Multi-disaster point emergency evacuation method, planning method, planning system and evacuation system

Abstract

The present invention discloses a multi-disaster point emergency evacuation method, a planning method, a planning system and an evacuation system. In the mathematical model, the disaster victims crossing the dangerous area, the load balance of the shelters and the redundancy of the evacuation distance of the disaster victims are considered; in view of the emergency evacuation problem of multiple disaster points, a multi-disaster emergency evacuation algorithm is established; the emergency evacuation algorithm can effectively embed the information of the disaster victims, shelters and disaster points in the evacuation problem, extract their relationship characteristics, and generate an efficient evacuation plan; the emergency evacuation algorithm adopts the optimization method of the policy gradient algorithm, effectively realizes the gradient descent of the algorithm parameters, and improves the algorithm training efficiency; in view of the emergency evacuation problem of multiple disaster points, a cloud-edge network emergency evacuation system is established, and the evacuation system adopts the "cloud-edge-end" three-layer architecture, which effectively improves the evacuation planning efficiency, and comprehensively considers the algorithm training and evacuation response processes, laying a solid foundation for the practical application of the system.

Description

Technical Field

[0001] The present invention relates to the fields of cloud-edge fusion computing, deep reinforcement learning and urban disaster management, and specifically to a multi-disaster point emergency evacuation method, a planning method, a planning system and an evacuation system. Background Art

[0002] Crowd evacuation is one of the important measures of disaster management. Its purpose is to transfer the victims to a safe area within a limited time when a disaster occurs, so as to protect the lives and property of the people. At present, the means that can effectively improve the efficiency of emergency evacuation can be divided into two categories, namely, evacuation drills and evacuation planning. Evacuation drills are designed to simulate the situation of disaster escape in advance and improve the disaster response ability of personnel. However, this method requires real people to participate, the implementation cost is high, and it is difficult to conduct large-scale drills. Evacuation planning aims to provide the best evacuation route for the victims, design an efficient scheduling plan, reasonably arrange the evacuation time, and fairly allocate the required resources. In the design process of evacuation planning, no real people are required to participate, which can effectively overcome the cost and scale limitations of evacuation drills. Therefore, it has become one of the most concerned topics in the field of disaster management.

[0003] Currently, researchers have proposed many evacuation schemes for different evacuation scenarios, which can be roughly divided into two categories, namely model-driven methods and data-driven methods. Model-driven methods usually require the establishment of a comprehensive and detailed system model based on prior knowledge to analyze the impact of various factors on the evacuation efficiency of disaster-stricken people when a disaster occurs. However, these models are often based on predefined assumptions and rules, which make it difficult to simulate the real evacuation environment, severely limiting the generalization and stability of the model.

[0004] Data-driven methods can effectively solve the above problems. Such methods infer the changing trend of objects based on the statistical laws of data without analyzing the physical characteristics and dynamic behaviors of objects, and have high flexibility. The current mainstream data-driven methods include heuristic algorithms and reinforcement learning algorithms (RL). Heuristic algorithms usually model the problem as an optimization problem and construct an initial solution, based on which they search for the optimal solution in the feasible domain, such as ant colony optimization (ACO) and genetic algorithm (GA). However, heuristic algorithms are usually greatly affected by parameter selection and problem structure, and often find solutions in an iterative and search manner. When the problem is more complex and the scale of the solution is large, a large amount of computing resources is required. Emergency evacuation planning is a time-sensitive task, which requires timely provision of evacuation plans for reference by disaster victims within a limited time. Therefore, traditional heuristic algorithms are difficult to apply to large-scale emergency evacuation planning.

[0005] As one of the important branches of machine learning, reinforcement learning can effectively overcome the above problems. It can effectively handle a variety of data mining problems by training intelligent agents to interact with the environment to maximize reward feedback. In addition, a special variant of reinforcement learning, deep reinforcement learning (DRL), introduces deep learning (DL) on the basis of reinforcement learning. Thanks to the powerful learning ability of deep learning, a well-trained deep reinforcement learning model can provide solutions to problems without a lot of calculations, and has great potential in solving time-sensitive problems. However, the existing emergency evacuation planning methods based on deep reinforcement learning are often oriented to single-disaster point evacuation situations. In reality, disasters often present a trend of multi-point outbreaks, and existing solutions are difficult to cope with such situations. Moreover, these solutions focus on algorithm efficiency and ignore the practical application of the algorithm. They fail to build an effective evacuation system, which limits the practical application of evacuation algorithms.

[0006] In summary, the current work on large-scale emergency evacuation planning in cities still has the following shortcomings: First, the problem of emergency evacuation at multiple disaster points is ignored. Most current emergency evacuation methods mainly consider the scenario of a single disaster outbreak, that is, there is only one disaster outbreak in the problem, and the modeling is based on this premise. However, in actual situations, disasters often break out at multiple points. For example, when an earthquake occurs, secondary disasters such as tsunamis, aftershocks, and landslides may occur at the same time. These disasters may not occur in one place, so it is necessary to design a method that can cope with emergency evacuation in multi-disaster scenarios. Second, there is a lack of a system framework suitable for practical applications. At present, cities are developing rapidly, and smart cities are gradually becoming an important concept in urban planning and management. However, the existing emergency evacuation schemes focus on improving algorithm performance, but do not provide a system framework suitable for the future development of cities, which makes it difficult for the algorithm to be applied in practice. Third, the problem of load balancing of shelters is ignored. The existing schemes focus on evacuation efficiency, that is, evacuation distance, evacuation time, etc., but ignore the humanization of the evacuation process. As the evacuation destination, the degree of congestion in the shelters greatly affects the evacuation experience of the affected people. In addition, overcrowding in shelters will lead to tight resource allocation, while too few people will result in inefficient resource utilization. Summary of the invention

[0007] In order to solve the defects of the above-mentioned prior art, the present invention provides a multi-disaster point emergency evacuation method, planning method, planning system and evacuation system, which can solve the problem that the existing emergency evacuation planning method cannot cope with the evacuation of multiple disaster points, and can solve the problem that the existing emergency evacuation planning method does not construct a systematic evacuation framework and is difficult to apply in practice, and does not consider the load balancing of shelters and is difficult to improve the evacuation efficiency and humanization level.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solutions: a multi-disaster point emergency evacuation method, comprising:

[0009] Establish a dangerous area crossing detection mechanism to effectively evaluate the probability of crossing the dangerous area during the evacuation of disaster victims, and obtain the number of people crossing the dangerous area, where the number of people crossing the dangerous area represents the number of people who may cross the dangerous area;

[0010] Establish a shelter load balancing mechanism to accurately describe the shelter load situation and obtain the shelter load penalty, where the shelter load penalty represents the sum of the load penalty values ​​of all shelters. The load penalty value of a single shelter will be calculated based on the ratio between the number of refugees received by the shelter and its preset maximum capacity;

[0011] Establish a method for calculating the evacuation distance by taking into account the detour around the danger zone, and obtain the evacuation distance, wherein the evacuation distance represents the sum of the distances of all disaster victims evacuated to the corresponding shelters;

[0012] The evacuation cost is obtained according to the number of people crossing the dangerous area, the shelter load penalty, and the evacuation distance.

[0013] Furthermore, the dangerous area crossing detection mechanism comprises the following sub-steps:

[0014] S1-1: Determine the location of the disaster point and delineate the danger zone;

[0015] Establish a plane rectangular coordinate system and regard it as the experimental area;

[0016] Disaster Point k As a point in the experimental area And set a coverage radius r according to its intensity k ;

[0017] Disaster Point k The coverage is is the center of the circle and the radius is r k Circular area of

[0018] Repeat the above steps until all disaster points are traversed;

[0019] S1-2: Single evacuation allocation R i = {v i ,s j}Detection of crossing dangerous areas;

[0020] Assign R to any group in the evacuation plan S i = {v i ,s j}, indicating that the i-th disaster victim is assigned to shelter j, and v i With s j The line segment l i,j ;

[0021] Calculate the disaster point d k With v i ,s j , l i,j The Euclidean distance

[0022] like Any one of them is smaller than the disaster point d k The coverage radius r k , indicating that the i-th disaster victim crosses the disaster point d during the evacuation to shelter j k The possibility of coverage will cross the discriminant factor Set to 1;

[0023] like Both are larger than the radius r k, indicating that the number of victims v i To the shelters j During the evacuation process, the disaster site will not be crossed. k The coverage range will cross the discriminant factor Set to 0;

[0024] Calculate the disaster point d k+1 With v i ,s j , l i,j The Euclidean distance is calculated and the above steps are repeated until all disaster points are traversed;

[0025] S1-3: Detection of the overall evacuation plan crossing the dangerous area;

[0026] Single evacuation allocation R i = {v i ,s j After the detection of crossing the dangerous area is completed, a judgment sequence will be obtained Sum the elements in the sequence, that is

[0027] If c i,j ≥1, indicating that the number of victims v i To the shelters j During the evacuation process, there is a phenomenon of crossing the disaster coverage area, and the discriminant factor C i Set to 1;

[0028] If c i,j =0, indicating that the number of victims v i To the shelters j If there is no phenomenon of crossing the disaster coverage area during the evacuation process, the discriminant factor C i Set to 0;

[0029] Determine the next set of allocations R in the evacuation plan S i+1 Whether there is a cross-over phenomenon, repeat the above steps until the entire evacuation plan is traversed;

[0030] After traversing the evacuation plan, another discriminant factor sequence is obtained Sum the sequence elements, i.e. the number of people crossing the dangerous area

[0031] Furthermore, the shelter load balancing mechanism includes the following sub-steps:

[0032] Calculation of single shelter load penalty:

[0033]

[0034] Among them, rec j In the evacuation plan S, shelter sj The number of disaster victims can be calculated by evacuating the evacuation plan S, s j The number of occurrences of is calculated;

[0035] Indicates Shelters j Suitable receiving quantity range, is the receiving lower bound, is the upper bound of the reception, and the present invention sets it to [0.4×cap j ,0.8×cap j ];

[0036] α and β represent the penalty coefficients for too low and too high number of recipients, respectively. The sparse penalty coefficient α = -1 / 0.4cap j , congestion penalty coefficient β = 1 / 0.2cap j ;

[0037] and represents the capacity discrimination factor, if Set to 1, otherwise set to 0; if Set to 1, otherwise set to 0;

[0038] Repeat the above steps to calculate the shelter s j+1 The load penalty value is increased until all shelters are traversed;

[0039] Adding up the load penalty values ​​of all shelters, we can get the load penalty Penalty (S|I) of evacuation plan S:

[0040]

[0041] Furthermore, the evacuation distance calculation method considering detouring the dangerous area includes the following sub-steps:

[0042] When R i = {v i ,s j}'s cross-judgment factor C i When it is 1, it indicates that R is allocated. i This will cause disaster victims to cross dangerous areas during the evacuation process, increasing route redundancy, that is, increasing the allocation R i The corresponding evacuation distance is calculated as follows:

[0043]

[0044] To allocate R i = {v i ,s j}Evacuation distance of disaster victims, For disaster victims i With shelters j The straight-line distance is the Euclidean distance, where γ is the redundancy factor, and C i For R i The cross-judgment factor;

[0045] Then calculate the next allocation R i+1 The corresponding evacuation distance until the evacuation plan S is traversed;

[0046] By adding up all the assigned evacuation distances, we can get the total evacuation distance Length(S|I) of the evacuation plan:

[0047]

[0048] Furthermore, the evacuation cost = the number of people crossing the dangerous area + the shelter load penalty + the evacuation distance, that is:

[0049]

[0050] Among them, Cost(S|I) is the evacuation cost, Length(S|I) is the evacuation distance, Penalty(S|I) is the shelter load penalty, Count(S|I) is the number of people crossing the dangerous area, and Represent the weights of the three components in the objective function respectively.

[0051] A multi-hazard point emergency evacuation planning method, comprising:

[0052] T1: Simulate multi-disaster emergency evacuation scenarios to generate a large number of evacuation problem instances for training algorithm parameters;

[0053] T2: Perform data processing on the information of each object in a single evacuation instance and implement feature extraction to obtain an overall observation of the problem, and generate an evacuation plan based on the overall observation;

[0054] T3: Evaluate a single evacuation instance, generate the expected evacuation cost, compare it with the actual evacuation cost of the evacuation plan, and implement algorithm parameter updates based on the policy gradient algorithm.

[0055] Further, step T2 includes the following sub-steps:

[0056] T2-1: Map the disaster victim information to a specific dimension, discover the relational features in the disaster victim information, and combine the relational features into a global embedding of the disaster victim information;

[0057] T2-2: Map the shelter information to a specific dimension, explore the characteristics of the shelter information, and obtain the global embedding of the shelter information;

[0058] T2-3: Map the disaster point information to a specific dimension, explore the characteristics of the disaster point information, and obtain the global embedding of the disaster point information;

[0059] T2-4: Integrate the global embedding of the disaster victims' information, the global embedding of the shelter information and the global embedding of the disaster point information to form the algorithm's overall observation of the evacuation problem, and decode the overall observation to obtain the probability of disaster victims choosing different shelters, that is, the evacuation probability; establish a parameterized distribution based on the evacuation probability, and use sampling means to realize the allocation of disaster victims and shelters; combine the allocations to obtain the evacuation plan generated by the algorithm.

[0060] A multi-disaster point emergency evacuation planning system, comprising:

[0061] The environment module is used to simulate multi-disaster emergency evacuation scenarios and generate a large number of evacuation problem instances for training algorithm parameters;

[0062] A strategy module is used to process the information of each object in a single evacuation instance, extract features, obtain an overall observation of the problem, and generate an evacuation plan through the overall observation;

[0063] The evaluation module is used to evaluate a single evacuation instance, generate an expected evacuation cost, compare it with the actual evacuation cost of the evacuation plan, and implement algorithm parameter update based on the policy gradient algorithm.

[0064] Furthermore, the policy module includes:

[0065] A disaster victim information encoder is used to map the disaster victim information to a specific dimension, discover relational features in the disaster victim information, and combine the relational features into a global embedding of the disaster victim information;

[0066] A shelter information encoder is used to map the shelter information to a specific dimension, explore the characteristics of the shelter information, and obtain the global embedding of the shelter information;

[0067] The disaster point information encoder is used to map the disaster point information to a specific dimension, explore the characteristics of the disaster point information, and obtain the global embedding of the disaster point information;

[0068] A decision decoder is used to integrate the global embedding of the disaster victim information, the global embedding of the shelter information and the global embedding of the disaster site information into the algorithm's overall observation of the evacuation problem, and decode the overall observation to obtain the probability of disaster victims choosing different shelters, that is, the evacuation probability; a parameterized distribution is established based on the evacuation probability, and sampling is used to achieve the allocation of disaster victims and shelters; the allocation is combined to obtain the evacuation plan generated by the algorithm.

[0069] Furthermore, the disaster victim information encoder is composed of a plurality of attention layers, and the attention layers are used to receive the disaster victim information after passing through the initial embedding layer and being mapped to the specified dimensional space;

[0070] The attention layer contains two sub-layers, namely the multi-head self-attention layer and the feed-forward layer;

[0071] The multi-head self-attention layer is used to receive the disaster victim information and calculate the multi-head self-attention value; wherein the multi-head self-attention value is used to perform batch normalization processing and jump connection operations to obtain the multi-head self-attention layer output;

[0072] The feedforward layer is used to receive the output of the multi-head self-attention layer, and perform two linear mappings on the output of the multi-head self-attention layer, and activate it with a Relu function, wherein the linear mapping value of the feedforward layer is batch normalized and jump-connected to obtain the feedforward layer output, and the feedforward layer output is passed as input to the next attention layer. The above process is repeated until all the attention layers are traversed, and the final embedding of the disaster victim information can be obtained. The final embedding of the disaster victim information is integrated to obtain the global embedding of the disaster victim information.

[0073] Furthermore, the shelter information encoder and the disaster point information encoder have the same structure, including a plurality of GRU units, and the number of the GRU units is consistent with the number of shelters and the number of disaster points respectively;

[0074] The GRU unit is used to receive the initial embedding of the current shelter information or the initial embedding of the disaster point information, and the hidden state output by the previous GRU unit, calculate the hidden state of the current shelter information or the hidden state of the disaster point information, and send it to the next GRU unit;

[0075] The above process is repeated until all the GRU units are traversed, and the embedding information output by the last GRU unit is the global embedding of the shelter information or the global embedding of the disaster point information.

[0076] A multi-disaster point emergency evacuation system, consisting of a cloud-edge-end three-layer architecture, including a model training phase and an evacuation response phase.

[0077] The three-layer architecture includes a mobile terminal equipped with a locator and with communication capabilities, an edge node with computing and communication capabilities, a data server for generating evacuation instances, a machine learning server for training emergency evacuation models, and an evacuation response server for evacuation decision-making and with communication capabilities;

[0078] The model training phase is completed in the cloud. A large number of evacuation instances are generated by the data server and sent to the machine learning server. The machine learning server uses the evacuation instances to train and optimize the parameters in the evacuation model, and splits the evacuation model into an initial embedding model and a decision model. The initial embedding model will be sent to the edge node, and the decision model will be sent to the evacuation response server.

[0079] The evacuation response stage is jointly executed by cloud-edge-end. The mobile terminal will collect disaster victim information and upload it to the edge node. The edge node will collect disaster point information, complete the initial embedding of disaster victim information, and upload the disaster point information and the initial embedding of disaster victim information to the evacuation response server. The evacuation response server will integrate the disaster point information, select a safe shelter, and use the decision model to complete the embedding of disaster point information and shelter information, and integrate the disaster point information embedding, shelter information embedding and received disaster victim information embedding to obtain an overall observation of the problem. The decision model will decode the overall observation and finally obtain an evacuation plan. The evacuation plan will be sent to the edge node, and the edge node will complete the evacuation guidance of the victims.

[0080] In summary, the present invention has achieved the following technical effects:

[0081] 1. The emergency evacuation planning algorithm proposed in the present invention can effectively solve the problem of urban emergency evacuation in multi-disaster point scenarios. The algorithm can reduce the evacuation distance while reducing the number of disaster victims crossing dangerous areas, thereby improving the evacuation efficiency. The algorithm is based on deep reinforcement learning and can provide an evacuation plan without a large amount of calculation, which can effectively meet the real-time requirements of emergency evacuation.

[0082] 2. The emergency evacuation planning algorithm proposed in the present invention effectively realizes the load balancing of shelters. The present invention models the evacuation problem as a multi-objective optimization problem. By introducing the shelter load balancing mechanism, the algorithm can effectively balance the number of disaster victims received by each shelter, while ensuring the evacuation efficiency, improving the humanization of the evacuation and the efficiency of resource utilization;

[0083] 3. The disaster emergency evacuation planning system framework designed by the present invention is implemented based on the cloud-edge network. The framework is beneficial to train the evacuation model in the cloud computing center, form an evacuation plan, use the edge nodes to share the computing pressure of the cloud, and maintain the communication between the mobile terminal and the cloud. The evacuation framework proposed by the present invention effectively reduces the evacuation response time, and the location information of the disaster victims is encoded before uploading to the cloud computing center, which protects the location privacy of the disaster victims and improves the fairness of the evacuation; the emergency evacuation planning system designs a "cloud-edge-end" three-layer architecture, involving data servers, machine learning servers, evacuation response servers, edge servers, mobile terminals and disaster sensors, which effectively realizes the clear division of evacuation tasks; the emergency evacuation planning system designs two operation stages, including evacuation model training and evacuation response stages. The model training stage is coordinated by the data server and the machine learning server to complete the evacuation model training. In the evacuation response stage, the evacuation response service, edge server, mobile terminal and disaster sensor collaborate to complete the evacuation guidance of the disaster victims. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 A schematic diagram of an algorithm flow of a cloud-edge network-based disaster emergency evacuation planning method and system according to an embodiment of the present invention;

[0085] Figure 2 A schematic diagram of a process for embedding disaster victim information in a cloud-edge network-based disaster emergency evacuation planning method and system according to an embodiment of the present invention;

[0086] Figure 3 A schematic diagram of a process for embedding shelter / disaster point information in a cloud-edge network-based disaster emergency evacuation planning method and system according to an embodiment of the present invention;

[0087] Figure 4 A schematic diagram of the training phase of the "cloud-edge-end" evacuation framework model of a cloud-edge network-based disaster emergency evacuation planning method and system according to an embodiment of the present invention;

[0088] Figure 5 This is a schematic diagram of the evacuation response phase of the "cloud-edge-end" evacuation framework of the cloud-edge network-based disaster emergency evacuation planning method and system described in an embodiment of the present invention. DETAILED DESCRIPTION

[0089] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0090] This specific embodiment is only an explanation of the present invention, and it is not a limitation of the present invention. After reading this specification, those skilled in the art can make non-creative modifications to this embodiment as needed, but they are protected as long as they are within the scope of the claims of the present invention.

[0091] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0092] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0093] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0094] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0095] Embodiment:

[0096] The present invention provides a multi-disaster point emergency evacuation method, planning method, planning system and evacuation system based on cloud-edge network, including establishing an accurate mathematical model for the multi-disaster point emergency evacuation planning problem, establishing an emergency evacuation planning algorithm based on deep reinforcement learning for the multi-disaster point emergency evacuation planning problem, and establishing an evacuation system framework based on cloud-edge network for emergency evacuation planning.

[0097] Three embodiments of the present invention will be described in detail below:

[0098] Embodiment 1:

[0099] In Example 1, the mathematical model of the present invention establishes a dangerous area crossing detection mechanism and a shelter load balancing mechanism in a multi-hazard scenario.

[0100] The mathematical model includes modeling the emergency evacuation of multiple disaster points as a multi-objective optimization problem, and its optimization objective is the evacuation cost Cost (S|I). The evacuation cost is composed of the weighted sum of three components, including the evacuation distance Length (S|I), the shelter load penalty Penalty (S|I) and the number of people crossing the dangerous area Count (S|I), which is calculated as follows:

[0101]

[0102] Wherein, S represents the evacuation plan, and I represents the problem instance. The evacuation distance Length (S|I) represents the sum of the distances of all disaster victims evacuated to the corresponding shelters. The evacuation distance of a single disaster victim is the Euclidean distance, and whether to add evacuation distance redundancy for it will be selected based on whether it will cross the dangerous area during the evacuation process to achieve detour around the dangerous area. The shelter load penalty Penalty (S|I) is obtained by the shelter load balancing mechanism, reflecting the shelter load balance degree of the evacuation plan. The load penalty value of a single shelter will be calculated based on the proportional relationship between the number of disaster victims received by the shelter and its preset maximum capacity. The sum of all shelter load penalty values ​​is the shelter load penalty Penalty (S|I). The number of people crossing the dangerous area Count (S|I) is obtained by the dangerous area crossing detection mechanism, reflecting the possibility of disaster victims crossing the dangerous area during the evacuation process. The Count (S|I) is initially 0. In the evacuation plan, disaster victims are assigned to shelters. The detection mechanism will establish a line connecting the disaster victims and the shelters. If the line intersects with the dangerous area near the disaster point, it means that the evacuation route may cross the dangerous area, and the Count (S|I) increases by 1, otherwise it increases by 0. This process is repeated until all disaster victims in the evacuation plan are traversed, and at this time Count (S|I) is the number of people crossing the dangerous area.

[0103] Specifically include:

[0104] A multi-disaster point emergency evacuation method based on cloud-edge network, comprising:

[0105] Establish a dangerous area crossing detection mechanism to effectively evaluate the probability of crossing the dangerous area during the evacuation of disaster victims, and obtain the number of people crossing the dangerous area, where the number of people crossing the dangerous area represents the number of people who may cross the dangerous area;

[0106] Establish a shelter load balancing mechanism to accurately describe the shelter load situation and obtain the shelter load penalty, where the shelter load penalty represents the sum of the load penalty values ​​of all shelters. The load penalty value of a single shelter will be calculated based on the ratio between the number of refugees received by the shelter and its preset maximum capacity;

[0107] Establish a method for calculating the evacuation distance by taking into account the detour around the danger zone, and obtain the evacuation distance, wherein the evacuation distance represents the sum of the distances of all disaster victims evacuated to the corresponding shelters;

[0108] The evacuation cost is obtained according to the number of people crossing the dangerous area, the shelter load penalty, and the evacuation distance.

[0109] Specifically, in order to solve the emergency evacuation planning problem in multi-disaster scenarios, the present invention models it as a multi-objective optimization problem, which is also an NP-hard problem. Therefore, in order to simplify the problem and avoid the influence of minor factors on the problem modeling, the present invention proposes the following three assumptions:

[0110] Assumption 1: Each disaster victim can only be assigned to one shelter and has basic mobility capabilities to complete evacuation according to the evacuation plan.

[0111] Assumption 2: The disaster will not cause a devastating blow to urban traffic, and urban traffic can maintain basic traffic conditions, that is, the victims will not be unable to evacuate.

[0112] Assumption 3: When a disaster occurs, the communication between the cloud computing center and the edge nodes, and the communication between the edge nodes and the mobile terminals within their jurisdiction can be basically guaranteed, that is, the "cloud-edge-end" architecture can basically operate.

[0113] Based on the above assumptions, once a disaster occurs, people in the disaster-stricken area can be quickly evacuated to surrounding shelters through the transportation network under the guidance of the emergency evacuation system.

[0114] In the present invention, each evacuation instance I includes n v 10 victims, n s Shelters and n d The disaster site, disaster victims, shelters and disaster sites are the main objects of the problem, which are defined as follows:

[0115] Definition 1: Disaster victims v i For any disaster victim vi , the mobile terminal it carries will capture its location information Where i represents the number of the victims, and i=1,2,…,n v Without considering other information, the location information is the attribute information of the victims, that is, All disaster victims constitute the disaster victim set Vic, that is,

[0116] Definition 2: Shelters j For any shelter s j , whose attribute information includes location information and capacity information cap j , so the shelter can be expressed as Where j represents the shelter number, and j = {1,2,...,n s}, all shelters constitute the shelter set She, that is Shelter information is stored as known information in the cloud computing center for a long time.

[0117] Definition 3: Disaster point d k For any disaster point d k , whose attribute information includes location information and coverage radius r k , so the disaster point can be expressed as Where k represents the disaster point number, and k={1,2,…,n d}, all disaster points constitute the disaster point set Dis, that is Disaster information is collected by edge nodes and estimated by the cloud computing center.

[0118] The purpose of the present invention is to find an evacuation plan Where R i = {v i ,s j}, represents a set of allocations in the evacuation plan, that is, the number of victims v i Assigned to shelters j The evacuation plan needs to avoid the disaster site as much as possible. k and its coverage to maximize evacuation efficiency.

[0119] In the emergency evacuation problem faced by the present invention, there are multiple disaster points, each of which covers a certain range according to its intensity. During the evacuation process, the affected people should avoid these areas as much as possible. Therefore, the present invention designs a dangerous area crossing detection mechanism, which is as follows:

[0120] S1-1: Determine the location of the disaster point and delineate the danger zone.

[0121] Establish a plane rectangular coordinate system and regard it as the experimental area;

[0122] Disaster Point k As a point in the experimental area And set a coverage radius r according to its intensity k ;

[0123] Disaster Point k The coverage is is the center of the circle and the radius is r k Circular area of

[0124] Repeat the above steps until all disaster points are traversed.

[0125] S1-2: Single evacuation allocation R i = {v i ,s j}Detection of crossing a hazardous area.

[0126] Assign R to any group in the evacuation plan S i = {v i ,s j}, indicating that the i-th disaster victim is assigned to shelter j, and v i With s j The line segment l i,j ;

[0127] Calculate the disaster point d k With v i ,s j , l i,j The Euclidean distance

[0128] like Any one of them is smaller than the disaster point d k The coverage radius r k , indicating that the i-th disaster victim crosses the disaster point d during the evacuation to shelter j k The possibility of coverage will cross the discriminant factor Set to 1;

[0129] like Both are larger than the radius r k , indicating that the number of victims v i To the shelters j During the evacuation process, the disaster site will not be crossed. k The coverage range will cross the discriminant factor Set to 0;

[0130] Calculate the disaster point d k+1 With vi ,s j , l i,j The Euclidean distance is calculated and the above steps are repeated until all disaster points are traversed;

[0131] S1-3: Detection of the overall evacuation plan across the danger zone.

[0132] Single evacuation allocation R i = {v i ,s j After the detection of crossing the dangerous area is completed, a judgment sequence will be obtained Sum the elements in the sequence, that is

[0133] If c i,j ≥1, indicating that the number of victims v i To the shelters j During the evacuation process, there is a phenomenon of crossing the disaster coverage area, and the discriminant factor C i Set to 1;

[0134] If c i,j =0, indicating that the number of victims v i To the shelters j If there is no phenomenon of crossing the disaster coverage area during the evacuation process, the discriminant factor C i Set to 0;

[0135] Determine the next set of allocations R in the evacuation plan S i+1 Whether there is a cross-over phenomenon, repeat the above steps until the entire evacuation plan is traversed;

[0136] After traversing the evacuation plan, another discriminant factor sequence can be obtained Sum the elements of the sequence, that is, Count(S|I) represents the number of evacuation allocations that may cross the disaster coverage area in the evacuation plan S of problem instance I, and is an important variable in the multi-disaster point emergency evacuation problem.

[0137] In the present invention, in addition to reducing the evacuation distance, the emergency evacuation planning algorithm also needs to consider whether the shelter load is balanced. Therefore, the present invention designs a shelter load balancing mechanism, which is as follows:

[0138] S2-1: Establish a shelter load penalty function.

[0139] For any shelters j , in addition to location information, it also contains capacity information cap j ,In order to ensure the balance of its load, it is necessary to control the number of disaster victims it receives within a certain range;

[0140] To this end, the present invention designs a penalty function to guide the algorithm to achieve load balancing. The penalty function is a linear function, and the calculation method is as follows:

[0141]

[0142] S2-2: Calculation of single shelter load penalty value.

[0143] rec j In the evacuation plan S, shelter s j The number of disaster victims can be calculated by s j The number of occurrences of is calculated;

[0144] Indicates Shelters j Suitable receiving quantity range, is the receiving lower bound, is the upper bound of the reception, and the present invention sets it to [0.4×cap j ,0.8×cap j ];

[0145] α and β represent the penalty coefficients for too low and too high number of recipients, respectively. The sparse penalty coefficient α = -1 / 0.4cap j , congestion penalty coefficient β = 1 / 0.2cap j ;

[0146] and represents the capacity discrimination factor, if Set to 1, otherwise set to 0; if Set to 1, otherwise set to 0;

[0147] Calculate the shelter s based on the above parameters j The load penalty value is calculated as follows:

[0148]

[0149] Repeat the above steps to calculate the shelter s j+1 The load penalty value is increased until all shelters are traversed;

[0150] Adding up the load penalty values ​​of all shelters, we can get the load penalty Penalty (S|I) of the evacuation plan S shelter.

[0151] In the present invention, the emergency evacuation planning algorithm needs to improve the overall efficiency of emergency evacuation as much as possible in addition to coping with the impact of multiple disaster points on evacuation and achieving shelter load balancing. Therefore, the present invention introduces the evacuation distance Length (S|I) as the third important variable into the objective function of the problem.

[0152] The present invention will allocate R i = {v i ,s j}V i With shelters j Straight line distance As a disaster victim i However, although using straight-line distance as the evacuation distance reduces the complexity of calculation, it will lead to partial allocation of R i It is inevitable to cross the disaster area. For this reason, the present invention does not prohibit the evacuation path from crossing the disaster area, but adds route redundancy to the evacuation plan with crossing situations, as follows:

[0153] When R i = {v i ,s j}'s cross-judgment factor C i When it is 1, it indicates that R is allocated. i This will cause disaster victims to cross dangerous areas during the evacuation process. Therefore, in order to cope with complex disaster environments or achieve detours, it is necessary to increase route redundancy, that is, to increase the allocation R i The corresponding evacuation distance is calculated as follows:

[0154]

[0155] To allocate R i = {v i ,s j}Evacuation distance of disaster victims, For disaster victims i With shelters j The straight-line distance is the Euclidean distance, where γ is the redundancy factor, and C i For R i The cross-judgment factor.

[0156] Then calculate the next allocation R i+1 The corresponding evacuation distance is calculated until the evacuation plan S is traversed. By adding up all the assigned evacuation distances, the total evacuation distance Length(S|I) of the evacuation plan can be obtained, that is:

[0157]

[0158] By combining Length(S|I), Penalty(S|I) and Count(S|I), the evacuation cost Cost(S|I) of the evacuation scheme S can be obtained. The calculation function of the evacuation cost is the objective function of the multi-disaster point emergency evacuation planning algorithm proposed in the present invention, and the calculation method is as follows:

[0159]

[0160] in, and They represent the weights of the three components in the objective function, so it can be said that the emergency evacuation problem is modeled as a multi-objective optimization problem. The goal of the evacuation algorithm optimization is to find an evacuation plan S for the problem instance I, so that the objective function value Cost(S|I) is as small as possible.

[0161] In summary, the mathematical model includes: modeling the multi-disaster emergency evacuation planning problem as a multi-objective optimization problem; three assumptions are proposed to simplify the multi-disaster emergency evacuation planning problem and eliminate the influence of minor factors on the problem modeling; standard definitions are made for the main objects such as disaster victims, shelters and disaster points in the multi-disaster emergency evacuation planning problem; a dangerous area crossing detection mechanism is established for the multi-disaster emergency evacuation planning problem to achieve an effective assessment of the probability of crossing the dangerous area during the evacuation of disaster victims; a shelter load balancing mechanism is established for the multi-disaster emergency evacuation planning problem to achieve an accurate description of the shelter load situation; a method for calculating the evacuation distance that considers detouring the dangerous area is established for the multi-disaster emergency evacuation planning problem; and an objective function that comprehensively considers the evacuation distance, shelter load and the probability of crossing the dangerous area is designed for the multi-disaster emergency evacuation planning problem.

[0162] Embodiment 2:

[0163] In Example 2, the algorithm of the present invention includes an environment module, a strategy module and an evaluation module; the environment module can generate multiple evacuation problem instances for model training; the strategy module includes a disaster victim information encoder, a shelter information encoder, a disaster point information encoder and a decision decoder; the disaster victim encoder can embed the disaster victim information and obtain a disaster victim feature vector; the shelter information encoder can embed the shelter information and obtain a shelter feature vector; the disaster point information encoder can embed the disaster point information and obtain a disaster point feature vector; the decision decoder can integrate the disaster victim feature vector, the shelter feature vector and the disaster point feature vector to form an overall observation of the evacuation problem and decode it, obtain the action probability distribution of the intelligent agent, then sample the distribution, obtain the action, and finally integrate the intelligent agent action to obtain the evacuation plan; the evaluation module can evaluate the evacuation problem and the evacuation plan, and realize the update of the algorithm parameters.

[0164] The emergency evacuation planning algorithm is implemented based on a deep reinforcement learning algorithm, and includes an environment module, a strategy module, and an evaluation module. The environment module will generate a number of multi-disaster point emergency evacuation instances for the strategy module to solve, so as to implement parameter updates of the strategy module. The multi-disaster point emergency evacuation instances include a certain amount of disaster victim information, shelter information, and disaster point information.

[0165] The strategy module includes a disaster victim information encoder, a shelter information encoder, a disaster site information encoder and a decision decoder. The disaster victim information encoder is a hybrid neural network. The hybrid neural network can map the disaster victim information to a specific dimension, explore the relationship features in the disaster victim information, and combine the features into a global embedding of the disaster victim information. The shelter information encoder is a neural network based on a gated recurrent unit (GRU), which can map the shelter information to a specific dimension, explore the features of the shelter information, and obtain the global embedding of the shelter information. The disaster site information encoder is a neural network based on a gated recurrent unit (GRU), which can map the disaster site information to a specific dimension, explore the features of the disaster site information, and obtain the global embedding of the disaster site information. The decision decoder is a neural network based on an attention mechanism, which can integrate the global embedding of disaster victims, shelters and disaster sites into the algorithm's overall observation of the evacuation problem, and can decode this observation to obtain the probability of disaster victims choosing different shelters, that is, the evacuation probability. A parameterized distribution is established based on the evacuation probability, and the allocation (Allocation) of disaster victims and shelters is achieved by sampling means. By combining the allocations, the evacuation plan (Solution) generated by the algorithm can be obtained. The evaluation module is a hybrid neural network, which is used to evaluate the problem instance and assist in updating the algorithm parameters. The hybrid neural network will receive the problem instance I generated by the environment module, complete the embedding of the problem instance, and output the corresponding estimated evacuation cost pred(I). The evacuation algorithm will calculate the actual evacuation cost Cost(S|I) of the evacuation plan S it generates, and optimize the evacuation algorithm parameters by minimizing the mean square error between the estimated evacuation cost pred(I) and the actual evacuation cost Cost(S|I).

[0166] Specifically include:

[0167] A multi-disaster point emergency evacuation planning method based on cloud-edge network, comprising:

[0168] T1: Simulate multi-disaster emergency evacuation scenarios to generate a large number of evacuation problem instances for training algorithm parameters;

[0169] T2: Perform data processing on the information of each object in a single evacuation instance and implement feature extraction to obtain an overall observation of the problem, and generate an evacuation plan based on the overall observation;

[0170] T3: Evaluate a single evacuation instance, generate the expected evacuation cost, compare it with the actual evacuation cost of the evacuation plan, and implement algorithm parameter updates based on the policy gradient algorithm.

[0171] Further, step T2 includes the following sub-steps:

[0172] T2-1: Map the disaster victim information to a specific dimension, discover the relational features in the disaster victim information, and combine the relational features into a global embedding of the disaster victim information;

[0173] T2-2: Map the shelter information to a specific dimension, explore the characteristics of the shelter information, and obtain the global embedding of the shelter information;

[0174] T2-3: Map the disaster point information to a specific dimension, explore the characteristics of the disaster point information, and obtain the global embedding of the disaster point information;

[0175] T2-4: Integrate the global embedding of the disaster victims' information, the global embedding of the shelter information and the global embedding of the disaster point information to form the algorithm's overall observation of the evacuation problem, and decode the overall observation to obtain the probability of disaster victims choosing different shelters, that is, the evacuation probability; establish a parameterized distribution based on the evacuation probability, and use sampling means to realize the allocation of disaster victims and shelters; combine the allocations to obtain the evacuation plan generated by the algorithm.

[0176] A multi-disaster point emergency evacuation planning system based on cloud-edge network, including:

[0177] The environment module is used to simulate multi-disaster emergency evacuation scenarios and generate a large number of evacuation problem instances for training algorithm parameters;

[0178] A strategy module is used to process the information of each object in a single evacuation instance, extract features, obtain an overall observation of the problem, and generate an evacuation plan through the overall observation;

[0179] The evaluation module is used to evaluate a single evacuation instance, generate an expected evacuation cost, compare it with the actual evacuation cost of the evacuation plan, and implement algorithm parameter update based on the policy gradient algorithm.

[0180] Furthermore, the policy module includes:

[0181] A disaster victim information encoder is used to map the disaster victim information to a specific dimension, discover relational features in the disaster victim information, and combine the relational features into a global embedding of the disaster victim information;

[0182] A shelter information encoder is used to map the shelter information to a specific dimension, explore the characteristics of the shelter information, and obtain the global embedding of the shelter information;

[0183] The disaster point information encoder is used to map the disaster point information to a specific dimension, explore the characteristics of the disaster point information, and obtain the global embedding of the disaster point information;

[0184] A decision decoder is used to integrate the global embedding of the disaster victim information, the global embedding of the shelter information and the global embedding of the disaster site information into the algorithm's overall observation of the evacuation problem, and decode the overall observation to obtain the probability of disaster victims choosing different shelters, that is, the evacuation probability; a parameterized distribution is established based on the evacuation probability, and sampling is used to achieve the allocation of disaster victims and shelters; the allocation is combined to obtain the evacuation plan generated by the algorithm.

[0185] Furthermore, the disaster victim information encoder is composed of a plurality of attention layers, and the attention layers are used to receive the disaster victim information after passing through the initial embedding layer and being mapped to the specified dimensional space;

[0186] The attention layer contains two sub-layers, namely the multi-head self-attention layer and the feed-forward layer;

[0187] The multi-head self-attention layer is used to receive the disaster victim information and calculate the multi-head self-attention value; wherein the multi-head self-attention value is used to perform batch normalization processing and jump connection operations to obtain the multi-head self-attention layer output;

[0188] The feedforward layer is used to receive the output of the multi-head self-attention layer, and perform two linear mappings on the output of the multi-head self-attention layer, and activate it with a Relu function, wherein the linear mapping value of the feedforward layer is batch normalized and jump-connected to obtain the feedforward layer output, and the feedforward layer output is passed as input to the next attention layer. The above process is repeated until all the attention layers are traversed, and the final embedding of the disaster victim information can be obtained. The final embedding of the disaster victim information is integrated to obtain the global embedding of the disaster victim information.

[0189] Furthermore, the shelter information encoder and the disaster point information encoder have the same structure, including a plurality of GRU units, and the number of the GRU units is consistent with the number of shelters and the number of disaster points respectively;

[0190] The GRU unit is used to receive the initial embedding of the current shelter information or the initial embedding of the disaster point information, and the hidden state output by the previous GRU unit, calculate the hidden state of the current shelter information or the hidden state of the disaster point information, and send it to the next GRU unit;

[0191] The above process is repeated until all the GRU units are traversed, and the embedding information output by the last GRU unit is the global embedding of the shelter information or the global embedding of the disaster point information.

[0192] Specifically, refer to Figure 1-Figure 3 , which is the second embodiment of the present invention. This embodiment is different from the first embodiment in that it provides a detailed implementation process of an emergency evacuation planning algorithm based on deep reinforcement learning for multi-disaster scenarios, specifically including:

[0193] The emergency evacuation planning algorithm in this embodiment is based on deep reinforcement learning and mainly includes three modules, namely, environment module, strategy module and evaluation module. Figure 1 .

[0194] The environment module will generate a large number of problem instances for model training. It is the environment that interacts with the agent. The generation process of an evacuation instance is as follows:

[0195] T1-1: The environment module evenly generates the locations of disaster victims within the [0,1]×[0,1] evacuation range.

[0196] T1-2: In order to simplify the problem and make the evacuation effect more obvious, this paper separates the disaster site and the shelter. The specific operations are as follows:

[0197] A [0.35, 0.65] × [0.35, 0.65] rectangular area was delineated in the center of the experimental area, and the coordinates of the disaster points were randomly generated within this area;

[0198] The coverage radius of the disaster point is intermittently taken in the range of [0.05, 0.15];

[0199] The location of the shelter is randomly generated outside a circle with (0.5, 0.5) as the center and a radius of 0.3;

[0200] It is required that the linear position of any two shelters is not less than 0.1, so that the shelters are evenly distributed at the edge of the evacuation area, and there is no need to consider the situation where the shelters are covered by the disaster points;

[0201] Shelter capacity cap is A random integer in the range makes the shelter capacity redundant compared to the number of disaster victims, which is consistent with common shelter designs;

[0202] The suitable reception capacity interval of the shelter is [0.4×cap, 0.8×cap], the sparse penalty coefficient α = -1 / 0.4cap, and the crowded penalty coefficient β = 1 / 0.2cap.

[0203] T1-3: Repeat step T1-2 until enough evacuation instances are generated, and send the evacuation instances to the policy module, which will generate the corresponding evacuation plan.

[0204] The strategy module receives problem instances from the environment module, including disaster victim information, shelter information, and disaster site information, and derives an evacuation plan based on this information. The strategy module adopts an "encoder-decoder" architecture, that is, it uses several encoders to encode the original data, and integrates the encoded data to form an overall observation of the problem. Finally, the decoder is used to decode the problem observation and generate an evacuation plan. The strategy module of the present invention includes a disaster victim encoder, a shelter / disaster site encoder, and a decision decoder. The specific process of the strategy module generating an evacuation route is as follows:

[0205] Reference Figure 2 ,The strategy module first uses the disaster victim encoder to embed the disaster victim information and obtains the global embedding of the disaster victim information. The specific operations are as follows:

[0206] T2-1-1: Use linear transformation to map it to the specified dimension as the initial embedding of the model. The calculation method is as follows:

[0207]

[0208] in, For the victims i The initial embedding of w0 represents the weight and b0 represents the bias. Although this mapping method is simple, it still takes a lot of time when the problem scale is large. To ensure the real-time operation of the algorithm, the initial embedding operation of the disaster victims is undertaken by the edge server, which is the initial embedding model delegated to the edge server as mentioned above.

[0209] T2-1-2: Reference Figure 2 The disaster victim encoder constructs multiple attention layers to extract the relationship features between the disaster victim nodes from the initial embedding of the disaster victims. The specific operations are as follows:

[0210] Each attention layer contains two sub-layers, namely the Multi-head Self-attention Layer (MHSA Layer) and the Feed Forward Layer (FF Layer), and batch normalization (BN) and skip connection (Skip-Connection) are added to each sub-layer;

[0211] The operation process of the multi-head self-attention layer in the attention layer is as follows:

[0212] Calculate the Query, Key, and Value in the attention mechanism as follows:

[0213]

[0214] in, They correspond to Query, Key and Value in the attention mechanism, and their dimensions are dim q 、dim k and dim v , for the convenience of calculation, let dim q =dim k =dim v =dim embed / n h , dim embed is the initial embedding dimension;

[0215] Calculate the value of a single attention head as follows:

[0216]

[0217] in, represents the value of the pth attention head in the mth attention layer. There are n attention heads in an attention layer. h The algorithm multiplies the input of the softmax function by To make its gradient more reasonable, are all weight parameter matrices.

[0218] All After integration, batch normalization and skip connection processing, the calculation is as follows:

[0219]

[0220] in, represents the output of the multi-head self-attention sub-layer in the m-th attention layer, It represents the embedding of the i-th disaster victim through this layer. m-1 represents the initial embedding of the disaster-affected person node or the output of the previous attention layer. BN represents batch normalization. is the weight parameter matrix.

[0221] The operation process of the feedforward layer in the attention layer is as follows:

[0222] The output of the multi-head self-attention layer is linearly mapped and activated with the Relu function, which is calculated as follows:

[0223]

[0224] Use the Dropout operation to put some neurons to sleep to avoid overfitting of the neural network

[0225] Perform the second linear mapping and calculate as follows:

[0226]

[0227] The mapped values ​​are batch normalized and skipped, and the calculation is as follows:

[0228]

[0229] in, is the embedding of the ith disaster victim after the mth attention layer, Dropout is the neuron release operation, Relu is the activation function, is the weight value, is the bias value.

[0230] T2-1-3: The disaster victim encoder is set to n l In other words, the input victim information will repeat step S2 for a total of n l times, and the final embedding of each disaster victim node

[0231] T2-1-4: Based on the final embedding of each disaster victim node, calculate the context embedding of the disaster victim node in the problem instance, that is, the global embedding, which is calculated as follows:

[0232]

[0233] Reference Figure 3 After embedding the disaster victims’ information, the strategy module will use the shelter / disaster site encoder to process the shelter and disaster site information and obtain the global embedding of the shelter and disaster site. The specific operations are as follows:

[0234] T2-2-1: Use linear transformation to map the original shelter information to the specified dimension as the initial embedding of the model. The calculation method is as follows:

[0235]

[0236] in, represents the initial embedding of the j-th shelter information, s j Initial information for the shelter, represents the weight, Indicates bias.

[0237] T2-2-2: Reference Figure 3 ,The shelter / disaster site encoder is based on the GRU neural network, which contains ,multiple GRU units, which can effectively extract the dependencies in the data ,sequence. And fewer parameters can reduce the computational ,overhead of model training and improve the model convergence ,speed.

[0238] Here, the number of GRU units is the same as the number of shelters. Each GRU unit will calculate the feature information of the corresponding shelter. The calculation method is as follows:

[0239]

[0240] H j =Up j ⊙H j-1 +(1-Up j )⊙Co j ;

[0241] in, Represents the hidden state of the previous GRU unit output, Re j Reset Gate is used to control the degree of neglect of the hidden state of the previous GRU unit. j The update gate (UpdateGate) is used to control the degree of retention of the hidden state of the previous GRU unit. j The control gate is used to control the degree of retention of the current GRU unit state, and ⊙ represents element multiplication.

[0242] T2-2-3: Pass the hidden state output by the current GRU unit to the subsequent GRU unit, and repeat the calculation process of S2 until the last GRU unit is calculated.

[0243] T2-2-4: Embedding information output by the last GRU unit This is the global embedding of the shelter.

[0244] T2-2-5: Use linear transformation to map the original information of the disaster point to the specified dimension as the initial embedding of the model. The calculation method is as follows:

[0245]

[0246] in, represents the initial embedding of the k-th disaster point information, d k Initial information for the shelter, represents the weight, Indicates bias.

[0247] T2-2-6: Build a second shelter / disaster point encoder. The number of GRU units is the same as the number of disaster points. Each GRU unit will calculate the feature information of the corresponding disaster point. The calculation method is as follows:

[0248]

[0249] H k =Up k ⊙H k-1 +(1-Up k )⊙Co k ;

[0250] T2-2-7: Pass the hidden state output by the current GRU unit to the subsequent GRU unit, and repeat the calculation process of T2-2-6 until the last GRU unit is calculated.

[0251] T2-2-8: Embedding information output by the last GRU unit This is the global embedding of the disaster point.

[0252] Reference Figure 1 After embedding the information of disaster victims, shelters and disaster sites, the strategy module will construct an overall observation of the problem based on their global embedding, and then use the decoder to decode it to obtain the evacuation plan. The specific process is as follows:

[0253] T2-4-1: Based on the global embedding of disaster victims, shelters and disaster sites, the overall observation of the evacuation problem is constructed as follows:

[0254]

[0255] Among them, O I represents the overall observation of problem instance I, Represent the global embeddings of disaster victims, shelters, and disaster sites, respectively.

[0256] T2-4-2: Decoder is established based on pointer network. The pointer network designed by the present invention is a simplified attention mechanism. I Decode and assign appropriate shelters to each disaster victim in turn. The decoder is calculated as follows:

[0257]

[0258] Qu j =W Qu O I ;

[0259]

[0260] in, is the decoding result of the decoder on the context embedding, Qu j Query, observe the problem as a whole I Calculated, it represents the algorithm's observation value for shelter j, Ke i is the Key, representing the embedded value of victim i.

[0261] T2-4-3: Use the softmax function to calculate the evacuation probability of victim i, that is, the probability that the i-th victim chooses the j-th shelter. The calculation method is as follows:

[0262]

[0263] T2-4-4: Construct a parameterized distribution based on the evacuation probability, and obtain the shelter numbers corresponding to the disaster victims through sampling to form the evacuation distribution R i = {v i ,s j}, the higher the evacuation probability of a shelter, the higher its probability of being sampled.

[0264] T2-4-5: Repeat steps T2-4-2 to T2-4-4 to calculate the evacuation probability of the next disaster victim and generate evacuation assignments until all disaster victims have been assigned.

[0265] T2-4-6: Evacuate all assigned R i = {v i ,s j} combination to form the evacuation plan of evacuation instance I

[0266]

[0267] Reference Figure 1 ,After the policy module generates the evacuation plan, the evaluation module will evaluate the problem instance and assist the algorithm parameter update. The specific process is as follows:

[0268] P1: Establish an instance evaluation neural network based on the Attention mechanism, and obtain the attention value based on the overall observation of the instance. The calculation method is as follows:

[0269]

[0270] Among them, O I is the overall observation of instance I, VI represents the parameter matrix and tanh represents the activation function.

[0271] P2: The attention value is combined with the overall observation and fed into two linear layers, calculated as follows:

[0272]

[0273] Among them, pred(I) represents the expected value of the evaluation module for the problem instance, and Relu represents the activation function.

[0274] P3: Establish a parameter optimization function. For any problem instance I, let θ represent the neural network parameters. Then the training goal is:

[0275]

[0276] Among them, S represents the evacuation plan generated by the model, p θ (·|I) represents the probability of instance I generating a solution, which is different from the probability of the decoder generating a solution. related.

[0277] P4: Establish parameter optimization gradient based on Policy Gradient (PG), as follows:

[0278]

[0279] P5: Adjust the gradient algorithm for batch training strategy. That is, during the training process, a batch of problem instances are usually sent to the model for training. Therefore, the gradient can be calculated using the average value of the Monte Carlo sampling in this batch of problem instances. The calculation method is as follows:

[0280]

[0281] Where B represents the number of instances, i.e., the batch size. The model starts with a random policy to learn the parameters θ of the policy module and uses the Adam optimizer to iteratively optimize the instances generated in real time.

[0282] In summary, the emergency evacuation planning algorithm establishes a disaster victim encoder, which is a hybrid neural network based on a multi-head self-attention mechanism and a feedforward neural network; and establishes a shelter / disaster point encoder, which is a gated recurrent unit (Gated Recurrent Unit). A hybrid neural network of GRU (GRU) is established; a decision decoder is a hybrid neural network based on a pointer network; the decision decoder can embed and combine disaster victims, shelters and disaster points as a whole to calculate the overall observation of the evacuation problem; the decision decoder sends the overall observation of the problem and the final embedding of the disaster victims into the hybrid neural network, and uses the pointer network decoding to obtain the evacuation probability of the disaster victims, that is, the probability of the disaster victims choosing different shelters; the decision decoder will form a probability distribution based on the evacuation probability of the disaster victims, and use sampling to form an evacuation allocation; the evacuation allocation is integrated to obtain the evacuation plan; the evaluation module is a hybrid neural network based on the attention mechanism; the evaluation module initially embeds the problem information, maps it to the specified dimension, and sends it to the hybrid neural network to generate the estimated evacuation cost of the problem; the evaluation module will compare the actual evacuation cost of the evacuation plan generated by the calculation policy module with the estimated evacuation cost; the evaluation module will use the policy gradient (PG) algorithm to achieve the descent of the algorithm parameter gradient according to the difference between the actual evacuation cost and the estimated evacuation cost, and finally complete the algorithm parameter training.

[0283] Embodiment 3:

[0284] In Example 3, the system framework of the present invention is a three-level "cloud-edge-end" architecture, which includes two stages: model training and evacuation response. The model training stage is completed in the cloud computing center, and the cloud computing center implements the training of the evacuation model. The evacuation response stage is completed by the cloud computing center, the edge server and the mobile terminal in collaboration. The mobile terminal collects data, the edge server implements initial data embedding, and the cloud computing center generates an evacuation plan.

[0285] The evacuation system based on the cloud-edge network consists of a three-layer architecture of "cloud-edge-end". Five types of entities are involved in the three-layer architecture of "cloud-edge-end", including mobile terminals (such as smart phones) equipped with locators and with communication capabilities, edge nodes with computing and communication capabilities, data servers for generating evacuation instances, machine learning servers for training emergency evacuation models, and evacuation response servers for evacuation decisions and with communication capabilities. The evacuation system based on the cloud-edge network includes two operating stages, including a model training stage and an evacuation response stage. The model training stage is completed in the cloud, and a large number of evacuation instances are generated by the data server and sent to the machine learning server. The machine learning server will use the evacuation instances to train and optimize the parameters in the evacuation model. The evacuation model is split into an initial embedding model and a decision model. The initial embedding model will be sent to the edge node, and the decision model will be sent to the evacuation response server. The evacuation response stage is jointly executed by "cloud-edge-end", and the mobile terminal will collect disaster victim information and upload it to the edge node. The edge node will collect disaster point information, complete the initial embedding of disaster victim information, and upload the disaster point information and disaster victim information to the evacuation response server. The evacuation response server will integrate the disaster point information, select a safe shelter, and use the decision model to complete the embedding of disaster point information and shelter information, and integrate the disaster point information embedding, shelter information embedding and received disaster victim information embedding to obtain an overall observation of the problem. The decision model will decode the overall observation and finally obtain an evacuation plan. The evacuation plan will be sent to the edge node, and the edge node will complete the evacuation guidance of the disaster victims.

[0286] A multi-disaster point emergency evacuation system based on cloud-edge network, which consists of a three-layer architecture of cloud-edge-end, including model training stage and evacuation response stage.

[0287] The three-layer architecture includes a mobile terminal equipped with a locator and with communication capabilities, an edge node with computing and communication capabilities, a data server for generating evacuation instances, a machine learning server for training emergency evacuation models, and an evacuation response server for evacuation decision-making and with communication capabilities;

[0288] The model training phase is completed in the cloud. A large number of evacuation instances are generated by the data server and sent to the machine learning server. The machine learning server uses the evacuation instances to train and optimize the parameters in the evacuation model, and splits the evacuation model into an initial embedding model and a decision model. The initial embedding model will be sent to the edge node, and the decision model will be sent to the evacuation response server.

[0289] The evacuation response stage is jointly executed by cloud-edge-end. The mobile terminal will collect disaster victim information and upload it to the edge node. The edge node will collect disaster point information, complete the initial embedding of disaster victim information, and upload the disaster point information and the initial embedding of disaster victim information to the evacuation response server. The evacuation response server will integrate the disaster point information, select a safe shelter, and use the decision model to complete the embedding of disaster point information and shelter information, and integrate the disaster point information embedding, shelter information embedding and received disaster victim information embedding to obtain an overall observation of the problem. The decision model will decode the overall observation and finally obtain an evacuation plan. The evacuation plan will be sent to the edge node, and the edge node will complete the evacuation guidance of the victims.

[0290] Specifically, refer to Figure 4-Figure 5 , which is the third embodiment of the present invention. This embodiment is different from the first embodiment in that it provides an implementation method of a disaster emergency evacuation planning system based on a cloud-edge network, specifically comprising:

[0291] The disaster emergency evacuation planning system based on cloud-edge network in this embodiment involves a three-layer architecture of "cloud-edge-end", including two stages: model training and evacuation response.

[0292] The "cloud-edge-end" three-layer architecture of the disaster emergency evacuation planning system based on the cloud-edge network described in this embodiment involves the following entities, and their functions are as follows:

[0293] H1: Cloud Layer, used to implement the training of emergency evacuation planning algorithms and to generate evacuation plans in a timely manner to guide disaster victims to evacuate when a disaster occurs. The entities included in the cloud are as follows:

[0294] The data server is equipped with the environment module in the evacuation algorithm, which can generate a large number of evacuation problem instances for training the evacuation algorithm;

[0295] The machine learning server is equipped with the strategy module and evaluation module of the evacuation algorithm. It has powerful computing capabilities and can realize effective training of the evacuation algorithm.

[0296] The evacuation response server is equipped with a decision model split from the policy module. It can receive various types of data and generate evacuation plans, and has the ability to communicate with edge servers.

[0297] H2: Edge Layer, which is composed of a large number of edge servers and assists the cloud layer in completing evacuation planning. The details are as follows:

[0298] Some edge servers are equipped with the initial embedding model split from the policy module, which is used to receive disaster victim information from the terminal layer and implement the initial embedding of disaster victim information;

[0299] Some edge servers receive evacuation plans from the cloud and issue evacuation directions to terminals;

[0300] Some edge servers receive disaster point information from the terminal layer and integrate and upload it to the cloud layer.

[0301] H3: Terminal Layer, which consists of a large number of mobile terminals and sensors, is used to collect information about disaster sites and victims, as follows:

[0302] Mobile terminals, carried by disaster victims, have locators and the ability to communicate with edge nodes. They can collect the location information of disaster victims and upload it to the edge server. They can also receive evacuation instructions from the edge server and guide the evacuation of disaster victims.

[0303] Disaster sensors, which have the ability to detect disasters and communicate with edge nodes, can collect disaster information and upload it to edge servers.

[0304] Reference Figure 4 In the operation process of the cloud-edge network-based disaster emergency evacuation planning system in this embodiment, a model training phase is included. In the model training phase, the model operation is as follows:

[0305] G1: The data server equipped with the environment module will generate a series of problem instances, including the generated information of disaster victims, shelters and disaster points, as well as the corresponding reward calculation rules.

[0306] G2: The data server sends the generated problem instance to the machine learning server.

[0307] G3: The machine learning server is equipped with a strategy module and an evaluation module, in which the decision module can embed the problem input and input it into the decision module to obtain the evacuation plan.

[0308] G4: The evaluation module in the machine learning server will calculate the actual cost of the evacuation plan and the estimated cost of the problem. Based on the actual cost and the estimated cost, the gradient loss is calculated.

[0309] G5: The machine learning server updates the model parameters in an iterative manner to reduce the gradient and thus achieve algorithm optimization.

[0310] G6: The machine learning server splits the trained policy module into an initial embedding model and a decision model. The initial embedding model will be transmitted to some edge servers, and the decision model will be transmitted to the evacuation response server.

[0311] Reference Figure 5In the present embodiment, the cloud-edge network-based disaster emergency evacuation planning system includes an evacuation response phase during its operation. The model operation in the evacuation response phase is as follows:

[0312] K1: When a disaster occurs, the terminal layer disaster sensor detects the disaster outbreak and transmits the disaster information to the edge server.

[0313] K2: The edge node collects information about the disaster point, including the location and intensity of the disaster point, and uploads the disaster information to the evacuation response server.

[0314] K3: The evacuation response server will analyze the disaster information, determine the location and coverage of the disaster center, define the evacuation range, retrieve the shelter information, and issue evacuation orders.

[0315] K4: After receiving the evacuation order, the edge node will issue a disaster alert to the mobile terminals of the disaster victims within its coverage area and collect the location information of the disaster victims through the mobile terminals.

[0316] K5: The edge node uses the initial embedding model to initially embed the collected disaster victim information and upload it to the evacuation response server.

[0317] K6: The evacuation response server uses the disaster victim information encoder to perform more complex feature extraction on the disaster victim information and obtain the overall embedding of the disaster victim information.

[0318] K7: The evacuation response server uses the shelter / disaster encoder to embed the disaster information and the shelter information to obtain the overall embedding of the shelter information and the overall embedding of the disaster point information.

[0319] K8: The evacuation response server integrates the overall embedding of disaster victim information, shelter information, and disaster site information to obtain a holistic view of the current evacuation problem.

[0320] K9: The evacuation response server uses the decision decoder to decode the overall observation to obtain the evacuation plan, and sends the evacuation plan to the edge server.

[0321] K10: The edge server forms an evacuation route based on the evacuation plan and sends the route to the mobile terminals carried by the victims, which then guide the victims to evacuate.

[0322] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multi-disaster point emergency evacuation method, characterized by: include, Establish a dangerous area crossing detection mechanism to effectively evaluate the probability of crossing the dangerous area during the evacuation of disaster victims, and obtain the number of people crossing the dangerous area, where the number of people crossing the dangerous area represents the number of people who may cross the dangerous area; Establish a shelter load balancing mechanism to accurately describe the shelter load situation and obtain the shelter load penalty, where the shelter load penalty represents the sum of the load penalty values ​​of all shelters. The load penalty value of a single shelter will be calculated based on the ratio between the number of refugees received by the shelter and its preset maximum capacity; Establish a method for calculating the evacuation distance by taking into account the detour around the danger zone, and obtain the evacuation distance, wherein the evacuation distance represents the sum of the distances of all disaster victims evacuated to the corresponding shelters; Obtaining an evacuation cost according to the number of people crossing the danger zone, the shelter load penalty, and the evacuation distance; The dangerous area crossing detection mechanism comprises the following sub-steps: S1-1: Determine the location of the disaster point and delineate the danger zone; Establish a plane rectangular coordinate system and regard it as the experimental area; Disaster Point k As a point in the experimental area And set a coverage radius r according to its intensity k ; Disaster Point k The coverage is is the center of the circle and the radius is r k Circular area of Repeat the above steps until all disaster points are traversed; S1-2: Single evacuation allocation R i = {v i ,s j }Detection of crossing dangerous areas; Assign R to any group in the evacuation plan S i = {v i ,s j }, indicating that the i-th disaster victim is assigned to shelter j, and v i With s j The line segment l i,j ; where v i Indicates the victims, j Indicates refuge, R i It means that the i-th disaster victim is assigned to shelter j; Calculate the disaster point d k With v i ,s j , l i,j The Euclidean distance like Any one of them is smaller than the disaster point d k The coverage radius r k , indicating that the i-th disaster victim crosses the disaster point d during the evacuation to shelter j k The possibility of coverage will cross the discriminant factor Set to 1; like Both are larger than the radius r k , indicating that the number of victims v i To the shelters j During the evacuation process, the disaster site will not be crossed. k The coverage range will cross the discriminant factor Set to 0; Calculate the disaster point d k+1 With v i ,s j , l i,j The Euclidean distance is calculated and the above steps are repeated until all disaster points are traversed; S1-3: Detection of the overall evacuation plan crossing the dangerous area; Single evacuation allocation R i = {v i ,s j After the detection of crossing the dangerous area is completed, a judgment sequence will be obtained Sum the elements in the sequence, that is If c i,j ≥1, indicating that the number of victims v i To the shelters j During the evacuation process, there is a phenomenon of crossing the disaster coverage area, and the discriminant factor C i Set to 1; If c i,j =0, indicating that the number of victims v i To the shelters j If there is no phenomenon of crossing the disaster coverage area during the evacuation process, the discriminant factor C i Set to 0; Determine the next set of allocations R in the evacuation plan S i+1 Whether there is a cross-over phenomenon, repeat the above steps until the entire evacuation plan is traversed; After traversing the evacuation plan, another discriminant factor sequence is obtained Sum the sequence elements, i.e. the number of people crossing the dangerous area The shelter load balancing mechanism comprises the following sub-steps: Calculation of single shelter load penalty: Among them, rec j In the evacuation plan S, shelter s j The number of disaster victims is calculated by evacuation plan S, s j The number of occurrences of cap is calculated; j Indicates capacity information; Indicates Shelters j Suitable receiving quantity range, is the receiving lower bound, is the upper bound of the receiver, set it to [0.4×cap j ,0.8×cap j ]; α and β represent the penalty coefficients for too low and too high number of recipients, respectively. The sparse penalty coefficient α = -1 / 0.4cap j , congestion penalty coefficient β = 1 / 0.2cap j ; and represents the capacity discrimination factor, if Set to 1, otherwise set to 0; if Set to 1, otherwise set to 0; Repeat the above steps to calculate the shelter s j+1 The load penalty value is increased until all shelters are traversed; Add up the load penalty values ​​of all shelters to get the load penalty Penalty (S|I) of evacuation plan S: The method for calculating the evacuation distance by considering detouring the dangerous area comprises the following sub-steps: When R i = {v i ,s j }'s cross-judgment factor C i When it is 1, it indicates that R is allocated. i This will cause disaster victims to cross dangerous areas during the evacuation process, increasing route redundancy, that is, increasing the allocation R i The corresponding evacuation distance is calculated as follows: To allocate R i = {v i ,s j }Evacuation distance of disaster victims, For disaster victims i With shelters j The straight-line distance is the Euclidean distance, where γ is the redundancy factor, and C i For R i The cross-judgment factor; Then calculate the next allocation R i+1 The corresponding evacuation distance until the evacuation plan S is traversed; Add up all the assigned evacuation distances to obtain the total evacuation distance Length(S|I) of the evacuation plan:

2. The multi-disaster point emergency evacuation method according to claim 1, characterized in that: The evacuation cost = number of people crossing the dangerous area + the shelter load penalty + evacuation distance, that is: Among them, Cost(S|I) is the evacuation cost, Length(S|I) is the evacuation distance, Penalty(S|I) is the shelter load penalty, Count(S|I) is the number of people crossing the dangerous area, and Represent the weights of the three components in the objective function respectively.

3. A multi-disaster point emergency evacuation planning method, characterized by: The method for emergency evacuation of multiple disaster points as claimed in any one of claims 1 to 2 comprises the following sub-steps: T1: Simulate multi-disaster emergency evacuation scenarios to generate a large number of evacuation problem instances for training algorithm parameters; T2: Perform data processing on the information of each object in a single evacuation instance and implement feature extraction to obtain an overall observation of the problem, and generate an evacuation plan based on the overall observation; T3: Evaluate a single evacuation instance, generate an expected evacuation cost, compare it with the actual evacuation cost of the evacuation scheme, and implement algorithm parameter updates based on the policy gradient algorithm; Step T2 includes the following sub-steps, T2-1: Map the disaster victim information to a specific dimension, discover the relational features in the disaster victim information, and combine the relational features into a global embedding of the disaster victim information; T2-2: Map the shelter information to a specific dimension, explore the characteristics of the shelter information, and obtain the global embedding of the shelter information; T2-3: Map the disaster point information to a specific dimension, explore the characteristics of the disaster point information, and obtain the global embedding of the disaster point information; T2-4: Integrate the global embedding of the disaster victims' information, the global embedding of the shelter information and the global embedding of the disaster point information to form the algorithm's overall observation of the evacuation problem, and decode the overall observation to obtain the probability of disaster victims choosing different shelters, that is, the evacuation probability; establish a parameterized distribution based on the evacuation probability, and use sampling means to realize the allocation of disaster victims and shelters; combine the allocations to obtain the evacuation plan generated by the algorithm. 4.Multi-disaster point emergency evacuation planning system, characterized by: Applied to the multi-disaster point emergency evacuation planning method as claimed in claim 3, include, The environment module is used to simulate multi-disaster emergency evacuation scenarios and generate a large number of evacuation problem instances for training algorithm parameters; A strategy module is used to process the information of each object in a single evacuation instance, extract features, obtain an overall observation of the problem, and generate an evacuation plan through the overall observation; The evaluation module is used to evaluate a single evacuation instance, generate an expected evacuation cost, compare it with the actual evacuation cost of the evacuation plan, and implement algorithm parameter update based on the policy gradient algorithm.

5. The multi-disaster point emergency evacuation planning system according to claim 4, characterized in that: The policy module includes: A disaster victim information encoder is used to map the disaster victim information to a specific dimension, discover relational features in the disaster victim information, and combine the relational features into a global embedding of the disaster victim information; A shelter information encoder is used to map the shelter information to a specific dimension, explore the characteristics of the shelter information, and obtain the global embedding of the shelter information; The disaster point information encoder is used to map the disaster point information to a specific dimension, explore the characteristics of the disaster point information, and obtain the global embedding of the disaster point information; A decision decoder is used to integrate the global embedding of the disaster victim information, the global embedding of the shelter information and the global embedding of the disaster site information into the algorithm's overall observation of the evacuation problem, and decode the overall observation to obtain the probability of disaster victims choosing different shelters, that is, the evacuation probability; establish a parameterized distribution based on the evacuation probability, and use sampling means to realize the allocation of disaster victims and shelters; combine the allocations to obtain the evacuation plan generated by the algorithm.

6. The multi-disaster point emergency evacuation planning system according to claim 5, characterized in that: The disaster victim information encoder is composed of a plurality of attention layers, and the attention layers are used to receive the disaster victim information after passing through the initial embedding layer and being mapped to the specified dimensional space; The attention layer contains two sub-layers, namely the multi-head self-attention layer and the feed-forward layer; The multi-head self-attention layer is used to receive the disaster victim information and calculate the multi-head self-attention value; wherein the multi-head self-attention value is used to perform batch normalization processing and jump connection operations to obtain the multi-head self-attention layer output; The feedforward layer is used to receive the output of the multi-head self-attention layer, and perform two linear mappings on the output of the multi-head self-attention layer, and activate it with a Relu function, wherein the linear mapping value of the feedforward layer is batch normalized and jump-connected to obtain the output of the feedforward layer, and the output of the feedforward layer is passed as input to the next attention layer, and the above process is repeated until all the attention layers are traversed, that is, the final embedding of the disaster victim information is obtained, and the final embedding of the disaster victim information is integrated, that is, the global embedding of the disaster victim information is obtained; The shelter information encoder and the disaster point information encoder have the same structure, and include a plurality of GRU units, and the number of the GRU units is consistent with the number of shelters and the number of disaster points respectively; The GRU unit is used to receive the initial embedding of the current shelter information or the initial embedding of the disaster point information, and the hidden state output by the previous GRU unit, calculate the hidden state of the current shelter information or the hidden state of the disaster point information, and send it to the next GRU unit; The above process is repeated until all the GRU units are traversed, and the embedding information output by the last GRU unit is the global embedding of the shelter information or the global embedding of the disaster point information. 7.Multi-disaster point emergency evacuation system, characterized by: The multi-disaster point emergency evacuation planning method as claimed in claim 3 is composed of a cloud-edge-end three-layer architecture, including a model training phase and an evacuation response phase. The three-layer architecture includes a mobile terminal equipped with a locator and having communication capabilities, an edge node with computing and communication capabilities, a data server for generating evacuation instances, a machine learning server for training emergency evacuation models, and an evacuation response server for evacuation decision-making and having communication capabilities; The model training phase is completed in the cloud. A large number of evacuation instances are generated by the data server and sent to the machine learning server. The machine learning server uses the evacuation instances to train and optimize the parameters in the evacuation model, and splits the evacuation model into an initial embedding model and a decision model. The initial embedding model will be sent to the edge node, and the decision model will be sent to the evacuation response server. The evacuation response stage is jointly executed by cloud-edge-end. The mobile terminal will collect disaster victim information and upload it to the edge node. The edge node will collect disaster point information, complete the initial embedding of disaster victim information, and upload the disaster point information and the initial embedding of disaster victim information to the evacuation response server. The evacuation response server will integrate the disaster point information, select a safe shelter, and use the decision model to complete the embedding of disaster point information and shelter information, and integrate the disaster point information embedding, shelter information embedding and received disaster victim information embedding to obtain an overall observation of the problem. The decision model will decode the overall observation and finally obtain an evacuation plan. The evacuation plan will be sent to the edge node, and the edge node will complete the evacuation guidance of the victims.

Images