Rainstorm disaster loss assessment method and system based on machine learning

Abstract

The invention discloses a rainstorm disaster loss assessment method and system based on machine learning, and the method comprises the steps: taking data of a rainstorm disaster data source in a preset region in a specified time period as to-be-analyzed data; dividing the preset area according to the to-be-analyzed data based on the rainfall, taking the preset area greater than a rainfall threshold as a disaster area, otherwise, taking the preset area as an uncertain area, and calculating the destructive power of the uncertain area according to a hydrological mechanism and a drainage system. Adding the uncertain area greater than a destructive power threshold value into the disaster area; performing comprehensive risk analysis on the to-be-analyzed data of the disaster area to obtain a comprehensive risk index, calculating a damage degree index of the disaster area, and performing real-time dynamic evaluation on the disaster area according to the damage degree index and the comprehensive risk index to obtain a loss index; and constructing a rainstorm disaster loss evaluation model according to the loss index, and outputting an evaluation result.

Description

Technical Field

[0001] The present invention relates to the field of disaster loss assessment, and particularly to a method and system for assessing rainstorm disaster losses based on machine learning. Background Art

[0002] With the acceleration of global climate change and urbanization, the frequent occurrence of rainstorm disasters has become a major challenge threatening urban safety and social and economic stability. Traditional disaster assessment methods mostly rely on historical disaster statistics and empirical models, with defects such as single data dimension and lag in dynamic response. The regional division method based on rainfall intensity thresholds often ignores the hydrological cycle mechanism and the spatial heterogeneity of urban drainage systems, resulting in insufficient accuracy in identifying the disaster scope. At the same time, existing research mostly focuses on physical damage indicators and fails to effectively integrate social perception data such as real-time public opinion on social media and regional economic vulnerability, making there be a significant deviation between the loss assessment results and the post-disaster recovery needs.

[0003] In recent years, machine learning technology has made breakthrough progress in the field of disaster prediction. Among them, the application of convolutional neural networks in identifying disaster losses in remote sensing images and the processing ability of the LSTM model for hydrological time series data have significantly improved the efficiency of disaster monitoring. However, existing models generally face two major bottlenecks: the lack of physical mechanism constraints in the feature fusion of multi-source heterogeneous data leads to a decrease in model interpretability; the model generalization ability for different regional disaster scenarios is insufficient, and a large amount of labeled data is required for repeated training. However, the meta-learning framework shows unique transfer advantages in small-sample scenarios by constructing feature representations shared among tasks, which provides a new idea for constructing a cross-regional and multi-scenario rainstorm disaster assessment model.

[0004] The current research gaps are mainly reflected in three aspects: the dynamic coupling mechanism of multi-modal data has not been established, especially the spatio-temporal correlation modeling of geological permeability coefficients and urban pipe network topologies is insufficient; the dynamic quantification method of social and economic vulnerability in the process of disaster chain evolution needs to be improved urgently; the knowledge reuse efficiency of existing assessment models for historical disaster scenarios is low. Therefore, there is an urgent need for a new method for assessing rainstorm disaster losses, which breaks through the static assessment limitations of traditional methods by integrating hydrological mechanism models and meta-learning paradigms, and realizes multi-dimensional dynamic perception and cross-regional transfer assessment of disaster losses. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for assessing rainstorm disaster losses based on machine learning.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0007] The present invention includes the following steps:

[0008] Take the data of the heavy rain disaster data source in the preset area within the specified time period as the data to be analyzed; the data sources to be analyzed include meteorological data, geological data, topographic and geomorphic data, social and economic data, social media data, historical disaster data, ecological data, and building data;

[0009] Based on the rainfall, divide the preset area according to the data to be analyzed, take the preset area with a rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area. Calculate the destructive power of the uncertain area according to the hydrological mechanism and drainage system, and add the uncertain area with a destructive power greater than the destructive power threshold to the disaster area; including:

[0010] Calculate the exposure, sensitivity and adaptability of the building according to the building data, and calculate the destructive power through objective weighting:

[0011]

[0012] BR i = ρ4WE i + ρ5WN i - ρ6WS i

[0013] where the building vulnerability of the i-th uncertain area is BR i , the exposure of the i-th uncertain area is WE i , the sensitivity of the i-th uncertain area is WN i , the adaptability of the i-th uncertain area is WS i , the exposure weight is ρ4, the sensitivity weight is ρ5, the adaptability is ρ6, and the destructive power of the i-th uncertain area is The rainfall of the i-th uncertain area is P i ;

[0014] Conduct a comprehensive risk analysis on the data to be analyzed in the disaster area to obtain a comprehensive risk index, calculate the damage degree index of the disaster area, and conduct real-time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index to obtain a loss index;

[0015] Construct a heavy rain disaster loss assessment model based on meta-learning according to the loss index, input the data to be evaluated into the heavy rain disaster loss assessment model, and output the evaluation result.

[0016] Further, the method for dividing the preset area according to the data to be analyzed based on rainfall includes:

[0017] Perform spatio-temporal alignment on the data to be analyzed and the preset area. Use the expert method to conduct a primary division of the preset area to obtain the initial area. Calculate the rainfall in the initial area based on the data to be analyzed. Take the initial area with rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area.

[0018] Further, the method for the disaster area includes:

[0019] The rainfall threshold is divided into the ordinary rainfall threshold and the super-threshold heavy rainfall amount. The disaster area is divided into the ordinary disaster area and the super heavy rainfall disaster area;

[0020] The disaster area with rainfall higher than the ordinary rainfall threshold and lower than the super-threshold heavy rainfall amount is the ordinary disaster area; the disaster area with rainfall higher than the super-threshold heavy rainfall amount is the super heavy rainfall disaster area;

[0021] The ordinary rainfall threshold and the super-threshold heavy rainfall amount are dynamically adjusted according to the rainfall threshold drift due to climate warming.

[0022] Further, the method for comprehensively analyzing the risk of the data to be analyzed in the disaster area includes:

[0023] Calculate the disaster-induced factor index of the ordinary disaster area through meteorological data, calculate the geological vulnerability index of the disaster area through geological data, calculate the terrain stability index through topographic and geomorphic data, and calculate the building vulnerability through building data;

[0024] Calculate the ecological sensitivity, ecological resilience, and ecological stress degree through ecological data, and calculate the ecological vulnerability index based on the ecological sensitivity, ecological resilience, and ecological stress degree:

[0025]

[0026] Among them, the ecological vulnerability index of the u-th ordinary disaster area is The ecological sensitivity is C1, the ecological stress degree is C2, the ecological resilience is C3, the sensitivity weight is ρ1, the stress degree weight is ρ2, and the resilience weight is ρ3;

[0027] Use disaster data and social media data to correct the disaster-induced factor index, geological vulnerability index, and terrain stability index;

[0028] Calculate the comprehensive risk index of the super heavy rainfall disaster area based on the disaster-induced factor index, geological vulnerability index, terrain stability index, and ecological vulnerability index:

[0029]

[0030] Among them, the comprehensive risk index of the u-th ordinary disaster area is The disaster-induced factor index of the u-th ordinary disaster area is MHu For the u-th ordinary disaster area, the geological vulnerability index is DX u For the u-th ordinary disaster area, the terrain stability index is ZX u The induced weight is ω1, the geological weight is ω2, the terrain weight is ω3, and the ecological weight is The building weight is ω5, and the building vulnerability of the u-th ordinary disaster area is BR u ;

[0031] Calculate the comprehensive risk index of the extreme rainstorm disaster area:

[0032]

[0033] Among them, the comprehensive risk index of the u-th extreme rainstorm disaster area is Perform risk annotation on the disaster area according to the comprehensive risk index, and output the annotated disaster area.

[0034] Furthermore, the method for calculating the damage degree index of the disaster area includes:

[0035] Calculate the life loss of the disaster area based on the data to be analyzed:

[0036]

[0037] Among them, the life loss is LM s The number of disaster areas is N a The number of at-risk population is N p The mortality rate of the at-risk population is η, and the influence degree of the b-th main influencing factor on the mortality rate of the at-risk population is γ b The weight coefficient of the b-th main influencing factor is φ b The influence degree of the b-th secondary factor on the mortality rate of the at-risk population is ξ b The weight coefficient value of the b-th secondary factor is The number of secondary factors is N b2 The number of main factors is N b1 The weight coefficient is β, and the correction coefficient is

[0038] Calculate the economic loss of the disaster area:

[0039]

[0040] Among them, the economic loss is D s The number of pixel disaster areas where property is distributed in the disaster area is N1, the property types of the v-th pixel disaster area are N2, the number of different flood depth levels is N3, and the loss rate of the v-th type of property in the a-th disaster area corresponding to the x-th flood depth is The property value of the v-th type of property in the a-th disaster area corresponding to the x-th inundation depth is

[0041] Calculate the indirect economic loss of the disaster area:

[0042]

[0043] Among them, the indirect coefficient corresponding to the r-th type of property is K a (s), the number of property categories is M1, and the r-th type of property is A r , and the indirect economic loss is

[0044] Calculate the ecological environment loss rate:

[0045]

[0046] Among them, the ecological environment loss rate is E, and the loss rate of the c-th type of ecosystem service function is Q c , the number of ecosystem service functions is N4, the number of ecological products is N5, and the service volume of the i-th product in the c-th type of ecosystem service function is H ck , and the unit price of the i-th product in the c-th type of ecosystem service function is R ck ;

[0047] Calculate the damage degree index based on the loss of life, economic loss, indirect economic loss, and ecological environment loss rate:

[0048]

[0049] Among them, the total economic volume is KS, and the damage degree index is

[0050] Furthermore, a method for obtaining a loss index by performing real-time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index includes:

[0051] Input the damage degree index and the comprehensive risk index into a real-time dynamic assessment model based on a Bayesian network, and discretize the disaster state into three levels: extremely heavy rain, heavy rain, and rainstorm using a state-time curve;

[0052] Take time as a node and calculate the conditional probability of the transfer network of the node:

[0053]

[0054] Among them, the time interval is Δt, and the offset rate when the node d changes from the w state to the m state is The state of the node at time t0 is w, the state of the node at time t1 is j, and the probability that the state at time t1 is j given that the state at time t0 is w is k(G t1 = j|G t0 = w), the damage degree index at time t0 is The comprehensive risk index at time t0 is The damage degree index of node d in state w is The damage degree index of node d in state m is The comprehensive risk index of node d in state w is The comprehensive risk index of node d in state m is

[0055] Calculate the loss index according to the conditional probability of the transfer network:

[0056]

[0057] Among them, the damage degree index at time t in state j is The comprehensive risk index at time t in state j is The loss index at time t is The damage degree index at time t in state w is The comprehensive risk index at time t in state w is

[0058] Furthermore, a method for constructing a rainstorm disaster loss assessment model based on meta-learning according to the loss index includes:

[0059] Taking the mean square error of the actual rainstorm disaster loss assessment and the predicted disaster loss as the evaluation loss function, and using the weighted sum of the loss index and the evaluation loss function as the objective function of the rainstorm disaster loss assessment model;

[0060] Initializing the parameters of the rainstorm disaster loss assessment model, and using a long short-term memory network as the predictor for the rainstorm disaster loss assessment model;

[0061] Randomly sampling the data to be analyzed from the meta-training tasks, calculating the inner gradient using the support set, and updating the parameters of the rainstorm disaster loss assessment model. The expression is:

[0062]

[0063] Among them, the q-th meta-training task is T q , the inner learning rate is α, and the initial parameter is The inner updated parameter of the meta-training task T q is The meta-training task T q The loss function of is about the initial parameter The gradient of Initial parameters The heavy rain disaster loss assessment model is

[0064] Use the support set to calculate the meta-loss:

[0065]

[0066] Where the meta-loss is The number of meta-training tasks is Use the parameters on the query set Calculate the meta-training task T q The loss of Update the outer layer parameters, and the expression is:

[0067]

[0068]

[0069] Where the outer layer learning rate is ψ, and the updated outer layer parameters are Parameters The heavy rain disaster loss assessment model is The meta-loss with respect to the parameter The gradient of

[0070] Output the heavy rain disaster loss assessment model after adding meta-learning.

[0071] Second, a heavy rain disaster loss assessment system based on machine learning, including:

[0072] Data acquisition module: used to take the data of the heavy rain disaster data source in the preset area within the specified time period as the data to be analyzed; the data source to be analyzed includes meteorological data, geological data, social and economic data, disaster data, social media data, historical disaster data; the historical disaster data includes historical meteorological data, historical geological data, historical social and economic data, historical disaster data;

[0073] Disaster area division module: used to divide the preset area based on the rainfall according to the data to be analyzed, take the preset area with rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area, calculate the destructive power of the uncertain area according to the hydrological mechanism and drainage system, and add the uncertain area with destructive power greater than the destructive power threshold to the disaster area;

[0074] Analysis and Real-time Dynamic Assessment Module: It is used to conduct comprehensive risk analysis on the data to be analyzed in the disaster area to obtain a comprehensive risk index, calculate the damage degree index of the disaster area, and conduct real-time dynamic assessment on the disaster area based on the damage degree index and the comprehensive risk index to obtain a loss index;

[0075] Modeling and Output Module: It is used to construct a rainstorm disaster loss assessment model based on the loss index, input the data to be evaluated into the rainstorm disaster loss assessment model, and output the evaluation result.

[0076] The beneficial effects of the present invention are:

[0077] The present invention is a rainstorm disaster loss assessment method and system based on machine learning. Compared with the prior art, the present invention has the following technical effects:

[0078] Through steps such as regional division, obtaining the disaster area, calculating the damage degree index, real-time dynamic assessment, and model construction, the present invention can improve the accuracy of the rainstorm disaster loss assessment of machine learning, thereby improving the precision of the rainstorm disaster loss assessment of machine learning. Optimizing the rainstorm disaster loss assessment of machine learning can greatly save resources, improve work efficiency, realize the intelligent assessment of the rainstorm disaster loss of machine learning, conduct regional division and real-time dynamic assessment on the rainstorm disaster loss assessment of machine learning in real time, which is of great significance for the rainstorm disaster loss assessment of machine learning, and can adapt to different standards and requirements of the rainstorm disaster loss assessment of machine learning, having a certain universality. Description of the Drawings

[0079] Figure 1 It is the flowchart of the steps of the rainstorm disaster loss assessment method based on machine learning of the present invention. Detailed Embodiments

[0080] The present invention will be further described below through specific embodiments. The illustrative embodiments and explanations of the present invention are used to explain the present invention, but do not limit the present invention.

[0081] The rainstorm disaster loss assessment method and system based on machine learning of the present invention include the following steps:

[0082] As Figure 1 shown, in this embodiment, it includes the following steps:

[0083] Take the data of the rainstorm disaster data source in the preset area within the specified time period as the data to be analyzed; the data source to be analyzed includes meteorological data, geological data, topographic and geomorphic data, social and economic data, social media data, historical disaster data, ecological data, and building data;

[0084] In the actual evaluation, the living area of a certain city was taken as the research object, and the data to be analyzed for the heavy rain at the end of summer 2023 was collected. The meteorological data included a rainfall of 35 mm in the past 12 hours, a rainfall of 55 mm in the past 24 hours, a wind speed of 15 m / s, and a southeast wind direction; the geological data included a clay soil type, a groundwater level of 2 m, and a stable geological structure; the topographic and geomorphic data included an altitude of 50 m, a slope of 5 degrees, and a plain terrain type; the social and economic data included a population density of 1000 people per square kilometer, an average annual income per capita of 50,000 yuan, a secondary industry proportion of 40%, and a tertiary industry proportion of 60%; the social media data included a high discussion heat regarding heavy rain in the past 12 hours, 5000 likes for related topics, and 2000 reposts; the historical disaster data included 1 small-scale waterlogging event in the past week, 0 flood disasters in the past month, 3 heavy rain disaster occurrences in this area in the past 5 years, and an average economic loss of 79.04 million yuan caused by disasters; the ecological data included a vegetation coverage rate decreased from 57% to 32%, and a wetland area proportion decreased from 21% to 10%; the building data included 500 buildings, 309 residential buildings, 157 commercial buildings, 44 industrial buildings, a building exposure of 0.6, a building sensitivity of 0.53, a building adaptability of 0.409, an exposure weight of 0.41, a sensitivity weight of 0.297, and an adaptability weight of 0.3;

[0085] Based on the rainfall, the preset area is divided according to the data to be analyzed, and the preset area with a rainfall greater than the rainfall threshold is taken as the disaster area, and vice versa as the uncertain area. The destructive power of the uncertain area is calculated according to the hydrological mechanism and drainage system, and the uncertain area with a destructive power greater than the destructive power threshold is added to the disaster area; including:

[0086] Calculate the exposure, sensitivity and adaptability of the building according to the building data, and calculate the destructive power through objective weighting:

[0087]

[0088] BR i =ρ4WE i +ρ5WN i -ρ6WS i

[0089] where the building vulnerability of the i-th uncertain area is BR i , the exposure of the i-th uncertain area is WE i , the sensitivity of the i-th uncertain area is WN i , the adaptability of the i-th uncertain area is WS i , the exposure weight is ρ4, the sensitivity weight is ρ5, the adaptability is ρ6, and the destructive power of the i-th uncertain area is The rainfall in the i-th uncertain area is P i ;

[0090] In the actual assessment, the initial areas are A1, A2, A3, B1, B2, B3, the disaster areas are A1, A2, B2, and the uncertain areas are A3, B1, B3; the uncertain area added to the disaster area is B3;

[0091] The rainfall threshold is 29 mm of rainfall within 12 hours and 49 mm of rainfall within 24 hours; the destructive power threshold is 7.12;

[0092] Perform a comprehensive risk analysis on the data to be analyzed in the disaster area to obtain a comprehensive risk index, calculate the damage degree index of the disaster area, and perform real-time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index to obtain a loss index;

[0093] In the actual assessment, the comprehensive risk indexes of A1, A2, B2, and B3 are 0.91, 0.63, 0.74, and 0.85 respectively; the damage degree indexes of A1, A2, B2, and B3 are 0.74, 0.69, 0.72, and 0.89 respectively; the loss indexes of A1, A2, B2, and B3 are 0.73, 0.67, 0.75, and 0.903;

[0094] Construct a heavy rain disaster loss assessment model based on meta-learning according to the loss index, input the data to be evaluated into the heavy rain disaster loss assessment model, and output the evaluation result;

[0095] In the actual assessment, after adding meta-learning, the evaluation error of the heavy rain disaster loss assessment model is reduced by 29%,

[0096] In this embodiment, the method for dividing the preset area according to the data to be analyzed based on rainfall includes:

[0097] Perform spatio-temporal alignment on the data to be analyzed and the preset area, use the expert method to perform primary division on the preset area to obtain the initial area, calculate the rainfall in the initial area according to the data to be analyzed, and take the initial area with rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area.

[0098] In this embodiment, the method for the disaster area includes:

[0099] The rainfall threshold is divided into an ordinary rainfall threshold and an over-threshold heavy rainfall amount, and the disaster area is divided into an ordinary disaster area and a super heavy rain disaster area;

[0100] The disaster area with rainfall higher than the ordinary rainfall threshold and less than the over-threshold heavy rainfall amount is the ordinary disaster area; the disaster area with rainfall higher than the over-threshold heavy rainfall amount is the super heavy rain disaster area;

[0101] The ordinary rainfall threshold and the amount of rainfall above the threshold are dynamically adjusted according to the rainfall threshold drift due to climate change.

[0102] In this embodiment, the method for comprehensively analyzing the risk of the data to be analyzed in the disaster area includes:

[0103] Calculating the disaster inducing factor index of the ordinary disaster area through meteorological data, calculating the geological vulnerability index of the disaster area through geological data, calculating the terrain stability index through topographic and geomorphic data, and calculating the building vulnerability through building data;

[0104] Calculating the ecological sensitivity, ecological resilience, and ecological stress degree through ecological data, and calculating the ecological vulnerability index based on the ecological sensitivity, ecological resilience, and ecological stress degree:

[0105]

[0106] Among them, the ecological vulnerability index of the u-th ordinary disaster area is The ecological sensitivity is C1, the ecological stress degree is C2, the ecological resilience is C3, the sensitivity weight is ρ1, the stress degree weight is ρ2, and the resilience weight is ρ3;

[0107] Using disaster data and social media data to correct the disaster inducing factor index, geological vulnerability index, and terrain stability index;

[0108] Calculating the comprehensive risk index of the ordinary rainstorm disaster area based on the disaster inducing factor index, geological vulnerability index, terrain stability index, and ecological vulnerability index:

[0109]

[0110] Among them, the comprehensive risk index of the u-th ordinary disaster area is The disaster inducing factor index of the u-th ordinary disaster area is MH u , the geological vulnerability index of the u-th ordinary disaster area is DX u , the terrain stability index of the u-th ordinary disaster area is ZX u , the inducing weight is ω1, the geological weight is ω2, the terrain weight is ω3, the ecological weight is The building weight is ω5, and the building vulnerability of the u-th ordinary disaster area is BR u ;

[0111] Calculating the comprehensive risk index of the super rainstorm disaster area:

[0112]

[0113] Among them, the comprehensive risk index of the u-th super rainstorm disaster area is Mark the risk of the disaster area according to the comprehensive risk index and output the marked disaster area.

[0114] In this embodiment, the method for calculating the damage degree index of the disaster area includes:

[0115] Calculate the loss of life in the disaster area based on the data to be analyzed:

[0116]

[0117] where the loss of life is LM s , the number of disaster areas is N a , the number of at-risk population is N p , the mortality rate of the at-risk population is η, and the influence degree of the b-th main influencing factor on the mortality rate of the at-risk population is γ b , the weight coefficient of the b-th main influencing factor is φ b , the influence degree of the b-th secondary factor on the mortality rate of the at-risk population is ξ b , the weight coefficient value of the b-th secondary factor is The number of secondary factors is N b2 , the number of main factors is N b1 , the weight coefficient is β, and the correction coefficient is

[0118] Calculate the economic loss of the disaster area:

[0119]

[0120] where the economic loss is D s , the number of pixel disaster areas where property is distributed in the disaster area is N1, the number of property types in the v-th pixel disaster area is N2, the number of different flood depth levels is N3, and the loss rate of the v-th type of property in the a-th disaster area corresponding to the x-th flood depth is The property value of the v-th type of property in the a-th disaster area corresponding to the x-th flood depth is

[0121] Calculate the indirect economic loss of the disaster area:

[0122]

[0123] where the indirect coefficient corresponding to the r-th type of property is K a (s), the number of property categories is M1, and the r-th type of property is A r , and the indirect economic loss is

[0124] Calculate the ecological environment loss rate:

[0125]

[0126] Among them, the ecological environment loss rate is E, and the loss rate of the c-th type of ecosystem service function is Q c , the number of ecosystem service functions is N4, the number of ecological products is N5, and the service volume of the i-th product in the c-th type of ecosystem service function is H ck , the unit price of the i-th product in the c-th type of ecosystem service function is R ck ;

[0127] Calculate the damage degree index according to the loss of life, economic loss, indirect economic loss, and ecological environment loss rate:

[0128]

[0129] Among them, the total economic volume is KS, and the damage degree index is

[0130] In this embodiment, the method for obtaining the loss index by performing real-time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index includes:

[0131] Input the damage degree index and the comprehensive risk index into the real-time dynamic assessment model based on the Bayesian network, and discretize the disaster state into three levels: extremely heavy rain, heavy rain, and rainstorm by using the state-time curve;

[0132] Take time as a node and calculate the conditional probability of the transfer network of the node:

[0133]

[0134] Among them, the time interval is Δt, and the offset rate when the node d changes from the w state to the m state is The state of the node at time t0 is w, the state of the node at time t1 is j, and the probability that the state at time t1 is j under the condition that the state at time t0 is w is k(G t1 =j|G t0 =w), the damage degree index at time t0 is The comprehensive risk index at time t0 is The damage degree index when the node d is in the w state is The damage degree index when the node d is in the m state is The comprehensive risk index when the node d is in the w state is The comprehensive risk index when the node d is in the m state is

[0135] Calculate the loss index according to the conditional probability of the transfer network:

[0136]

[0137] The damage degree index at time t in state j is The comprehensive risk index at time t in state j is The loss index at time t is The damage degree index at time t in state w is The comprehensive risk index at time t in state w is

[0138] In this embodiment, a method for constructing a rainstorm disaster loss assessment model based on meta - learning according to the loss index includes:

[0139] Taking the mean square error of the actual rainstorm disaster loss assessment and the predicted disaster loss as the evaluation loss function, and using the weighted sum of the loss index and the evaluation loss function as the objective function of the rainstorm disaster loss assessment model;

[0140] Initializing the parameters of the rainstorm disaster loss assessment model, and using a long - short - term memory network as the predictor for the rainstorm disaster loss assessment model;

[0141] Randomly sampling the data to be analyzed from the meta - training tasks, calculating the inner - layer gradient using the support set, and updating the parameters of the rainstorm disaster loss assessment model. The expression is:

[0142]

[0143] Where the q - th meta - training task is T q , the inner - layer learning rate is α, and the initial parameter is The meta - training task T q The parameter after inner - layer update of is The meta - training task T q The gradient of the loss function of with respect to the initial parameter is The initial parameter The rainstorm disaster loss assessment model of is

[0144] Calculating the meta - loss using the support set:

[0145]

[0146] Where the meta - loss is The number of meta - training tasks is Using the parameter Calculating the loss of the meta - training task T q on the query set is Updating the outer - layer parameters. The expression is:

[0147]

[0148] where the outer learning rate is ψ, and the updated outer parameter is parameter the heavy rain disaster loss assessment model is the loss of yuan with respect to the parameter the gradient is

[0149] Output the heavy rain disaster loss assessment model after adding meta - learning.

[0150] Second, a heavy rain disaster loss assessment system based on machine learning includes:

[0151] Data acquisition module: used to take the data of the heavy rain disaster data source in a preset area within a specified time period as the data to be analyzed; the data sources to be analyzed include meteorological data, geological data, social and economic data, disaster data, social media data, historical disaster data; the historical disaster data includes historical meteorological data, historical geological data, historical social and economic data, historical disaster data;

[0152] Disaster area division module: used to divide the preset area based on the rainfall according to the data to be analyzed, take the preset area with rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area, calculate the destructive power of the uncertain area according to the hydrological mechanism and drainage system, and add the uncertain area with destructive power greater than the destructive power threshold to the disaster area;

[0153] Analysis and real - time dynamic assessment module: used to conduct a comprehensive risk analysis on the data to be analyzed in the disaster area to obtain a comprehensive risk index, calculate the damage degree index of the disaster area, and conduct a real - time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index to obtain a loss index;

[0154] Modeling and output module: used to construct a heavy rain disaster loss assessment model according to the loss index, input the data to be evaluated into the heavy rain disaster loss assessment model, and output the evaluation result.

[0155] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for evaluating the losses caused by rainstorm disasters based on machine learning, characterized in that, It includes the following steps: Taking the data of the heavy rain disaster data source in the preset area within the specified time period as the data to be analyzed; the data source to be analyzed includes meteorological data, geological data, topographic and geomorphic data, social and economic data, social media data, historical disaster data, ecological data, and building data; Based on the rainfall, dividing the preset area according to the data to be analyzed, taking the preset area with a rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area, calculating the destructive power of the uncertain area according to the hydrological mechanism and drainage system, and adding the uncertain area with a destructive power greater than the destructive power threshold to the disaster area; including: Calculating the exposure, sensitivity and adaptability of buildings according to building data, and calculating the destructive power through objective weighting: BR i = ρ4WE i + ρ5WN i - ρ6WS i where the building vulnerability of the i-th uncertain area is BR i , the exposure of the i-th uncertain area is WE i , the sensitivity of the i-th uncertain area is WN i , the adaptability of the i-th uncertain area is WS i , the exposure weight is ρ4, the sensitivity weight is ρ5, the adaptability is ρ6, and the destructive power of the i-th uncertain area is The rainfall of the i-th uncertain area is P i ; Conducting a comprehensive risk analysis on the data to be analyzed in the disaster area to obtain a comprehensive risk index, calculating the damage degree index of the disaster area, and conducting a real-time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index to obtain a loss index; Constructing a heavy rain disaster loss assessment model based on meta-learning according to the loss index, inputting the data to be evaluated into the heavy rain disaster loss assessment model, and outputting the assessment result.

2. The method for evaluating heavy rain disaster losses based on machine learning according to claim 1, wherein The method for dividing the preset area according to the data to be analyzed based on rainfall includes: Aligning the data to be analyzed and the preset area in space and time, using the expert method to conduct a primary division of the preset area to obtain the initial area, calculating the rainfall of the initial area according to the data to be analyzed, taking the initial area with a rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area.

3. The method for evaluating rainstorm disaster losses based on machine learning according to claim 1, wherein The method for the disaster area includes: The rainfall threshold is divided into an ordinary rainfall threshold and an ultra-threshold heavy rain amount, and the disaster area is divided into an ordinary disaster area and a super heavy rain disaster area; The disaster area with a rainfall higher than the ordinary rainfall threshold and lower than the ultra-threshold heavy rain amount is the ordinary disaster area; the disaster area with a rainfall higher than the ultra-threshold heavy rain amount is the super heavy rain disaster area; The ordinary rainfall threshold and the ultra-threshold heavy rain amount are dynamically adjusted according to the rainfall threshold drift caused by climate warming.

4. The method for evaluating the losses caused by rainstorm disasters based on machine learning according to claim 1, characterized in that, The method for conducting a comprehensive risk analysis on the data to be analyzed in the disaster area includes: Calculating the disaster inducing factor index of the ordinary disaster area through meteorological data, calculating the geological vulnerability index of the disaster area through geological data, calculating the terrain stability index through topographic and geomorphic data, and calculating the building vulnerability through building data; Calculating the ecological sensitivity, ecological resilience, and ecological stress degree through ecological data, and calculating the ecological vulnerability index according to the ecological sensitivity, ecological resilience, and ecological stress degree: where the ecological vulnerability index of the \(u\)-th ordinary disaster area is \(\theta\). u , the ecological sensitivity is \(C1\), the ecological stress degree is \(C2\), the ecological resilience is \(C3\), the sensitivity weight is \(\rho1\), the stress degree weight is \(\rho2\), and the resilience weight is \(\rho3\); Using disaster data and social media data to correct the disaster inducing factor index, geological vulnerability index, and terrain stability index; Calculating the comprehensive risk index of the ordinary heavy rain disaster area according to the disaster inducing factor index, geological vulnerability index, terrain stability index, and ecological vulnerability index: where the comprehensive risk index of the $u$-th general disaster area is the disaster inducing factor index of the $u$-th general disaster area is MH u , the geological vulnerability index of the $u$-th general disaster area is DX u , the terrain stability index of the $u$-th general disaster area is ZX u , the inducing weight is $\omega_1$, the geological weight is $\omega_2$, the terrain weight is $\omega_3$, and the ecological weight is $\theta$ u , the building weight is $\omega_5$, and the building vulnerability of the $u$-th general disaster area is BR u ; Calculating the comprehensive risk index of the super heavy rain disaster area: where the comprehensive risk index of the $u$-th super rainstorm disaster area is Perform risk annotation on the disaster area according to the comprehensive risk index, and output the annotated disaster area.

5. The method for evaluating the losses caused by rainstorm disasters based on machine learning according to claim 1, characterized in that, The method for calculating the damage degree index of the disaster area includes: Calculating the life loss in the disaster area according to the data to be analyzed: Among them, the loss of life is LM s , the number of disaster areas is N a , the number of at-risk population is N p , the mortality rate of the at-risk population is η, and the influence degree of the b-th main influencing factor on the mortality rate of the at-risk population is γ b , the weight coefficient of the b-th main influencing factor is φ b , the influence degree of the b-th secondary factor on the mortality rate of the at-risk population is ξ b , the value of the weight coefficient of the b-th secondary factor is , the number of secondary factors is N b2 , the number of main factors is N b1 , the weight coefficient is β, and the correction coefficient is Calculating the economic loss in the disaster area: Among them, the economic loss is D s , the number of pixel disaster areas where property is distributed in the disaster area is N1, the property type of the v-th pixel disaster area is N2, the number of different flood depth levels is N3, and the loss rate of the v-th type of property in the a-th disaster area corresponding to the x-th flood depth is The property value of the v-th type of property in the a-th disaster area corresponding to the x-th flood depth is Calculating the indirect economic loss in the disaster area: Among them, the indirect coefficient corresponding to the r-th type of property is K a (s), the number of property categories is M1, and the r-th type of property is A r , the indirect economic loss is Calculate the ecological environment loss rate: Among them, the ecological environment loss rate is E, and the loss rate of the c-th type of ecosystem service function is Q c , the number of ecosystem service functions is N4, the number of ecological products is N5, and the service volume of the i-th product in the c-th type of ecosystem service function is H ck , the unit price of the i-th product in the c-th type of ecosystem service function is R ck ; Calculate the damage degree index based on the loss of life, economic loss, indirect economic loss, and ecological environment loss rate: Among them, the total economic volume is KS, and the damage degree index is 6. The method for evaluating losses caused by rainstorm disasters based on machine learning according to claim 1, wherein, A method for obtaining a loss index by performing real-time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index, including: Input the damage degree index and the comprehensive risk index into a real-time dynamic assessment model based on a Bayesian network, and discretize the disaster state into three levels: extremely heavy rain, heavy rain, and rainstorm using a state-time curve; Take time as a node and calculate the conditional probability of the transfer network of the node: where the time interval is Δt, and the offset rate when node d changes from state w to state m is χ dwm , the state of the node at time t0 is w, the state of the node at time t1 is j, and the probability that the state at time t1 is j under the condition that the state at time t0 is w is k(G t1 = j|G t0 = w), the damage degree index at time t0 is the comprehensive risk index at time t0 is the damage degree index when node d is in state w is the damage degree index when node d is in state m is the comprehensive risk index when node d is in state w is the comprehensive risk index when node d is in state m is Calculate the loss index according to the conditional probability of the transfer network: The damage degree index in state j at time t is The comprehensive risk index in state j at time t is The loss index at time t is The damage degree index in state w at time t is The comprehensive risk index in state w at time t is 7. The method for evaluating rainstorm disaster losses based on machine learning according to claim 1, wherein A method for constructing a rainstorm disaster loss assessment model based on meta-learning according to the loss index, including: Take the mean square error of the actual rainstorm disaster loss assessment and the predicted disaster loss as the evaluation loss function, and use the weighted sum of the loss index and the evaluation loss function as the objective function of the rainstorm disaster loss assessment model; Initialize the parameters of the rainstorm disaster loss assessment model, and use a long short-term memory network as the predictor for the rainstorm disaster loss assessment model; Randomly sample the data to be analyzed from the meta-training task, calculate the inner gradient using the support set, and update the parameters of the rainstorm disaster loss assessment model. The expression is: where the q-th meta-training task is T q , the inner learning rate is α, and the initial parameters are The inner updated parameters of the meta-training task T q are The loss function of the meta-training task T q with respect to the initial parameters is the gradient The initial parameters of the rainstorm disaster loss assessment model are Calculate the meta-loss using the support set: where the meta-loss is the number of meta-training tasks is Using the parameter on the query set, the calculated meta-training task T q has a loss of Update the outer-layer parameters, the expression is: where the outer learning rate is ψ, and the updated outer parameter is parameter The heavy rain disaster loss assessment model is The loss of yuan with respect to the parameter The gradient is Output the rainstorm disaster loss assessment model after adding meta-learning.

8. A rainstorm disaster loss assessment system based on machine learning for performing the method according to any one of claims 1-7, characterized in that, Including: Data acquisition module: used to take the data of the rainstorm disaster data source in a preset area within a specified time period as the data to be analyzed; the data source to be analyzed includes meteorological data, geological data, social and economic data, disaster data, social media data, historical disaster data; the historical disaster data includes historical meteorological data, historical geological data, historical social and economic data, historical disaster data; Disaster area division module: used to divide the preset area based on the rainfall according to the data to be analyzed, take the preset area with rainfall greater than the rainfall threshold as the disaster area, and vice versa as the uncertain area, calculate the destructive power of the uncertain area according to the hydrological mechanism and drainage system, and add the uncertain area with destructive power greater than the destructive power threshold to the disaster area; Analysis real-time dynamic assessment module: used to perform comprehensive risk analysis on the data to be analyzed in the disaster area to obtain a comprehensive risk index, calculate the damage degree index of the disaster area, and perform real-time dynamic assessment on the disaster area according to the damage degree index and the comprehensive risk index to obtain a loss index; Modeling output module: used to construct a rainstorm disaster loss assessment model according to the loss index, input the data to be evaluated into the rainstorm disaster loss assessment model, and output the evaluation result.

Images