Disaster emergency material distribution method based on multi-agent collaborative optimization

Abstract

The invention discloses a disaster emergency material distribution method based on multi-agent collaborative optimization, and the method comprises the following steps: S1, obtaining the real-time data of a disaster area, and generating a multi-dimensional data set; s2, constructing a collaborative optimization model based on the game theory, and generating a multi-level risk assessment model; s3, demand prediction is carried out by using the space-time convolutional neural network; s4, optimizing a material distribution path and resource scheduling; s5, automatically adjusting a scheduling strategy through a strategy gradient algorithm; s6, performing global optimization through particle swarm optimization, and adjusting a transportation path; and S7, dynamically adjusting a material distribution scheme and path planning. According to the invention, through the multi-level risk assessment model and the optimization algorithm, the material distribution path and the scheduling strategy are adjusted in real time, and the efficiency, the accuracy and the emergency response capability of disaster emergency material distribution are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency management and material dispatching, and in particular to a disaster emergency material distribution method based on multi-agent collaborative optimization. Background Art

[0002] With the frequent occurrence of natural disasters and the advancement of urbanization, the importance of emergency material dispatch and distribution has become increasingly prominent. After a disaster occurs, timely and accurate transportation of materials to the disaster area and reasonable distribution are the key to reducing the impact of the disaster, protecting the lives of disaster victims and restoring post-disaster order. Traditional material dispatch methods usually rely on experience and manual judgment, and often have problems such as uneven material distribution, transportation delays, and waste of resources. Especially in the face of large-scale, multi-type, and cross-regional disasters, the existing emergency material dispatch system cannot effectively respond to the changing disaster situation, resulting in inefficient emergency response and may even aggravate the negative impact of the disaster. Therefore, how to optimize the distribution and dispatch of emergency materials and improve dispatch efficiency has become an important issue that needs to be urgently solved in the current emergency management field.

[0003] At present, many traditional emergency material dispatch systems rely on static dispatch strategies and simple material allocation models, which cannot fully consider the dynamic changes and real-time feedback of the disaster area. For example, many systems only distribute materials based on the preset material demand, ignoring the changes in the actual demand in the disaster area, such as changes in traffic conditions, weather changes, personnel safety, material inventory status, etc. These factors will affect the timely supply and reasonable allocation of materials. In addition, many systems use simple optimization methods, such as linear programming or rule-based allocation strategies. Although they can provide a certain degree of dispatch support, it is difficult to achieve optimal allocation when dealing with complex multi-objective and multi-constrained emergency material dispatch problems.

[0004] In order to make up for the shortcomings of traditional methods, in recent years, more and more studies have begun to try to introduce intelligent technologies, especially material dispatch methods based on artificial intelligence (AI) and big data technology. These methods mainly analyze the real-time data of the disaster area, dynamically adjust the material distribution plan, and continuously optimize the dispatch strategy through intelligent algorithms, striving to achieve efficient distribution of emergency materials. For example, some studies use machine learning algorithms to predict the material demand in the disaster area and distribute materials based on the predicted results. However, although these methods can take into account the real-time data and demand changes in the disaster area, there are still some problems. For example, existing methods often rely on a single model, which makes it difficult to comprehensively consider the complex situation of multiple dimensions in the disaster area, resulting in low efficiency and lack of flexibility in practical applications.

[0005] Especially in the aspects of multi-level risk assessment and multi-objective optimization, the deficiencies of the existing technologies are particularly prominent. The risk factors in the disaster area are multi-level and multi-dimensional, including aspects such as traffic conditions, weather changes, resource inventory, personnel safety, and social stability. Existing technologies often ignore the interrelationships and dynamic changes among these factors and are unable to make efficient and flexible decisions in complex emergency environments. For example, some existing scheduling systems may rely on a single risk assessment model, ignoring the comprehensive consideration of multi-level risks in the disaster area, resulting in the inability of the material scheduling plan to fully address the actual situation.

[0006] In addition, the existing optimization methods also face certain challenges when dealing with large-scale and multi-objective scheduling problems. Traditional optimization methods usually have difficulty considering multiple optimization objectives simultaneously, such as transportation cost, delivery time, resource utilization rate, and risk control. Moreover, many existing systems lack adaptive scheduling capabilities and are unable to make dynamic adjustments and optimizations according to the real-time situation in the disaster area. For example, when sudden situations occur in the disaster area, such as traffic jams, weather changes, or insufficient supplies in a certain warehouse, existing systems often fail to make timely adjustments, resulting in a decrease in the efficiency of material distribution and even shortages of materials in some areas of the disaster area.

[0007] Therefore, how to provide a disaster emergency material distribution method based on multi-agent collaborative optimization is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a disaster emergency material distribution method based on multi-agent collaborative optimization. The present invention makes full use of intelligent optimization technologies such as multi-level risk assessment models, particle swarm optimization algorithms, genetic algorithms, and deep reinforcement learning, and details how to combine multi-dimensional risk factors such as real-time data, traffic, weather, safety, and inventory in the disaster area to dynamically optimize the material distribution plan. Through this method, it is possible to adjust the material distribution path in real time when a disaster occurs, thereby ensuring the efficient, timely, and accurate distribution of materials to the disaster area.

[0009] The present invention has many remarkable advantages. First, through the multi-level risk assessment model, it is possible to comprehensively consider multiple factors such as traffic, weather, safety, and inventory in the disaster area, avoiding the shortcoming of ignoring important risk factors in traditional methods. Second, by combining the particle swarm optimization algorithm and the genetic algorithm, the present invention can optimize the material distribution path in a complex scheduling environment, thereby reducing transportation costs and resource waste and improving the distribution efficiency. In addition, the application of the deep reinforcement learning algorithm enables the system to make dynamic adjustments according to the real-time changes in the disaster area, ensuring the flexibility of material distribution and the efficiency of emergency response. Through this comprehensive method, the present invention ensures that the material distribution plan not only has high accuracy but also can respond within the shortest time, improving the response speed and effect of material scheduling.

[0010] A disaster emergency material distribution method based on multi-agent collaborative optimization according to an embodiment of the present invention includes the following steps:

[0011] S1. Obtain real-time data of the disaster area, and generate a multi-dimensional data set through data cleaning and standardization processing;

[0012] S2. Construct a multi-agent collaborative optimization model based on game theory, define the multi-agent collaborative optimization goal through an information sharing and dynamic adjustment mechanism, and at the same time generate a multi-level risk assessment model;

[0013] S3. Based on the multi-dimensional data set and the multi-agent collaborative optimization goal, use a spatio-temporal convolutional neural network to perform multi-scale spatio-temporal feature analysis on the disaster area to generate a dynamic demand prediction model for the disaster area;

[0014] S4. Combine the dynamic demand prediction model of the disaster area and spatio-temporal features, optimize the material distribution path and resource scheduling, generate a material distribution plan, and adjust the resource configuration and distribution strategy in real time;

[0015] S5. Based on the preliminary material distribution plan, adopt a policy gradient algorithm to automatically adjust the scheduling strategy according to the changes in the disaster situation and resource requirements;

[0016] S6. Combine the multi-level risk assessment model and the scheduling strategy, and globally optimize the material distribution plan through a particle swarm optimization algorithm, predict potential risks and bottlenecks, and adjust the transportation path and resource allocation;

[0017] S7. Dynamically adjust the material distribution plan according to the real-time changes in the disaster situation and feedback information, and perform path planning and resource reallocation on unforeseen emergencies through a self-repair mechanism.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Define multiple agent roles, including logistics companies and rescue teams, and determine the resources, goals, and constraints of each agent;

[0020] S22. Establish a multi-agent collaborative optimization model and set the profit function of each agent:

[0021] R i =f i (x1,x2,...,x n ,t)+λ i ·g i (x1,x2,...,x n );

[0022] Wherein, R i represents the profit of the i-th agent, f i () is the objective function of the i-th agent, gi () is the constraint function of the i-th agent, where x1, x2,..., x n are resource allocation and scheduling policy variables, t represents the time factor, and λ i is the Lagrange multiplier;

[0023] S23. An adaptive game algorithm based on multi-agent reinforcement learning is adopted. The strategies of each agent in the game model are dynamically adjusted through a deep Q-network to optimize the cooperation and competition relationships among multiple agents, and the Q-value function of each agent is obtained:

[0024] Q i (x1, x2,..., x n , t) = r i + γ·max a' Q i (x1', x2',..., x n ');

[0025] where Q i () is the Q-value function of the i-th agent, r i is the immediate reward of the current strategy, γ is the discount factor, x1', x2',..., x n ' represents the future state, t' is the future time, a' is the future action selection, and max is the maximum operation;

[0026] S24. Through the learning mechanism of the deep Q-network, the Q-value function is updated according to real-time data feedback and environmental changes, and the ε-greedy strategy is used for action selection:

[0027] Q i (x t , a t ) ← Q i (x t , a t ) + α(r t + γ·max a' Q i (x t+1 , a′) - Q i (x t , a t ));

[0028] where Q i (x t , a t ) is the Q-value of the i-th agent when choosing action a t in state x t , r t is the immediate reward, α is the learning rate, and max a' Q i (xt+1 , a') is the maximum Q value in the future state, x t+1 is the state;

[0029] S25. According to the dynamic adjustment mechanism, generate a multi-level risk assessment model. Through comprehensive evaluation of the material needs, transportation capacity, and time constraints in the disaster area, predict potential resource bottlenecks and risk points, and combine the feedback adjustment strategy of the game model to form an optimization goal.

[0030] Optionally, the S3 specifically includes:

[0031] S31. Obtain real-time data of the disaster area, including historical material needs, traffic conditions, weather information, and geographical information in the disaster area;

[0032] S32. Use a spatio-temporal convolutional neural network to perform spatio-temporal feature analysis on the disaster area. Extract spatial features through the spatial convolutional layer, capture temporal features through the temporal convolutional layer, and generate a spatio-temporal feature expression of the disaster area demand:

[0033] X spatial = Conv(X input , W1);

[0034] where X input is the input data, W1 is the spatial convolution kernel, X spatial is the spatial feature extracted through the convolution operation, and Conv is the convolution operation;

[0035] S33. Combine the graph attention network to model the geographical information and transportation network in the disaster area, extract the heterogeneous features of the graph data, and use the self-attention mechanism to perform weighted learning on different adjacent nodes of the graph data to generate a graph representation of the disaster area traffic and geographical features:

[0036] X geo = GAT(X spatial , A, W2);

[0037] where A is the adjacency matrix of the graph, X geo represents the geographical and transportation features extracted by graph convolution, W2 is the weight matrix in the graph attention network, and GAT is the graph attention network;

[0038] S34. Through the combination of the spatio-temporal convolutional neural network and the graph attention network, fuse the spatio-temporal and graph features to generate a predicted value of the material demand in the disaster area;

[0039] S35. Combine the adaptive weighting mechanism to perform weighted processing on the spatio-temporal features and geographical traffic features of the material demand in the disaster area, adjust the influence of different features on the prediction result, and generate a dynamic demand prediction model for the disaster area:

[0040]

[0041] Among them, ω1 and ω2 are adaptive weighting factors, W3 and W4 are weight matrices for spatio-temporal and geographical feature weighting, and X temporal is the temporal feature extracted by spatio-temporal convolution, and is the predicted value of material demand.

[0042] Optionally, the S4 specifically includes:

[0043] S41. Based on the generated dynamic demand prediction model for the disaster area, obtain the material demand of each area in the disaster area, and construct a preliminary material distribution objective function F:

[0044]

[0045] Among them, c ij represents the transportation cost from warehouse i to destination j, x ij is the allocated quantity of materials, m is the number of warehouses, n is the number of destinations, λ i is the Lagrange multiplier, s i is the storage capacity of warehouse i, and min is the minimum value operation;

[0046] S42. According to the traffic conditions, infrastructure damage degree, and weather in the disaster area, combined with spatio-temporal feature data, construct material transportation path constraint conditions;

[0047] S43. Adopt a multi-objective reinforcement learning algorithm to globally optimize the material distribution path and resource scheduling:

[0048] Q i (x t ,a t )=r t +γ·max a' Q i (x t+1 ,a′);

[0049] Among them, Q i (x t ,a t ) is the Q value of the i-th agent choosing action a t in state x t , γ is the discount factor, a' is the future action selection, r t is the immediate reward, max is the maximum value operation, and x t+1 is the state;

[0050] S44. Add Kalman filtering to the particle swarm optimization algorithm for dynamic path adjustment, reduce the prediction error through state estimation, and update the material distribution path in real time:

[0051]

[0052] P t|t = (I - K t H t )P t|t-1 ;

[0053] Wherein, is the estimated state, K t is the Kalman gain, z t is the measured value, P t|t-1 is the previous estimated covariance, H t is the observation matrix, I is the identity matrix, is the state estimate before time t, P t|t is the updated covariance matrix;

[0054] S45. Based on the material distribution plan, adjust the resource allocation and distribution strategy in real time, and re-plan the path through the scheduling optimization algorithm.

[0055] Optionally, the S5 specifically includes:

[0056] S51. According to the generated preliminary material distribution plan, obtain the real-time data feedback of the disaster area, including traffic information, weather changes, and resource inventory, establish a real-time adjustment model, and generate the adjustment factors of the resource demand and material scheduling in the disaster area by combining different data sources through the weighted summation method:

[0057] D adjusted = τ1·D traffic + τ2·D weather + τ3·D inventory ;

[0058] Wherein, D adjusted is the adjusted resource demand vector in the disaster area, D traffic , D weather and D inventory respectively represent the traffic information, weather changes, and resource inventory data in the disaster area, and τ1, τ2, and τ3 are the weighting coefficients;

[0059] S52. Based on the real-time adjustment model and the disaster area feedback data, establish the dynamic scheduling objective function D f :

[0060]

[0061] Wherein, c ij is the transportation cost from warehouse i to destination j, t ij is the delivery delay, u ij is the resource utilization rate, α1, α2, and α3 are the weighting coefficients, x ijis the quantity of allocated materials, m is the number of warehouses, and n is the number of destinations;

[0062] S53. Adopt the policy gradient algorithm in deep reinforcement learning, optimize the material scheduling plan by updating the policy function, learn the optimal scheduling policy, and dynamically adjust the scheduling policy according to the real-time feedback of the disaster area to optimize the material distribution efficiency and resource utilization:

[0063]

[0064] Among them, is the policy gradient, E is the expected operation, π θ (a t |x t ) is the current policy, a t is the action selected at time t, x t is the state of the system at time t, V(x t+1 ) is the value of the next state, V(x t is the value of the current state, r t is the immediate reward, γ is the discount factor, ln is the logarithmic function, θ is the parameter of the policy, x t+1 is the next state;

[0065] S54. Based on the optimization result of the policy gradient algorithm, adjust the material scheduling plan, and the generated scheduling policy automatically adjusts the distribution path in real-time feedback;

[0066] S55. Re-plan the material distribution and scheduling plan for the disaster area according to the material scheduling policy.

[0067] Optionally, the specific content of S6 includes:

[0068] S61. Based on the generated material scheduling policy and real-time disaster area demand prediction data, obtain the risk assessment data of each area in the disaster area, fuse risk factors at different levels through a multi-level risk assessment model, and obtain the fuzzy degree of each risk factor through the fuzzy logic reasoning method, and finally obtain the comprehensive risk assessment value R total as the input of material scheduling:

[0069]

[0070] Among them, μ1, μ2, μ3, and μ4 are weighting coefficients, and are the traffic, weather, safety, and inventory risks of the i-th area in the disaster area respectively, and n is the number of areas in the disaster area;

[0071] S62. Based on the comprehensive risk assessment, use the multi-objective genetic algorithm to optimize the material distribution path, construct a fitness function to evaluate the advantages and disadvantages of each material distribution plan, and the fitness function comprehensively considers the transportation cost, risk assessment, and resource utilization efficiency:

[0072]

[0073] Among them, F i represents the fitness of the i-th material distribution plan, c ik is the transportation cost from warehouse i to destination k, x ik is the quantity of materials distributed from warehouse i to destination k, u ik is the resource utilization rate, and ρ1, ρ2, and ρ3 are weighting coefficients;

[0074] S63. Adopt the particle swarm optimization algorithm for global optimization and adjust the distribution path:

[0075]

[0076] Among them, represents the velocity of particle i in the k-th generation, represents the position of particle i in the k-th generation, is the individual historical optimal position of particle i, gbest k is the historical optimal position of all particles, ω is the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random factors, is the updated particle velocity, is the updated particle position;

[0077] S64. By optimizing the material distribution path, adjust the scheduling strategy in real time, and re-plan the distribution path, comprehensively considering the impacts of traffic and weather changes in the disaster area, and maximize the resource utilization rate and efficiency;

[0078] S65. According to the optimization results, adjust the material distribution plan through the adaptive path correction mechanism, and re-plan the material distribution path in combination with the real-time disaster situation changes.

[0079] The beneficial effects of the present invention are:

[0080] The global optimization method for the material allocation plan based on the multi-level risk assessment model and scheduling strategy optimization proposed by the present invention solves the problems of low efficiency, slow response, resource waste, etc. existing in traditional emergency material scheduling. The present invention utilizes the multi-level risk assessment model, comprehensively considering multi-dimensional risk factors such as transportation, weather, safety, and inventory in the disaster area, enabling the material allocation plan to more accurately reflect the actual needs and risk status of the disaster area, thereby optimizing the scheduling and allocation path of materials. This innovation effectively makes up for the deficiencies of the prior art in being unable to fully consider dynamic risk changes and multi-objective optimization in emergency material scheduling.

[0081] In addition, the present invention globally optimizes the material allocation path through the particle swarm optimization algorithm and genetic algorithm, ensuring that the optimal solution can be found in large-scale, multi-objective scheduling problems. This optimization not only reduces transportation costs and resource waste, but also improves the efficiency of material allocation, ensuring that materials can be delivered to the places in need in the disaster area more timely. By introducing the deep reinforcement learning algorithm, the present invention further improves the adaptive ability of the scheduling strategy, enabling the system to adjust the material allocation plan in real time according to the changes in the disaster situation and emergencies. This dynamic adjustment ability can maximize the guarantee of the timely supply and effective allocation of materials in the event of emergencies or environmental changes in the disaster area, avoiding material shortages or uneven distribution that may be caused by scheduling lags.

[0082] Generally speaking, by comprehensively utilizing the multi-level risk assessment model and advanced optimization algorithms, the present invention significantly improves the flexibility, accuracy, and efficiency of the material scheduling system, and can optimize the material allocation path and resource scheduling strategy in real time in a complex and dynamic disaster situation environment, thereby effectively enhancing the response speed and coping ability of the emergency material scheduling in the disaster area, and ensuring the safety of the lives of the victims and the restoration of social order. Brief Description of the Drawings

[0083] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0084] Figure 1 is a flowchart of the disaster emergency material allocation method based on multi-agent collaborative optimization proposed by the present invention;

[0085] Figure 2 is a schematic diagram of the material scheduling path of the disaster emergency material allocation method based on multi-agent collaborative optimization proposed by the present invention. Detailed Embodiment

[0086] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0087] Reference Figure 1 and Figure 2 , a disaster emergency material distribution method based on multi-agent collaborative optimization, includes the following steps:

[0088] S1. Obtain the real-time data of the disaster area, and generate a multi-dimensional data set through data cleaning and standardization processing;

[0089] S2. Build a multi-agent collaborative optimization model based on game theory, define the multi-agent collaborative optimization goal through information sharing and dynamic adjustment mechanism, and generate a multi-level risk assessment model at the same time;

[0090] S3. Based on the multi-dimensional data set and the multi-agent collaborative optimization goal, use a spatio-temporal convolutional neural network to perform multi-scale spatio-temporal feature analysis on the disaster area, and generate a dynamic demand prediction model for the disaster area;

[0091] S4. Combine the dynamic demand prediction model of the disaster area and spatio-temporal features, optimize the material distribution path and resource scheduling, generate a material distribution plan, and adjust the resource configuration and distribution strategy in real time;

[0092] S5. Based on the preliminary material distribution plan, adopt the policy gradient algorithm to automatically adjust the scheduling strategy according to the disaster situation changes and resource requirements;

[0093] S6. Combine the multi-level risk assessment model and the scheduling strategy, and globally optimize the material distribution plan through the particle swarm optimization algorithm, predict potential risks and bottlenecks, and adjust the transportation path and resource allocation;

[0094] S7. According to the real-time disaster situation changes and feedback information, dynamically adjust the material distribution plan, and perform path planning and resource reallocation on unforeseen emergencies through the self-repair mechanism.

[0095] In this embodiment, the S2 specifically includes:

[0096] S21. Define multiple agent roles, including logistics companies and rescue teams, and determine the resources, goals and constraints of each agent;

[0097] S22. Establish a multi-agent collaborative optimization model and set the revenue function of each agent:

[0098] R i = f i (x1,x2,...,x n ,t)+λ i ·g i (x1,x2,...,x n );

[0099] Among them, R idenotes the revenue of the $i$-th entity, $f$ i () is the objective function of the $i$-th entity, $g$ i () is the constraint function of the $i$-th entity, $x_1, x_2, \ldots, x$ n are resource allocation and scheduling policy variables, $t$ represents the time factor, $\lambda$ i is the Lagrange multiplier;

[0100] S23. An adaptive game algorithm based on multi-agent reinforcement learning is adopted to dynamically adjust the strategies of each entity in the game model through a deep Q-network, optimize the cooperation and competition relationships among multiple entities, and obtain the Q-value function of each entity:

[0101] Q i (x_1, x_2, \ldots, x n , t) = r i + \gamma \cdot \max a' Q i (x_1', x_2', \ldots, x n ';

[0102] where $Q$ i () is the Q-value function of the $i$-th entity, $r$ i is the immediate reward of the current strategy, $\gamma$ is the discount factor, $x_1', x_2', \ldots, x$ n ' represents the future state, $t'$ is the time of the future moment, $a'$ is the future action selection, and $\max$ is the maximum operation;

[0103] S24. Through the learning mechanism of the deep Q-network, the Q-value function is updated according to real-time data feedback and environmental changes, and the $\epsilon$-greedy strategy is adopted for action selection:

[0104] Q i (x t , a t ) \leftarrow Q i (x t , a t ) + \alpha (r t + \gamma \cdot \max a' Q i (x t+1 , a') - Q i (x t , a t ));

[0105] where $Q$ i (x t , a t ) is the Q-value of the $i$-th entity when choosing action $a$ t in state $x$ t , $r$ tFor immediate reward, α is the learning rate, max a' Q i (x t+1 , a') is the maximum Q-value in the future state, and x t+1 is the state;

[0106] S25. According to the dynamic adjustment mechanism, a multi-level risk assessment model is generated. By comprehensively evaluating the material needs, transportation capacity, and time constraints in the disaster area, potential resource bottlenecks and risk points are predicted, and combined with the feedback adjustment strategy of the game model, an optimization goal is formed.

[0107] In this embodiment, the specific steps of S3 include:

[0108] S31. Obtain the real-time data of the disaster area, including the historical material needs, traffic conditions, weather information, and geographical information of the disaster area;

[0109] S32. Use the spatio-temporal convolutional neural network to analyze the spatio-temporal characteristics of the disaster area. Extract spatial features through the spatial convolutional layer and capture temporal features through the temporal convolutional layer to generate the spatio-temporal feature expression of the disaster area demand:

[0110] X spatial = Conv(X input , W1);

[0111] Among them, X input is the input data, W1 is the spatial convolution kernel, and X spatial is the spatial feature extracted through the convolution operation, and Conv is the convolution operation;

[0112] S33. Combine the graph attention network to model the geographical information and traffic network of the disaster area, extract the heterogeneous features of the graph data, and use the self-attention mechanism to perform weighted learning on different adjacent nodes of the graph data to generate the graph representation of the traffic and geographical features of the disaster area:

[0113] X geo = GAT(X spatial , A, W2);

[0114] Among them, A is the adjacency matrix of the graph, X geo represents the geographical and traffic features extracted by the graph convolution, W2 is the weight matrix in the graph attention network, and GAT is the graph attention network;

[0115] S34. Through the combination of the spatio-temporal convolutional neural network and the graph attention network, fuse the spatio-temporal and graph features to generate the predicted value of the material demand in the disaster area;

[0116] S35. Combine with the adaptive weighting mechanism, weight the spatio-temporal characteristics and geographical traffic characteristics of the material requirements in the disaster area, adjust the influence of different characteristics on the prediction results, and generate a dynamic demand prediction model for the disaster area:

[0117]

[0118] Among them, ω1 and ω2 are adaptive weighting factors, W3 and W4 are weight matrices for spatio-temporal and geographical feature weighting, X temporal is the time series feature extracted by spatio-temporal convolution, is the predicted value of material demand.

[0119] In this embodiment, the S4 specifically includes:

[0120] S41. Based on the generated dynamic demand prediction model for the disaster area, obtain the material demand of each region in the disaster area, and construct a preliminary material distribution objective function F:

[0121]

[0122] Among them, c ij represents the transportation cost from warehouse i to destination j, x ij is the allocated quantity of materials, m is the number of warehouses, n is the number of destinations, λ i is the Lagrange multiplier, s i is the storage capacity of warehouse i, and min is the minimum value operation;

[0123] S42. According to the traffic conditions, infrastructure damage degree, and weather in the disaster area, combined with spatio-temporal feature data, construct material transportation path constraint conditions;

[0124] S43. Adopt a multi-objective reinforcement learning algorithm to globally optimize the material distribution path and resource scheduling:

[0125] Q i (x t ,a t )=r t +γ·max a' Q i (x t+1 ,a′);

[0126] Among them, Q i (x t ,a t ) is the Q value of the i-th agent choosing action a t in state x t , γ is the discount factor, a' is the future action selection, r t is the immediate reward, max is the maximum value operation, and x t+1 is the state;

[0127] S44. Incorporate Kalman filtering into the particle swarm optimization algorithm for dynamic path adjustment, reduce prediction errors through state estimation, and update the material distribution path in real time:

[0128]

[0129] P t|t =(I - K t H t )P t|t-1 ;

[0130] Wherein, is the estimated state, K t is the Kalman gain, z t is the measurement value, P t|t-1 is the previous estimated covariance, H t is the observation matrix, I is the identity matrix, is the state estimate before time t, P t|t is the updated covariance matrix;

[0131] S45. Based on the material distribution plan, adjust the resource allocation and distribution strategy in real time, and re-plan the path through the scheduling optimization algorithm.

[0132] In this embodiment, the S5 specifically includes:

[0133] S51. According to the generated preliminary material distribution plan, obtain real-time data feedback of the disaster area, including traffic information, weather changes, and resource inventory, establish a real-time adjustment model, and generate adjustment factors for resource demand and material scheduling in the disaster area by combining different data sources through the weighted summation method:

[0134] D adjusted =τ1·D traffic +τ2·D weather +τ3·D inventory ;

[0135] Wherein, D adjusted is the adjusted resource demand vector in the disaster area, D traffic , D weather and D inventory respectively represent traffic information, weather changes, and resource inventory data in the disaster area, and τ1, τ2, and τ3 are weighting coefficients;

[0136] S52. Based on the real-time adjustment model and the feedback data of the disaster area, establish a dynamic scheduling objective function D f :

[0137]

[0138] Among them, c ij is the transportation cost from warehouse i to destination j, t ij is the delivery delay, u ij is the resource utilization rate, α1, α2, and α3 are weighting coefficients, x ij is the quantity of allocated materials, m is the number of warehouses, and n is the number of destinations;

[0139] S53. Adopt the policy gradient algorithm in deep reinforcement learning to optimize the material scheduling plan by updating the policy function, learn the optimal scheduling policy, and dynamically adjust the scheduling policy according to the real-time feedback of the disaster area to optimize the material distribution efficiency and resource utilization:

[0140]

[0141] Among them, is the policy gradient, E is the expectation operation, π θ (a t |x t ) is the current policy, a t is the action selected at time t, x t is the state of the system at time t, V(x t+1 ) is the value of the next state, V(x t is the value of the current state, r t is the immediate reward, γ is the discount factor, ln is the logarithmic function, θ is the parameter of the policy, x t+1 is the next state;

[0142] S54. Based on the optimization result of the policy gradient algorithm, adjust the material scheduling plan, and automatically adjust the distribution path according to the generated scheduling policy in real-time feedback;

[0143] S55. Re-plan the material distribution and scheduling plan for the disaster area according to the material scheduling policy.

[0144] In this embodiment, the S6 specifically includes:

[0145] S61. Based on the generated material scheduling policy and real-time disaster area demand prediction data, obtain the risk assessment data of each area in the disaster area, fuse risk factors at different levels through a multi-level risk assessment model, and obtain the fuzzy degree of each risk factor through a fuzzy logic reasoning method, and finally obtain the comprehensive risk assessment value R total as the input of material scheduling:

[0146]

[0147] Among them, μ1, μ2, μ3, and μ4 are weighting coefficients, and They are respectively the traffic, weather, safety and inventory risks of the i-th area in the disaster area, and n is the number of areas in the disaster area;

[0148] S62. On the basis of comprehensive risk assessment, use the multi-objective genetic algorithm to optimize the material distribution path, and construct a fitness function to evaluate the advantages and disadvantages of each material distribution plan. The fitness function comprehensively considers the transportation cost, risk assessment and resource utilization efficiency:

[0149]

[0150] Among them, F i represents the fitness of the i-th material distribution plan, c ik is the transportation cost from warehouse i to destination k, x ik is the quantity of materials allocated from warehouse i to destination k, u ik is the resource utilization rate, and ρ1, ρ2 and ρ3 are weighting coefficients;

[0151] S63. Adopt the particle swarm optimization algorithm for global optimization and adjust the distribution path:

[0152]

[0153] Among them, represents the velocity of particle i in the k-th generation, represents the position of particle i in the k-th generation, pbest i k is the individual historical optimal position of particle i, gbest k is the historical optimal position of all particles, ω is the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random factors, is the updated particle velocity, is the updated particle position;

[0154] S64. By optimizing the material distribution path, adjust the scheduling strategy in real time, and re-plan the distribution path, comprehensively considering the impacts of traffic and weather changes in the disaster area, and maximize the resource utilization rate and efficiency;

[0155] S65. According to the optimization results, adjust the material distribution plan through the adaptive path correction mechanism, and re-plan the material distribution path in combination with the real-time disaster situation changes.

[0156] Example 1:

[0157] To verify the feasibility of the present invention in implementation, the present invention is applied to the post-disaster emergency material scheduling in a certain area. A strong typhoon occurred in a certain area, resulting in traffic interruption, house collapse, power outage and communication paralysis in the area, and the material demands in multiple cities and towns increased sharply. After the typhoon, the material demands in the affected areas included water, food, tents and medical equipment. Due to the damage of post-disaster transportation and the logistics bottleneck between regions, the material supply in some affected areas was delayed. How to optimize the material distribution plan and complete the scheduling efficiently in a short time became the key issue for emergency management in this area.

[0158] To solve this problem, the material scheduling optimization method proposed by the present invention is applied to the actual scenario. First, the post-disaster emergency material scheduling center collected real-time data of the disaster area, including road traffic conditions, weather forecasts, material inventory situations and the specific demands of the disaster area. There are differences in demands among multiple regions in the disaster area. The traffic in some regions is severely blocked, while other regions are in urgent need of relief materials such as water and food due to insufficient inventory. Through real-time data collection and a risk assessment model, the system comprehensively analyzes the material demands of each region.

[0159] In the specific application, the system first evaluates each affected area through a multi-level risk assessment model and calculates the risk value of each area. The risk assessment model takes into account multiple aspects such as the traffic conditions, weather changes, resource inventory and material demands in the disaster area. Through this model, the risk value of the disaster area reflects the severity of the disaster situation and the urgency of material scheduling. For example, due to the influence of heavy rain and storms, the roads in some areas are impassable, becoming high-risk areas; while in other areas, due to sufficient material inventory, the risk is relatively low.

[0160] Subsequently, the system applies the particle swarm optimization algorithm and the genetic algorithm to globally optimize the material distribution path. According to the risk assessment results, the material distribution path and the resource scheduling plan are re-planned to reduce the transportation cost, optimize the transportation route and avoid blocked areas. Through the action of the optimization algorithm, the timeliness of material transportation is significantly improved, ensuring that materials can reach the places in greatest need on time and accurately.

[0161] For example, the initial transportation path planning was expected to take 48 hours, while after optimization, the material delivery time was shortened to 36 hours. The optimized plan avoids traffic bottlenecks and takes into account the influence of weather changes in the disaster area, and can deliver essential materials to high-risk areas before heavy rain arrives. Through the deep reinforcement learning algorithm, the system can dynamically adjust the material distribution plan according to real-time feedback. When an emergency occurs in a certain area (such as the road being closed again or the weather changing), the system will immediately adjust the path and re-plan the resource scheduling strategy.

[0162] In this embodiment, the optimized material scheduling system has greatly improved the distribution efficiency and accuracy of emergency supplies. For example, within the first day after the disaster, the distribution route of the supplies was optimized by the system, and the actual transportation time was shortened from the initially planned 72 hours to 48 hours, saving 33% of the transportation time. For some areas greatly affected by traffic and weather, the material distribution time was shortened from the expected 96 hours to 72 hours, saving 24 hours of transportation time. By this method, supplies can reach the disaster area in the shortest time, greatly improving the efficiency of post-disaster emergency response.

[0163] Table 1 Material Requirements and Allocation Plans in the Disaster Area

[0164]

[0165] Based on the data in the above table, the material requirements and the improvement of scheduling efficiency can be analyzed. The table lists the material requirements, initial scheduling time, optimized scheduling time, and the improvement of scheduling efficiency in five disaster-stricken areas.

[0166] First of all, from the perspective of material requirements, the demand in Area 5 is the largest, including 300 tons of water, 200 tons of food, 800 tents, and 500 pieces of medical equipment. While the demand in Area 1 is less, with 150 tons of water demand, 50 tons of food, 300 tents, and 100 pieces of medical equipment. This shows that there are significant differences in material requirements among different regions, and it is particularly important to optimize the scheduling route.

[0167] In terms of the initial scheduling time, the material distribution time in Area 4 is the longest, at 120 hours, indicating that factors such as traffic and weather in this area have a greater impact on scheduling. The initial scheduling time in Area 2 is 72 hours, which is relatively short, indicating that the material scheduling situation in this area is better, but there is still room for optimization.

[0168] After the optimized scheduling time, the distribution time of all regions has been significantly shortened. For example, the optimized scheduling time in Area 1 is 36 hours, which is 12 hours less than the initial scheduling time, and the scheduling efficiency has increased by 25%. The optimized scheduling time in Area 2 is 48 hours, a reduction of 24 hours, and the scheduling efficiency has increased by 33.33%. These data show the remarkable effect of the present invention in optimizing the scheduling route and improving the material distribution efficiency.

[0169] For Area 3, Area 4, and Area 5, the optimized scheduling times are 72 hours, 90 hours, and 90 hours respectively. Although the optimization effects in these areas are relatively small, there is still a 25% increase in efficiency. Especially in areas with a relatively long initial scheduling time, the optimized scheduling has greatly improved the timeliness of material distribution.

[0170] In summary, the tabular data show that the optimization method of the present invention significantly improves the material scheduling efficiency, ensures that materials can be timely distributed to the disaster area, thereby enhancing the response speed and effect of post-disaster rescue.

[0171] As described above, only the preferred specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A disaster emergency material distribution method based on multi-agent collaborative optimization, characterized in that: The steps include: S1. Obtain real-time data from the disaster area and generate a multi-dimensional data set through data cleaning and standardization; S2. Construct a multi-agent collaborative optimization model based on game theory, define the multi-agent collaborative optimization objectives through information sharing and dynamic adjustment mechanisms, and generate a multi-level risk assessment model; S3. Based on multi-dimensional data sets and multi-agent collaborative optimization objectives, a spatiotemporal convolutional neural network is used to analyze the multi-scale spatiotemporal characteristics of the disaster area and generate a dynamic demand prediction model for the disaster area; S4. Combine the dynamic demand forecasting model and spatiotemporal characteristics of the disaster area to optimize the material distribution path and resource scheduling, generate material distribution plans, and adjust resource configuration and allocation strategies in real time; S5. Based on the preliminary material allocation plan, the policy gradient algorithm is used to automatically adjust the scheduling strategy according to the changes in the disaster situation and resource demand; S6. Combine the multi-level risk assessment model and scheduling strategy, use the particle swarm optimization algorithm to globally optimize the material distribution plan, predict potential risks and bottlenecks, and adjust the transportation route and resource allocation; S7. Dynamically adjust the material distribution plan based on real-time disaster changes and feedback information, and use the self-repair mechanism to perform path planning and resource reallocation for unforeseen emergencies.

2. The disaster emergency material distribution method based on multi-agent collaborative optimization according to claim 1 is characterized in that: The S2 specifically includes: S21. Define multiple subject roles, including logistics companies and rescue teams, and determine the resources, goals, and constraints of each subject; S22. Establish a multi-agent collaborative optimization model and set the profit function of each agent: R i =f i (x1,x2,…,x n ,t)+λ i ·g i (x1,x2,…,x n ): Among them, R i represents the income of the ith entity, f i () is the objective function of the ith subject, g i () is the constraint function of the ith subject, x1,x2,...,x n is the resource configuration and scheduling strategy variable, t represents the time factor, λ i is the Lagrange multiplier; S23, using an adaptive game algorithm based on multi-agent reinforcement learning, dynamically adjust the strategies of each subject in the game model through a deep Q network, optimize the cooperation and competition relationship between multiple subjects, and obtain the Q value function of each subject: Q i (x1,x2,...,x n ,t)=r i +γ·max a 'Q i (x1',x2',...,x n ',t'); Among them, Q i () is the Q value function of the i-th subject, r i is the immediate reward of the current strategy, γ is the discount factor, x1',x2',...,x n ' represents the future state, t' is the time of the future moment, a' is the future action selection, and max is the maximum value operation; S24, through the learning mechanism of the deep Q network, the Q value function is updated according to real-time data feedback and environmental changes, and the ε-greedy strategy is used for action selection: Q i (x t ,a t )←Q i (x t ,a t )+α(r t +γ·max a' Q i (x t+1 ,a)-Q i (x t ,a t )); Among them, Q i (x t ,a t ) is the state of the i-th subject in x t Next select action a t Q value, r t is the immediate reward, α is the learning rate, max a' Q i (x t+1 ,a') is the maximum Q value in the future state, x t+1 for status; S25. Generate a multi-level risk assessment model based on the dynamic adjustment mechanism. Through a comprehensive assessment of the disaster area's material demand, transportation capacity, and time constraints, predict potential resource bottlenecks and risk points, and combine the feedback adjustment strategy of the game model to form an optimization goal.

3. The disaster emergency material distribution method based on multi-agent collaborative optimization according to claim 1 is characterized in that: The S3 specifically includes: S31. Obtain real-time data of the disaster area, including historical material demand, traffic conditions, weather information, and geographic information of the disaster area; S32. Use the spatiotemporal convolutional neural network to analyze the spatiotemporal characteristics of the disaster area, extract the spatial features through the spatial convolution layer, capture the temporal features through the temporal convolution layer, and generate the spatiotemporal feature expression of the disaster area needs: X spatial =Conv(X input ,W1); Among them, X input is the input data, W1 is the spatial convolution kernel, X spatial is the spatial feature extracted by convolution operation, Conv is the convolution operation; S33. Combine the graph attention network to model the geographical information and traffic network of the disaster area, extract the heterogeneous features of the graph data, and use the self-attention mechanism to perform weighted learning on different adjacent nodes of the graph data to generate a graph representation of the traffic and geographical features of the disaster area: X geo =GAT(X spatial ,A,W2); Among them, A is the adjacency matrix of the graph, X geo represents the geographical and traffic features extracted by graph convolution, W2 is the weight matrix in the graph attention network, and GAT is the graph attention network; S34, by combining the spatiotemporal convolutional neural network with the graph attention network, the spatiotemporal and graph features are integrated to generate the predicted value of material demand in the disaster area; S35. Combined with the adaptive weighting mechanism, the temporal and spatial characteristics and geographical traffic characteristics of the disaster area material demand are weighted, the impact of different characteristics on the prediction results is adjusted, and a dynamic demand prediction model for the disaster area is generated: Among them, ω1 and ω2 are adaptive weighting factors, W3 and W4 are weight matrices for weighting spatiotemporal and geographic features, and X temporal is the temporal feature extracted by spatiotemporal convolution. The material demand forecast value.

4. The method for allocating disaster emergency materials based on multi-agent collaborative optimization according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on the generated dynamic demand forecasting model for the disaster area, obtain the material demand of each area in the disaster area, and construct a preliminary material allocation objective function F: Among them, c ij represents the transportation cost from warehouse i to destination j, x ij is the number of materials allocated, m is the number of warehouses, n is the number of destinations, λ i is the Lagrange multiplier, s i is the storage capacity of warehouse i, and min is the minimum value operation; S42. According to the traffic conditions, infrastructure damage, weather and time-space characteristic data in the disaster area, the material transportation path constraint conditions are established; S43, use multi-objective reinforcement learning algorithm to globally optimize material distribution path and resource scheduling: Q i (x t ,a t )=r t +γ·max a' Q i (x t+1 ,a); Among them, Q i (x t ,a t ) is the state of the i-th subject in x t Next select action a t Q value, γ is the discount factor, a' is the future action selection, r t is the immediate reward, max is the maximum value operation, x t+1 for status; S44. Add Kalman filtering to the particle swarm optimization algorithm for dynamic path adjustment, reduce prediction errors through state estimation, and update the material distribution path in real time: P t|t =(I-K t H t )P t|t-1 ; in, is the estimated state, K t is the Kalman gain, z t is the measured value, P t|t-1 is the previous estimated covariance, H t is the observation matrix, I is the identity matrix, is the state estimate before time t, P t|t is the updated covariance matrix; S45. Based on the material allocation plan, adjust the resource configuration and allocation strategy in real time, and re-plan the path through the scheduling optimization algorithm.

5. The method for allocating disaster emergency materials based on multi-agent collaborative optimization according to claim 1 is characterized in that: The S5 specifically includes: S51. Based on the generated preliminary material allocation plan, obtain real-time data feedback from the disaster area, including traffic information, weather changes, and resource inventory, establish a real-time adjustment model, combine different data sources through the weighted sum method, and generate adjustment factors for disaster area resource demand and material dispatch: D adjusted =τ1·D traffic +τ2·D weather +τ3·D inventory ; Among them, D adjusted is the adjusted resource demand vector of the disaster area, D traffic , D weather and D inventory They represent the traffic information, weather changes and resource inventory data in the disaster area, respectively, and τ1, τ2 and τ3 are weighting coefficients; S52. Based on the real-time adjustment model and the feedback data from the disaster area, establish a dynamic scheduling objective function D f : Among them, c ij is the transportation cost from warehouse i to destination j, t ij Due to delivery delay, u ij is the resource utilization rate, α1, α2 and α3 are weighted coefficients, x ij is the number of materials allocated, m is the number of warehouses, and n is the number of destinations; S53, using the policy gradient algorithm in deep reinforcement learning, optimizing the material dispatching plan by updating the policy function, learning the best dispatching strategy, and dynamically adjusting the dispatching strategy based on real-time feedback from the disaster area to optimize material distribution efficiency and resource utilization: in, is the policy gradient, E is the expected operation, π θ (a t |x t ) is the current strategy, a t is the action chosen at time t, x t is the state of the system at time t, V(x t+1 ) is the value of the next state, V(x t is the value of the current state, r t is the immediate reward, γ is the discount factor, ln is the logarithmic function, θ is the parameter of the strategy, x t+1 For the next state; S54, based on the optimization results of the policy gradient algorithm, adjust the material dispatching plan, and automatically adjust the distribution path based on the dispatching strategy generated by real-time feedback; S55. Re-plan the material distribution and dispatch plan in the disaster area based on the material dispatch strategy.

6. The method for allocating disaster emergency materials based on multi-agent collaborative optimization according to claim 1 is characterized in that: The S6 specifically includes: S61. Based on the generated material dispatch strategy and real-time disaster area demand forecast data, obtain the risk assessment data of each area in the disaster area, integrate the risk factors at different levels through the multi-level risk assessment model, and obtain the fuzziness of each risk factor through the fuzzy logic reasoning method, and finally obtain the comprehensive risk assessment value R total As input for material dispatch: Among them, μ1, μ2, μ3 and μ4 are weighting coefficients, and are the transportation, weather, safety and inventory risks of the i-th area in the disaster area, respectively, and n is the number of areas in the disaster area; S62. Based on comprehensive risk assessment, a multi-objective genetic algorithm is used to optimize the material distribution path, and a fitness function is constructed to evaluate the pros and cons of each material distribution plan. The fitness function comprehensively considers transportation costs, risk assessment and resource utilization efficiency: Among them, F i represents the fitness of the i-th material allocation plan, c ik is the transportation cost from warehouse i to destination k, x ik is the quantity of materials allocated from warehouse i to destination k, u ik is the resource utilization rate, ρ1, ρ2 and ρ3 are weighting coefficients; S63. Use particle swarm optimization algorithm for global optimization and adjust the delivery path: in, represents the velocity of particle i in the kth generation, represents the position of particle i in the kth generation, is the individual best historical position of particle i, gbest k is the historical optimal position of all particles, ω is the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random factors, is the updated particle velocity, is the updated particle position; S64. By optimizing the material distribution path, adjusting the dispatch strategy in real time, and replanning the distribution path, comprehensively considering the impact of traffic and weather changes in the disaster area, the resource utilization and efficiency can be maximized; S65. According to the optimization results, the material distribution plan is adjusted through the adaptive path correction mechanism, and the material distribution path is replanned in combination with real-time disaster changes.

Images