of the corresponding individuals in the multi-field coupling rate set are superimposed to obtain the instantaneous movement rate of each agent.

5. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 1, characterized in that, Step four includes: The maximum real-time water level of each grid within the coverage area of ​​each critical facility is compared with the critical water level for facility damage to determine the status set of critical facilities. Based on the ratio of local congestion density to a preset congestion rejection threshold, a nonlinear mapping of the game trigger probability is performed on the environmental perception pressure of each agent in the heterogeneous intelligent agent set to obtain a game trigger probability set. Independent random trials are performed on each agent using the probability values ​​in the game trigger probability set as parameters, and the target is redirected according to the three-level state machine logic in combination with the key facility state set to obtain the updated path target of each agent.

6. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 1, characterized in that, Step five includes: The product of the instantaneous movement rate of each agent and the simulation step size is gradually accumulated. When the accumulated amount exceeds the cell side length, a coordinate jump is performed along the target direction of the update path to obtain a discrete coordinate sequence. Based on the comparison results of the real-time water level of each grid in the discrete coordinate sequence and its risk tolerance, coordinate locking and trapped state marking are performed on the over-limit individuals to obtain the set of agent survival states and the final coordinate sequence. The frequency of all historical accesses in the final coordinate sequence is rendered using nonlinear color mapping, and the set of agent survival states and update path targets are serialized into structured time-series data frames according to a preset protocol to obtain batch statistical indicators and spatiotemporal trajectory heatmaps.

7. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 2, characterized in that, Two-dimensional meshing and color semantic interval matching are performed on the pixel drawing data of the subway station geometric model to obtain a discretized mesh matrix and a set of region attribute annotations, including: Based on a preset physical size ratio, the pixel drawing data of the subway station geometric model is divided into equally spaced two-dimensional grids to obtain a discretized grid matrix. The pixel color values ​​of each grid in the discretized grid matrix are matched with the preset color semantic intervals to obtain a set of region attribute annotations.

8. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 5, characterized in that, A nonlinear mapping of the game-triggered probability to the environmental perception pressure of each agent in a heterogeneous intelligent agent set is performed to obtain a set of game-triggered probabilities, including: Based on the ratio of local congestion density to a preset congestion rejection threshold, environmental pressure is calculated for each agent in the heterogeneous intelligent agent set to obtain the environmental pressure index of each agent. The game trigger probability set is obtained by calculating the game trigger probability based on the environmental pressure index of each individual.

9. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 6, characterized in that, The final coordinate sequence is rendered using nonlinear color mapping based on the total historical access frequency. The agent's survival state set and update path target are then serialized into structured time-series data frames according to a preset protocol to obtain batch statistical indicators and spatiotemporal trajectory heatmaps, including: Based on the contrast enhancement coefficient, nonlinear color mapping rendering is performed on the full-field historical access frequency of the final coordinate sequence to obtain a spatiotemporal trajectory heatmap. The set of agent survival states and update path goals are encapsulated into structured time-series data frames according to a preset heterogeneous data serialization protocol. Based on structured time-series data frames, statistics are collected on total evacuation time, the ratio of trapped personnel, evacuation success rate, and space utilization balance to obtain batch statistical indicators.

10. The simulation method for dynamic evacuation of subway stations based on multidimensional coupling according to claim 3, characterized in that, A set of family units is obtained by performing bidirectional kinship indexing pairing between vulnerable individuals and adult individuals in a heterogeneous set of intelligent agents, including: Based on preset action obstacle labeling rules, obstacle attribute identification and adult candidate screening are performed on a heterogeneous intelligent agent set to generate an auxiliary pairing set; Optimal matching calculations and bidirectional bindings are performed on the auxiliary pairing set to obtain the union parameter set; Based on the care gain coefficient and the risk transmission coefficient, joint velocity fusion and joint tolerance correction are performed on the set of parameters of the consortium to obtain an enhanced agent set. Facility dependency labeling and risk weight estimation are performed on the set of augmented agents to obtain the set of household units.

Images