A population evacuation path optimization method based on multi-model fusion

Abstract

This invention discloses a population evacuation route optimization method based on multi-model fusion, comprising: preprocessing collected data to obtain a standardized spatiotemporal matrix of population flow; extracting structured event element fields from textual public opinion information; constructing a population flow-event coupled prediction model; outputting the population flow distribution for future periods; performing route optimization calculations to generate an optimal evacuation route set; matching the optimal evacuation route set with an emergency resource database to establish a resource allocation optimization model; and solving the established resource allocation optimization model using a collaborative scheduling algorithm to generate time-segmented and region-segmented resource deployment plans. This invention comprehensively considers multiple factors such as population distribution, geographical environment, traffic conditions, and emergencies, and achieves real-time route optimization through a hierarchical weighted dynamic algorithm, improving evacuation efficiency and the scientific nature of decision-making, and maximizing the protection of people's lives and minimizing property losses in emergency situations.

Description

Technical Field

[0001] This invention belongs to the field of population evacuation planning and management technology, specifically involving a population evacuation path optimization method based on multi-model fusion. Background Technology

[0002] In the field of population dispersal planning, the most commonly used algorithmic solutions are those based on a single model or simple rules. Taking a solution based on Geographic Information System (GIS) combined with classic shortest path algorithms (such as Dijkstra's algorithm) as an example, its core structure includes a geographic data acquisition module, a path planning algorithm module, and a result output module.

[0003] The geographic data acquisition module primarily utilizes satellite remote sensing technology, ground surveying equipment (such as total stations and levels), and data from map data providers to collect geographic information within the evacuation area. This geographic information includes road networks (road topology, length, width, traffic restrictions, etc.), building distribution (location, type, height, and area of ​​buildings), and topography (distribution and characteristics of mountains, plains, rivers, etc.). This data is stored in a database in a specific format, such as the common geospatial data format Shapefile, providing basic geographic information for subsequent route planning.

[0004] Path planning algorithm module: typically employs the classic Dijkstra's algorithm or The algorithm, taking Dijkstra's algorithm as an example, starts from the set evacuation starting point and gradually searches for the shortest path to all other nodes. During the search process, the algorithm maintains a priority queue and continuously updates the shortest distance from each node to the starting point until the shortest path to the evacuation destination is found. This module receives geographical information provided by the geographical data acquisition module and performs path calculation based on the set starting and ending points. For example, in a city's population evacuation scenario, the residential area is taken as the starting point and the safe area around the city is taken as the ending point. Dijkstra's algorithm is used to calculate the shortest path from the residential area to the safe area.

[0005] Results output module: Displays the shortest path calculated by the path planning algorithm module on a map in a visual manner.

[0006] The data transmission of each of the above modules is unidirectional, focusing only on geographical factors and lacking comprehensive consideration of population distribution and real-time environmental changes.

[0007] Currently, existing algorithmic solutions based on a single model or simple rules have the following drawbacks:

[0008] Lack of comprehensive consideration: Focusing only on geographical factors, there is a lack of comprehensive analysis of population distribution characteristics (such as population density, age structure, and population mobility patterns), types of emergencies, and the scope of their impact (different types of emergencies have very different ways and degrees of impact on population evacuation; for example, fires may mainly affect the emergency evacuation of people in local buildings, while floods may lead to large-scale population transfers over a larger area; earthquakes may cause people to be trapped and have difficulty evacuating in the core damaged area, and the surrounding areas may also experience changes in population flow due to aftershocks and other factors).

[0009] Insufficient dynamic adaptability: It is difficult to quickly adjust evacuation routes and resource allocation according to real-time changes; during the evacuation process, new emergencies may occur, such as the sudden appearance of new danger zones or congestion of some evacuation routes due to the gathering of people; existing algorithms cannot respond to these changes in a timely manner, cannot replan the optimal route and reasonably adjust resource allocation, causing chaos in the evacuation work and increasing the risk of casualties and property damage.

[0010] Poor accuracy: The single model simplifies the actual scenario, and the planning results deviate significantly from the actual evacuation needs. In practical applications, it may lead to some people being evacuated to dangerous areas, or the evacuation route being too long and time-consuming, failing to achieve the goal of rapid and safe evacuation. Summary of the Invention

[0011] To address the shortcomings of existing technologies, the present invention aims to provide a population evacuation path optimization method based on multi-model fusion, thereby solving the problems of insufficient comprehensive consideration, poor dynamic adaptability, and inadequate accuracy in existing population evacuation algorithms. This invention comprehensively considers multiple factors such as population distribution, geographical environment, traffic conditions, and emergencies, and employs hierarchical weighted dynamic optimization. The algorithm enables real-time route optimization, improving evacuation efficiency and the scientific nature of decision-making, and maximizing the protection of people's lives and minimizing property losses in emergency situations.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] The present invention provides a population evacuation path optimization method based on multi-model fusion, comprising the following steps:

[0014] 1) Collect data from multiple sources, preprocess the collected data to obtain a standardized spatiotemporal matrix of population flow; extract text-based public opinion information in a structured manner to generate structured event element fields;

[0015] 2) Construct a macro-prediction model of population flow and a model of the impact of emergencies, and fuse them using the inverse error weighting method to form a population flow-event coupled prediction model;

[0016] 3) Input the standardized spatiotemporal matrix of population flow obtained in step 1) into the population flow-event coupled prediction model in step 2) to output the population flow distribution in the future period; and use this distribution as the initial value to design a hierarchical weighted dynamic algorithm to perform path optimization calculation and generate the optimal evacuation path set.

[0017] 4) Match the set of optimal evacuation routes generated in step 3) with the emergency resource database to establish a resource allocation optimization model; use the collaborative scheduling algorithm to solve the established resource allocation optimization model to generate resource deployment plans by time period and by region.

[0018] Further, step 1) specifically includes:

[0019] 11) Obtain raw numerical data on population distribution, geographical environment, and traffic conditions through temperature sensors, smoke sensors, displacement sensors, drone aerial photography, and mobile terminal positioning; at the same time, use online public opinion monitoring tools to capture text information related to emergencies from social media, news portals, and SMS platforms as raw data for emergencies.

[0020] 12) An outlier detection algorithm was used to clean the raw data and remove invalid samples; a coordinate transformation algorithm was used to unify the multi-source spatial data to the CGCS2000 coordinate system; text-based public opinion information was transformed into event element fields through structured extraction methods; population density was normalized to [0,1], and the normalized population density rasters for each time period were stacked in chronological order to form a three-dimensional array. Where H and W are the number of rows and columns of the raster, respectively, and T is the number of time slices; the three-dimensional array D is Z-score standardized so that the mean of each spatiotemporal location is 0 and the variance is 1, thus obtaining the standardized spatiotemporal matrix of population flow. The normalized population density value is calculated from the population density data using the following formula:

[0021] ;

[0022] in, The original population density of grid or traffic cell i; The normalized population density takes values ​​in the range [0,1]. This is the normalized value for population density; when node n falls on grid i, = , representing the normalized population density of the region where node n is located.

[0023] Among them, a grid refers to dividing the study area into square grids of equal size, with each grid cell called a grid cell, which is the smallest unit of spatial quantification; a traffic zone is a commonly used term in urban traffic planning, referring to a closed area divided by natural barriers such as roads and rivers, with an area generally of 0.1–1 km², which is larger than a grid cell and can contain multiple grid cells.

[0024] Furthermore, the structured event element fields in step 1) include: event type, center coordinates, influence radius, and intensity level.

[0025] Further, step 2) specifically includes:

[0026] 21) Construct a macro-prediction model for population flow: Stack the normalized population density values ​​obtained in step 1) by time slices to form a spatiotemporal matrix sequence; use the spatiotemporal graph convolutional network ST-GCN to train the spatiotemporal matrix sequence. The input is the spatiotemporal matrix of the past N time periods. After training by the spatiotemporal graph convolutional network ST-GCN, the output is the predicted spatiotemporal distribution of population density for the next M time periods.

[0027] 22) Constructing an impact model of the emergency: Extract the event type, center coordinates, impact radius, and intensity level from the structured event element fields in step 1). Generate a dynamic event mask matrix with the same spatial resolution as the population density matrix according to the grid size. Concatenate this mask matrix with the population density distribution matrix of the corresponding time period to form an event-population coupling tensor. Use a three-dimensional convolutional neural network to train the event-population coupling tensor and output the population density change of each grid under the condition of the emergency to obtain the impact field of the emergency on population flow.

[0028] 23) Model Fusion: Combine the spatiotemporal distribution predictions of population density obtained in step 21). , Compared with the impact field of the emergency on population movement obtained in step 22), Assign fusion weights based on the reciprocal of the prior accuracy of the scenario. , Pixel-level linear overlay is performed to obtain the population flow-event coupled prediction model, as follows:

[0029] ;

[0030] in, Let be the normalized population density prediction matrix for each grid at time t.

[0031] Furthermore, step 3) specifically includes:

[0032] 31) Construct a spatiotemporal search graph: Overlay the normalized population density matrix obtained in step 1) with the road vector layer to generate a directed graph G=(V,E) with a time dimension. The node attributes of the directed graph include coordinates, time, population density, traffic flow and event intensity.

[0033] 32) Calculate the actual cost The details are as follows:

[0034] 321) Extract the current node with its parent node The path length normalization value was obtained by using Min-Max normalization, Z-score congestion classification, direct reading, and event mask inflation algorithms to measure the segment length, real-time traffic flow, population density, and event intensity. Traffic congestion index Normalized population density and road traffic condition coefficient ;

[0035] Path length dimension normalization: The measured length of the road segment is linearly mapped to the (0–1) interval to eliminate dimension differences and maintain comparability. The mapping result is denoted as . as follows:

[0036] ;

[0037] in, Let n be the measured length of the road segment from node n to its parent node n'. , These are the minimum and maximum lengths of all road segments within the study area, respectively.

[0038] Traffic congestion index The calculation steps are as follows:

[0039] Calculate the traffic saturation of road segment (n,n′) at time t. The expression is as follows:

[0040] ;

[0041] in, Let be the real-time traffic flow of road segment (n,n′) at time t; C is the traffic capacity of road segment;

[0042] Take historical traffic flow samples of road segment (n, n′) over a certain period of time and calculate its mean. with standard deviation Then use real-time traffic flow By standardizing the distribution using the Z-score, we obtain the traffic fluctuation index at time t. The expression is as follows:

[0043] ;

[0044] Traffic congestion index The expression is as follows:

[0045] ;

[0046] Population density normalized value : ;

[0047] Road traffic condition coefficient The calculation steps are as follows:

[0048] Dynamic mask matrix for events Morphological dilation is expressed as follows:

[0049] ;

[0050] in, The event impact value of grid i after dilation at time t; The original event mask value for raster j at time t; Let i be the set of all grids j within the square structural element centered at grid i; structural element Adjacent raster index within;

[0051] Road traffic condition coefficient The calculation expression is as follows:

[0052] ;

[0053] 322) Four-dimensional features [ , , , Input an online logistic regression model and output normalized weights ω1, ω2, ω3, ω4, satisfying ω1+ω2+ω3+ω4=1;

[0054] 323) Starting from the evacuation point Based on the baseline, calculate The actual cost to reach the current node n ,as follows:

[0055] ;

[0056] 33) Calculate the heuristic cost estimate The details are as follows:

[0057] 331) Extract the planar coordinates of node n Coordinates of the center of the pre-set evacuation endpoint safety zone ;

[0058] 332) The Euclidean distance algorithm is used to calculate the straight-line distance from node n to the evacuation endpoint. ,as follows:

[0059] ;

[0060] 333) Based on the shortest road link from node n to the evacuation endpoint, and according to the event dynamic mask matrix, calculate the road traffic status coefficient of each grid on the link. , Then, the arithmetic mean algorithm is used to obtain the average traffic condition coefficient. , ,in This represents the total number of grid cells the link passes through.

[0061] 334) will and Input a linear correction model, whose coefficients are obtained through offline training using historical evacuation samples, and output a correction factor. ;

[0062] 335) Calculate the heuristic cost estimate from node n to the evacuation endpoint. ,as follows:

[0063] ;

[0064] 34) Dynamic weight adjustment, as detailed below:

[0065] 341) Extraction Real-time traffic flow, event intensity, and population density at time 1 are used to obtain the [data point name] using the Z-score normalization algorithm. Influence strength of class factors ;

[0066] 342) Will affect the intensity Input linear dynamic correction model, model coefficients The output correction term is obtained through offline training using historical evacuation cases. ;

[0067] 343) Calculate dynamic weights and normalization guarantee ; Use the updated Recalculate the actual cost ; This represents the baseline value of the k-th factor in the initial static weights; This represents the influence intensity of the k-th type of factor at time t after Z-score standardization, used to quantify the degree of disturbance of this type of factor to the current traffic state;

[0068] 35) The path update trigger conditions are as follows:

[0069] 351) Obtain the real-time traffic flow and event dynamic mask matrix, and calculate the absolute change in the traffic congestion index. and the absolute change of the road traffic condition coefficient The expression is as follows:

[0070] ;

[0071] ;

[0072] in, For the current moment Update time with the previous path The absolute change in the traffic congestion index between periods; For the current moment The traffic congestion index is set to a value of ; Update time for the previous path The traffic congestion index; It represents the absolute change in the road traffic condition coefficient within the same time period; For the current moment The road traffic condition coefficient is set to a value of . Update time for the previous path The road traffic condition coefficient;

[0073] 352) The 95th quantile was calculated for 100 samples of traffic congestion index change Δc and road traffic state coefficient change Δs using the sliding window statistical method, and adaptive thresholds were obtained respectively. , The periodic threshold T is taken as the 95th percentile of the response time of historical evacuation cases.

[0074] 353) Recalculate the path if any of the following conditions are met:

[0075] or or ;

[0076] Recalculate according to steps 31)-34). , , And adopt a hierarchical weighted dynamic The algorithm re-searches and obtains the updated set of optimal evacuation paths, while also letting Generate the optimal evacuation path set.

[0077] Further, step 4) specifically includes:

[0078] 41) Based on the population evacuation needs and the existing resource situation, establish a resource allocation optimization model; wherein, the existing resource situation includes: material reserves, number of vehicles and personnel allocation;

[0079] The resource allocation optimization model's material constraints include: the total number of evacuated people. Planned guarantee days (T), average daily demand per person Uncertainty margin coefficient The total reserve of Category j materials, Q j The unit volume of the j-th type of material The weight of a single piece of the jth type of material The unit price of the jth type of goods Available budget limit B, maximum load capacity W, maximum storage volume Number of items purchased for category j ;

[0080] The core constraints of the resource allocation optimization model include:

[0081] Demand guarantee constraint: , ;

[0082] Budget constraints: ;

[0083] Weight constraints: ;

[0084] Volumetric constraints: ;

[0085] Non-negative integer constraint: ;

[0086] 42) The material constraints of the resource allocation optimization model are as follows:

[0087] Demand guarantee constraint: , ;

[0088] Budget constraints: ;

[0089] Weight constraints: ;

[0090] Volumetric constraints: ;

[0091] Non-negative integer constraint: , ;

[0092] 43) The optimization objectives of the resource allocation optimization model include:

[0093] Minimize total cost: ;

[0094] Minimize total weight: ;

[0095] Minimize total volume: ;

[0096] Multi-objective weighted: , ;

[0097] 44) Model types and optional extensions;

[0098] 441) The resource allocation optimization model is an integer linear programming model, and its decision variables are... Take non-negative integers; both the objective function and constraints are linear.

[0099] 442) Optional extensions;

[0100] 4421) Introducing a time decay coefficient The expression is as follows:

[0101] ;

[0102] Where τ is the remaining shelf life, Permissible storage time for Category J materials;

[0103] Replace the demand guarantee constraint with: This allows inventory to be dynamically adjusted downwards as the shelf life expires.

[0104] 4422) The total number of people to be evacuated Divided by geographical region Corresponding to the 1st to the 2nd in sequence There are 1 distribution node, among which Total number of nodes For the first The number of people served by each node; then, assign all the demand guarantee, budget, weight, volume, and non-negative integer constraints, along with the cost, weight, and volume objective functions, to the node indices. This leads to a joint integer linear programming approach for inventory and transportation, enabling synchronous optimization of each node.

[0105] 45) Model solving and dynamic scheduling;

[0106] 451) By adding transportation capacity constraints to the resource allocation optimization model, a mixed-integer linear programming model is obtained, with the following expression:

[0107] ;

[0108] in, This refers to the vehicle type category number available for dispatch. ; For the first The number of vehicles of this type that can be called; For the first Passenger capacity per vehicle of this type; For the number of evacuees to be transported in the current batch, the number of evacuees to be transported in a single trip. ;

[0109] 452) In the resource allocation optimization model, material constraints, transportation capacity constraints, and personnel constraints are simultaneously added to form a minimization objective function. Maximize evacuation efficiency A bi-objective mixed-integer linear programming problem is solved using the NSGA-II fast non-dominated sorting genetic algorithm, with the fitness function as follows:

[0110] ;

[0111] in, To normalize evacuation efficiency, This represents the theoretical maximum evacuation efficiency. The total cost of the current solution is... The upper limit of the budget; It is set by the command center according to preferences;

[0112] The Pareto solution set is obtained by running the fast non-dominated sorting genetic algorithm NSGA-Ⅱ offline. Real-time road conditions, vehicle GPS locations and material inventory are obtained every ΔT minutes. Fitness is recalculated and selection, crossover, mutation and non-dominated sorting are performed. New solutions are output in a rolling manner to generate resource deployment plans by time period and region, and to complete the dynamic scheduling of materials, vehicles and personnel.

[0113] Furthermore, the regional resource allocation plan in step 4) specifically involves dividing the evacuation area into several geographical units, and for each unit and each time period, specifying how much material, how many vehicles, and how many rescue personnel to allocate; thus forming a spatiotemporal resource table.

[0114] The beneficial effects of this invention are:

[0115] 1. This invention, through multi-source data fusion, can comprehensively and accurately obtain real-time information such as the geographical environment, population distribution, and traffic conditions of the evacuation area. It can simulate the flow of population during emergencies, consider the uneven distribution of population and various influencing factors, and dynamically adjust the evacuation speed and direction to make the evacuation process more in line with the actual situation. Based on real-time traffic flow and population flow information, it can adjust the evacuation route in real time to avoid congested road sections and ensure that the evacuation route always remains in the optimal state, which greatly shortens the evacuation time and improves the evacuation efficiency.

[0116] 2. The resource allocation optimization model of the present invention aims to minimize resource usage costs and maximize evacuation efficiency. It comprehensively considers various factors such as material reserves, number of vehicles, and personnel allocation to obtain the optimal resource allocation scheme. Attached Figure Description

[0117] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0118] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0119] Reference Figure 1 As shown, the present invention provides a population evacuation path optimization method based on multi-model fusion, comprising the following steps:

[0120] 1) Collect data from multiple sources, preprocess the collected data to obtain a standardized spatiotemporal matrix of population flow; extract structured event element fields from text-based public opinion information; specifically including:

[0121] 11) Obtain raw numerical data on population distribution, geographical environment, and traffic conditions through temperature sensors, smoke sensors, displacement sensors, drone aerial photography, and mobile terminal positioning; at the same time, use online public opinion monitoring tools to capture text information related to emergencies from social media, news portals, and SMS platforms as raw data for emergencies.

[0122] 12) An outlier detection algorithm was used to clean the raw data and remove invalid samples; a coordinate transformation algorithm was used to unify the multi-source spatial data to the CGCS2000 coordinate system; text-based public opinion information was transformed into event element fields through structured extraction methods; population density was normalized to [0,1], and the normalized population density rasters for each time period were stacked in chronological order to form a three-dimensional array. Where H and W are the number of rows and columns of the raster, respectively, and T is the number of time slices; the three-dimensional array D is Z-score standardized so that the mean of each spatiotemporal location is 0 and the variance is 1, thus obtaining the standardized spatiotemporal matrix of population flow. The normalized population density value is calculated from the population density data using the following formula:

[0123] ;

[0124] in, The original population density of grid or traffic cell i; The normalized population density takes values ​​in the range [0,1]. This is the normalized value for population density; when node n falls on grid i, = , representing the normalized population density of the region where node n is located.

[0125] Among them, a grid refers to dividing the study area into square grids of equal size (such as 100m×100m), and each grid is called a grid, which is the smallest unit of spatial quantification; a traffic zone is a commonly used term in urban traffic planning, referring to a closed area divided by natural barriers such as roads and rivers, with an area of ​​generally 0.1–1km², which is larger than a grid and can contain multiple grids.

[0126] Specifically, the structured event element fields in step 1) include: event type, center coordinates, influence radius, and intensity level.

[0127] 2) Construct a macro-level population flow forecasting model and a model for the impact of unforeseen events, and fuse them using the inverse error weighting method to form a coupled population flow-event forecasting model; specifically including:

[0128] 21) Construct a macro-prediction model for population flow: Stack the normalized population density values ​​obtained in step 1) by time slices to form a spatiotemporal matrix sequence; use the spatiotemporal graph convolutional network ST-GCN to train the spatiotemporal matrix sequence. The input is the spatiotemporal matrix of the past N time periods. After training by the spatiotemporal graph convolutional network ST-GCN, the output is the predicted spatiotemporal distribution of population density for the next M time periods.

[0129] 22) Constructing an impact model of the emergency: Extract the event type, center coordinates, impact radius, and intensity level from the structured event element fields in step 1). Generate a dynamic event mask matrix with the same spatial resolution as the population density matrix according to the grid size. Concatenate this mask matrix with the population density distribution matrix of the corresponding time period to form an event-population coupling tensor. Train the event-population coupling tensor with a three-dimensional convolutional neural network (3D-CNN) to output the population density change of each grid under the condition of the emergency, and obtain the impact field of the emergency on population flow.

[0130] 23) Model Fusion: Combine the spatiotemporal distribution predictions of population density obtained in step 21). , Compared with the impact field of the emergency on population movement obtained in step 22), Assign fusion weights based on the reciprocal of the prior accuracy of the scenario. , ( (Given by the Bayesian accuracy estimate during the validation period), pixel-level linear superposition yields the population flow-event coupled prediction model, as follows:

[0131] ;

[0132] in, Let be the normalized population density prediction matrix for each grid at time t.

[0133] 3) Input the standardized spatiotemporal matrix of population flow obtained in step 1) into the population flow-event coupled prediction model in step 2), and output the population flow distribution for future periods; using this distribution as the initial value, design a hierarchical weighted dynamic algorithm to perform path optimization calculations and generate the optimal evacuation path set; specifically including:

[0134] 31) Construct a spatiotemporal search graph: Overlay the normalized population density matrix obtained in step 1) with the road vector layer to generate a directed graph G=(V,E) with a time dimension. The node attributes of the directed graph include coordinates, time, population density, traffic flow and event intensity.

[0135] 32) Calculate the actual cost The details are as follows:

[0136] 321) Extract the current node with its parent node The path length normalization value was obtained by using Min-Max normalization, Z-score congestion classification, direct reading, and event mask inflation algorithms to measure the segment length, real-time traffic flow, population density, and event intensity. Traffic congestion index Normalized population density and road traffic condition coefficient ;

[0137] Path length dimension normalization: The measured length of the road segment is linearly mapped to the (0–1) interval to eliminate dimension differences and maintain comparability. The mapping result is denoted as . as follows:

[0138] ;

[0139] in, Let n be the measured length of the road segment from node n to its parent node n'. , These are the minimum and maximum lengths of all road segments within the study area, respectively.

[0140] Traffic congestion index The calculation steps are as follows:

[0141] Calculate the traffic saturation of road segment (n,n′) at time t. The expression is as follows:

[0142] ;

[0143] in, Let be the real-time traffic flow of road segment (n,n′) at time t; C is the traffic capacity of road segment;

[0144] Take historical traffic flow samples of road segment (n, n′) over a certain period (30 days) and calculate its mean. with standard deviation Then use real-time traffic flow By standardizing the distribution using the Z-score, we obtain the traffic fluctuation index at time t. The expression is as follows:

[0145] ;

[0146] Traffic congestion index The expression is as follows:

[0147] ;

[0148] Population density normalized value : ;

[0149] Road traffic condition coefficient The calculation steps are as follows:

[0150] Dynamic mask matrix for events Morphological dilation is expressed as follows:

[0151] ;

[0152] in, The event impact value of grid i after dilation at time t; The original event mask value for raster j at time t; Let i be the set of all grids j within the square structuring element (3×3 neighborhood) centered at grid i; structural element Adjacent raster index within;

[0153] Road traffic condition coefficient The calculation expression is as follows:

[0154] ;

[0155] 322) Four-dimensional features [ , , , Input an online logistic regression model and output normalized weights ω1, ω2, ω3, ω4, satisfying ω1+ω2+ω3+ω4=1;

[0156] 323) Starting from the evacuation point Based on the baseline, calculate The actual cost to reach the current node n ,as follows:

[0157] ;

[0158] 33) Calculate the heuristic cost estimate The details are as follows:

[0159] 331) Extract the planar coordinates of node n Coordinates of the center of the pre-set evacuation endpoint safety zone ;

[0160] 332) The Euclidean distance algorithm is used to calculate the straight-line distance from node n to the evacuation endpoint. ,as follows:

[0161] ;

[0162] 333) Based on the shortest road link from node n to the evacuation endpoint, and according to the event dynamic mask matrix, calculate the road traffic status coefficient of each grid on the link. , Then, the arithmetic mean algorithm is used to obtain the average traffic condition coefficient. , ,in This represents the total number of grid cells the link passes through.

[0163] 334) will and Input a linear correction model, whose coefficients are obtained through offline training using historical evacuation samples, and output a correction factor. ;

[0164] 335) Calculate the heuristic cost estimate from node n to the evacuation endpoint. ,as follows:

[0165] ;

[0166] 34) Dynamic weight adjustment, as detailed below:

[0167] 341) Extraction Real-time traffic flow, event intensity, and population density at time 1 are used to obtain the [data point name] using the Z-score normalization algorithm. Influence strength of class factors ;

[0168] 342) Will affect the intensity Input linear dynamic correction model, model coefficients The output correction term is obtained through offline training using historical evacuation cases. ;

[0169] 343) Calculate dynamic weights and normalization guarantee ; Use the updated Recalculate the actual cost ; This represents the baseline value of the k-th factor in the initial static weights; This represents the influence intensity of the k-th type of factor at time t after Z-score standardization, used to quantify the degree of disturbance of this type of factor to the current traffic state;

[0170] 35) The path update trigger conditions are as follows:

[0171] 351) Obtain the real-time traffic flow and event dynamic mask matrix, and calculate the absolute change in the traffic congestion index. and the absolute change of the road traffic condition coefficient The expression is as follows:

[0172] ;

[0173] ;

[0174] in, For the current moment Update time with the previous path The absolute change in the traffic congestion index between periods; For the current moment The traffic congestion index is set to a value of ; Update time for the previous path The traffic congestion index; It represents the absolute change in the road traffic condition coefficient within the same time period; For the current moment The road traffic condition coefficient is set to a value of . Update time for the previous path The road traffic condition coefficient;

[0175] 352) The 95th quantile was calculated for 100 samples of traffic congestion index change Δc and road traffic state coefficient change Δs using the sliding window statistical method, and adaptive thresholds were obtained respectively. , The periodic threshold T is taken as the 95th percentile of the response time of historical evacuation cases.

[0176] 353) Recalculate the path if any of the following conditions are met:

[0177] or or ;

[0178] Recalculate according to steps 31)-34). , , And adopt a hierarchical weighted dynamic The algorithm re-searches and obtains the updated set of optimal evacuation paths, while also letting Generate the optimal evacuation path set.

[0179] 4) Match the optimal evacuation route set generated in step 3) with the emergency resource database to establish a resource allocation optimization model; solve the established resource allocation optimization model using a collaborative scheduling algorithm to generate time-segmented and region-segmented resource deployment plans; specifically including:

[0180] 41) Based on the population evacuation needs and the existing resource status, establish a resource allocation optimization model; wherein, the existing resource status includes: material reserves (such as the amount and location of reserves of food, water, medical supplies, etc.), the number of transportation vehicles (the number and passenger capacity of different types of transportation vehicles (such as cars, buses, helicopters, etc.)), and personnel allocation (such as the number and professional skills of rescue personnel, medical personnel, guides, etc.).

[0181] The resource allocation optimization model's material constraints include: the total number of evacuated people. Planned guarantee days (T), average daily demand per person Uncertainty margin coefficient The total reserve of Category j materials, Q j The unit volume of the j-th type of material The weight of a single piece of the jth type of material The unit price of the jth type of goods Available budget limit B, maximum load capacity W, maximum storage volume Number of items purchased for category j ;

[0182] The core constraints of the resource allocation optimization model include:

[0183] Demand guarantee constraint: , ;

[0184] Budget constraints: ;

[0185] Weight constraints: ;

[0186] Volumetric constraints: ;

[0187] Non-negative integer constraint: ;

[0188] 42) The material constraints of the resource allocation optimization model are as follows:

[0189] Demand guarantee constraint: , ;

[0190] Budget constraints: ;

[0191] Weight constraints: ;

[0192] Volumetric constraints: ;

[0193] Non-negative integer constraint: , ;

[0194] 43) The optimization objectives of the resource allocation optimization model include:

[0195] Minimize total cost: ;

[0196] Minimize total weight: ;

[0197] Minimize total volume: ;

[0198] Multi-objective weighted: , ;

[0199] 44) Model types and optional extensions;

[0200] 441) The resource allocation optimization model is an integer linear programming (ILP) model, and its decision variables are... Take non-negative integers; both the objective function and constraints are linear.

[0201] 442) Optional extensions;

[0202] 4421) Introducing a time decay coefficient The expression is as follows:

[0203] ;

[0204] Where τ is the remaining shelf life (i.e., the number of days remaining until the expiration of material j at the current moment). Permissible storage time for Category J materials;

[0205] Replace the demand guarantee constraint with: This allows inventory to be dynamically adjusted downwards as the shelf life expires.

[0206] 4422) The total number of people to be evacuated Divided by geographical region Corresponding to the 1st to the 2nd in sequence There are 1 distribution node, among which Total number of nodes For the first The number of people served by each node; then, assign all the demand guarantee, budget, weight, volume, and non-negative integer constraints, along with the cost, weight, and volume objective functions, to the node indices. This leads to a joint integer linear programming approach for inventory and transportation, enabling synchronous optimization of each node.

[0207] 45) Model solving and dynamic scheduling;

[0208] 451) By adding transportation capacity constraints to the resource allocation optimization model, a mixed-integer linear programming model is obtained, with the following expression:

[0209] ;

[0210] in, This refers to the vehicle type category number available for dispatch. ; For the first The number of vehicles of this type that can be called; For the first Passenger capacity per vehicle of this type; For the number of evacuees to be transported in the current batch, the number of evacuees to be transported in a single trip. ;

[0211] 452) In the resource allocation optimization model, material constraints, transportation capacity constraints, and personnel constraints are simultaneously added to form a minimization objective function. Maximize evacuation efficiency A bi-objective mixed-integer linear programming problem is solved using the NSGA-II fast non-dominated sorting genetic algorithm, with the fitness function as follows:

[0212] ;

[0213] in, To normalize evacuation efficiency, This represents the theoretical maximum evacuation efficiency. The total cost of the current solution is... The upper limit of the budget; It is set by the command center according to preferences;

[0214] The Pareto solution set is obtained by running the fast non-dominated sorting genetic algorithm NSGA-Ⅱ offline. Real-time road conditions, vehicle GPS locations and material inventory are obtained every ΔT minutes. Fitness is recalculated and selection, crossover, mutation and non-dominated sorting are performed. New solutions are output in a rolling manner to generate resource deployment plans by time period and region, and to complete the dynamic scheduling of materials, vehicles and personnel.

[0215] Specifically, the regional resource allocation plan in step 4) is as follows: the evacuation area is divided into several geographical units, and for each unit and each time period, the amount of materials, vehicles, and rescue personnel to be allocated is specified; thus forming a spatiotemporal-resource table.

[0216] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A population evacuation route optimization method based on multi-model fusion, characterized in that, The steps are as follows: 1) Collect data from multiple sources, preprocess the collected data to obtain a standardized spatiotemporal matrix of population flow; extract text-based public opinion information in a structured manner to generate structured event element fields; 2) Construct a macro-prediction model of population flow and a model of the impact of emergencies, and fuse them using the inverse error weighting method to form a population flow-event coupled prediction model; 3) Input the standardized spatiotemporal matrix of population flow obtained in step 1) into the population flow-event coupled prediction model in step 2) to output the population flow distribution in the future period; and use this distribution as the initial value to design a hierarchical weighted dynamic algorithm to perform path optimization calculation and generate the optimal evacuation path set. 4) Match the set of optimal evacuation routes generated in step 3) with the emergency resource database to establish a resource allocation optimization model; use the collaborative scheduling algorithm to solve the established resource allocation optimization model to generate resource deployment plans by time period and by region.

2. The population evacuation path optimization method based on multi-model fusion according to claim 1, characterized in that, Step 1) specifically includes: 11) Obtain raw numerical data on population distribution, geographical environment, and traffic conditions through temperature sensors, smoke sensors, displacement sensors, drone aerial photography, and mobile terminal positioning; at the same time, use online public opinion monitoring tools to capture text information related to emergencies from social media, news portals, and SMS platforms as raw data for emergencies. 12) An outlier detection algorithm was used to clean the raw data and remove invalid samples; a coordinate transformation algorithm was used to unify the multi-source spatial data to the CGCS2000 coordinate system; text-based public opinion information was transformed into event element fields through structured extraction methods; population density was normalized to [0,1], and the normalized population density rasters for each time period were stacked in chronological order to form a three-dimensional array. Where H and W are the number of rows and columns of the raster, respectively, and T is the number of time slices; the three-dimensional array D is Z-score standardized so that the mean of each spatiotemporal location is 0 and the variance is 1, thus obtaining the standardized spatiotemporal matrix of population flow. The normalized population density value is calculated from the population density data using the following formula: ； in, The original population density of grid or traffic cell i; The normalized population density takes values ​​in the range [0,1]. This is the normalized value for population density; when node n falls on grid i, = , representing the normalized population density of the region where node n is located; Among them, a grid refers to dividing the study area into square grids of equal size, with each grid cell being called a grid cell, which is the smallest unit of spatial quantification; a traffic zone refers to a closed area divided by roads, which contains multiple grid cells.

3. The population evacuation path optimization method based on multi-model fusion according to claim 2, characterized in that, Step 2) specifically includes: 21) Construct a macro-prediction model for population flow: Stack the normalized population density values ​​obtained in step 1) by time slices to form a spatiotemporal matrix sequence; use the spatiotemporal graph convolutional network ST-GCN to train the spatiotemporal matrix sequence. The input is the spatiotemporal matrix of the past N time periods. After training by the spatiotemporal graph convolutional network ST-GCN, the output is the predicted spatiotemporal distribution of population density for the next M time periods. 22) Constructing an impact model of the emergency: Extract the event type, center coordinates, impact radius, and intensity level from the structured event element fields in step 1). Generate a dynamic event mask matrix with the same spatial resolution as the population density matrix according to the grid size. Concatenate this mask matrix with the population density distribution matrix of the corresponding time period to form an event-population coupling tensor. Use a three-dimensional convolutional neural network to train the event-population coupling tensor and output the population density change of each grid under the condition of the emergency to obtain the impact field of the emergency on population flow. 23) Model Fusion: Combine the spatiotemporal distribution predictions of population density obtained in step 21). , Compared with the impact field of the emergency on population movement obtained in step 22), Assign fusion weights based on the reciprocal of the prior accuracy of the scenario. , Pixel-level linear overlay is performed to obtain the population flow-event coupled prediction model, as follows: ； in, Let be the normalized population density prediction matrix for each grid at time t.

4. The population evacuation path optimization method based on multi-model fusion according to claim 3, characterized in that, Step 3) specifically includes: 31) Overlay the normalized population density matrix obtained in step 1) with the road vector layer to generate a directed graph G=(V,E) with time dimension. The node attributes of the directed graph include coordinates, time, population density, traffic flow and event intensity. 32) Calculate the actual cost ; 33) Calculate the heuristic cost estimate ; 34) Dynamic weight adjustment; 35) Path update trigger conditions.

5. The population evacuation path optimization method based on multi-model fusion according to claim 4, characterized in that, Step 32) specifically includes: 321) Extract the current node with its parent node The path length normalization value was obtained by using Min-Max normalization, Z-score congestion classification, direct reading, and event mask inflation algorithms to measure the segment length, real-time traffic flow, population density, and event intensity. Traffic congestion index Normalized population density and road traffic condition coefficient ; Path length dimension normalization: The measured length of the road segment is linearly mapped to the (0–1) interval to eliminate dimension differences and maintain comparability. The mapping result is denoted as . as follows: ； in, Let n be the measured length of the road segment from node n to its parent node n'. , These are the minimum and maximum lengths of all road segments within the study area, respectively. Traffic congestion index The calculation steps are as follows: Calculate the traffic saturation of road segment (n,n′) at time t. The expression is as follows: ； in, Let be the real-time traffic flow of road segment (n,n′) at time t; C is the traffic capacity of road segment; Take historical traffic flow samples of road segment (n, n′) over a certain period of time and calculate its mean. with standard deviation Then use real-time traffic flow By standardizing the distribution using the Z-score, we obtain the traffic fluctuation index at time t. The expression is as follows: ； Traffic congestion index The expression is as follows: ； Population density normalized value : ; Road traffic condition coefficient The calculation steps are as follows: Dynamic mask matrix for events Morphological dilation is expressed as follows: ； in, The event impact value of grid i after dilation at time t; The original event mask value for raster j at time t; Let i be the set of all grids j within the square structural element centered at grid i; structural element Adjacent raster index within; Road traffic condition coefficient The calculation expression is as follows: ； 322) Four-dimensional features [ , , , Input an online logistic regression model and output normalized weights ω1, ω2, ω3, ω4, satisfying ω1+ω2+ω3+ω4=1; 323) Starting from the evacuation point Based on the baseline, calculate The actual cost to reach the current node n ,as follows: 。 6. The population evacuation path optimization method based on multi-model fusion according to claim 5, characterized in that, Step 33) specifically includes: 331) Extract the planar coordinates of node n Coordinates of the center of the pre-set evacuation endpoint safety zone ; 332) The Euclidean distance algorithm is used to calculate the straight-line distance from node n to the evacuation endpoint. ,as follows: ； 333) Based on the shortest road link from node n to the evacuation endpoint, and according to the event dynamic mask matrix, calculate the road traffic status coefficient of each grid on the link. , Then, the arithmetic mean algorithm is used to obtain the average traffic condition coefficient. , ,in This represents the total number of grid cells the link passes through. 334) will and Input a linear correction model, whose coefficients are obtained through offline training using historical evacuation samples, and output a correction factor. ; 335) Calculate the heuristic cost estimate from node n to the evacuation endpoint. ,as follows: 。 7. The population evacuation path optimization method based on multi-model fusion according to claim 6, characterized in that, Step 34) specifically includes: 341) Extraction Real-time traffic flow, event intensity, and population density at time 1 are used to obtain the [data point name] using the Z-score normalization algorithm. Influence strength of class factors ; 342) Will affect the intensity Input linear dynamic correction model, model coefficients The output correction term is obtained through offline training using historical evacuation cases. ; 343) Calculate dynamic weights and normalization guarantee ; Use the updated Recalculate the actual cost ; This represents the baseline value of the k-th factor in the initial static weights; This represents the influence intensity of the k-th type of factor at time t after Z-score standardization, used to quantify the degree of disturbance of this type of factor to the current traffic state.

8. The population evacuation path optimization method based on multi-model fusion according to claim 7, characterized in that, Step 35) specifically includes: 351) Obtain the real-time traffic flow and event dynamic mask matrix, and calculate the absolute change in the traffic congestion index. and the absolute change of the road traffic condition coefficient The expression is as follows: ； ； in, For the current moment Update time with the previous path The absolute change in the traffic congestion index between different periods; For the current moment The traffic congestion index is set to a value of ; Update time for the previous path The traffic congestion index; It represents the absolute change in the road traffic condition coefficient within the same time period; For the current moment The road traffic condition coefficient is set to a value of . Update time for the previous path The road traffic condition coefficient; 352) The 95th quantile was calculated for 100 samples of traffic congestion index change Δc and road traffic state coefficient change Δs using the sliding window statistical method, and adaptive thresholds were obtained respectively. , The periodic threshold T is taken as the 95th percentile of the response time of historical evacuation cases. 353) Recalculate the path if any of the following conditions are met: or or ; Recalculate according to steps 31)-34). , , And adopt a hierarchical weighted dynamic The algorithm re-searches and obtains the updated set of optimal evacuation paths, while also letting Generate the optimal evacuation path set.

9. The population evacuation path optimization method based on multi-model fusion according to claim 8, characterized in that, Step 4) specifically includes: 41) Based on the population evacuation needs and the existing resource situation, establish a resource allocation optimization model; wherein, the existing resource situation includes: material reserves, number of vehicles and personnel allocation; The resource allocation optimization model's material constraints include: the total number of evacuated people. Planned guarantee days (T), average daily demand per person Uncertainty margin coefficient The total reserve of Category j materials, Q j The volume of a single piece of the j-th type of material The weight of a single piece of the j-th category of materials The unit price of the jth type of goods Available budget limit B, maximum load capacity W, maximum storage volume Number of items purchased for category j ; The core constraints of the resource allocation optimization model include: Demand guarantee constraint: , ; Budget constraints: ; Weight constraints: ; Volumetric constraints: ; Non-negative integer constraint: ; 42) The material constraints of the resource allocation optimization model are as follows: Demand guarantee constraint: , ; Budget constraints: ; Weight constraints: ; Volumetric constraints: ; Non-negative integer constraint: , ; 43) The optimization objectives of the resource allocation optimization model include: Minimize total cost: ; Minimize total weight: ; Minimize total volume: ; Multi-objective weighted: , ; 44) Model types and optional extensions; 441) The resource allocation optimization model is an integer linear programming model, and its decision variables are... Take non-negative integers; both the objective function and constraints are linear. 442) Optional extensions; 4421) Introducing a time decay coefficient The expression is as follows: ； Where τ is the remaining shelf life, Permissible storage time for Category J materials; Replace the demand guarantee constraint with: This allows inventory to be dynamically adjusted downwards as the shelf life expires. 4422) The total number of people to be evacuated Divided by geographical region Corresponding to the 1st to the 2nd in sequence There are 1 distribution node, among which Total number of nodes For the first The number of people served by each node; then, assign all the demand guarantee, budget, weight, volume, and non-negative integer constraints, along with the cost, weight, and volume objective functions, to the node indices. This leads to a joint integer linear programming approach for inventory and transportation, enabling synchronous optimization of each node. 45) Model solving and dynamic scheduling; 451) By adding transportation capacity constraints to the resource allocation optimization model, a mixed-integer linear programming model is obtained, with the following expression: ； in, This refers to the vehicle type category number available for dispatch. ; For the first The number of vehicles of this type that can be called; For the first Passenger capacity per vehicle of this type; For the number of evacuees to be transported in the current batch, the number of evacuees to be transported in a single trip. ; 452) In the resource allocation optimization model, material constraints, transportation capacity constraints, and personnel constraints are simultaneously added to form a minimization objective function. Maximize evacuation efficiency A bi-objective mixed-integer linear programming problem is solved using the NSGA-II fast non-dominated sorting genetic algorithm, with the fitness function as follows: ； in, To normalize evacuation efficiency, This represents the theoretical maximum evacuation efficiency. The total cost of the current solution is... The upper limit of the budget; It is set by the command center according to preferences; The Pareto solution set is obtained by running the fast non-dominated sorting genetic algorithm NSGA-Ⅱ offline. Real-time road conditions, vehicle GPS locations and material inventory are obtained every ΔT minutes. Fitness is recalculated and selection, crossover, mutation and non-dominated sorting are performed. New solutions are output in a rolling manner to generate resource deployment plans by time period and region, and to complete the dynamic scheduling of materials, vehicles and personnel.

Images