A storm flood disaster chain risk inference method and system based on weather typing and bayesian network

Abstract

The application discloses a storm flood disaster chain risk inference method and system based on weather typing and a Bayesian network, and the method comprises the following steps: collecting atmospheric reanalysis, rainfall station, hydrological station, historical disaster record and design flow data of return period; after processing sea level pressure field and 500hPa potential height field, adopting empirical orthogonal function dimension reduction and K-means clustering to realize daily objective weather typing; automatically extracting a precipitation event and a flood event and matching them with disaster event records according to a time window, forming an event sample table through disaster grading and disaster factor discretization; taking a weather type as a root node to construct a Bayesian network; and performing forward inference of the weather type and reverse inference of a disaster scenario. The application can improve the objectivity, interpretability and robustness under a small sample condition of storm flood disaster chain risk inference.

Description

Technical Field

[0001] This invention relates to the field of natural disaster risk assessment and meteorological-hydrological interdisciplinary technology, and in particular to a method and system for inferring the risk of rainstorm and flood disaster chains based on weather classification and Bayesian networks. Background Technology

[0002] Rainstorms and floods are among the most frequent and devastating types of natural disasters in my country. Heavy rainfall events often do not trigger a single disaster, but rather amplify and propagate through a chain reaction of "rainstorms—floods—landslides / mudslides / waterlogging—loss of life and property," forming a rainstorm and flood disaster chain. Quantitatively inferring the probability of occurrence and the conditional dependencies of each link in this disaster chain is a crucial foundation for disaster risk assessment, early warning decision-making, and the allocation of disaster prevention and mitigation resources.

[0003] Traditional methods for regional rainstorm and flood disaster risk assessment primarily rely on static regional indicators for risk mapping, failing to characterize the conditional transmission relationships of disaster chains at the event scale. Furthermore, event-scale rainstorm and flood disaster risk assessment methods based on Bayesian networks are often geared towards single engineering projects, neglecting atmospheric circulation background and multi-hazard chain structures. Overall, existing rainstorm and flood disaster risk assessment technologies face the following main problems: First, the characterization of atmospheric circulation background is lacking or subjective. Existing methods often directly use rainfall indicators as input for risk assessment, failing to incorporate the atmospheric circulation patterns behind rainfall events into the modeling. A few methods that consider weather patterns rely on forecasters manually identifying circulation patterns, which is highly subjective, non-repeatable, and difficult to apply in batches to long-term historical data. Second, the modeling of each link in the disaster chain is fragmented. Existing technologies typically assess floods, landslides, and urban flooding independently, failing to characterize the chain-like conditional dependencies of "circulation background—hazard-causing factors—multiple hazards—multiple damages" within a unified probabilistic framework. This makes it impossible to answer forward and backward inference questions such as "Given a certain type of weather, what is the probability that each link in the disaster chain will reach a certain level?" and "When a disaster of a certain level occurs, what is the most likely combination of weather background and hazard-causing factors?" Third, the automation level of event sample construction is low. Constructing a disaster chain probabilistic model requires a large number of interconnected event samples of "rainfall events—flood events—disaster records," and existing methods largely rely on manual compilation. The rules for flood event identification and disaster matching are inconsistent, resulting in low efficiency and difficulty in reproduction. Fourth, the standards for disaster level classification are inconsistent. The units for indicators such as the number of affected people, collapsed houses, and economic losses in disaster records are chaotic and severely missing. The flood magnitude lacks objective classification basis that is consistent with engineering design standards, making it difficult to define the state of model nodes. Fifth, probability estimation is not robust under small sample conditions. High-level events in the disaster chain are rare. The number of parameter combinations in the multi-parent node conditional probability table far exceeds the sample size. Directly using frequency estimation will produce a large number of zero probabilities and ill-conditioned conditions. Moreover, the maximum likelihood fitting of commonly used software is prone to errors or missing values ​​under sparse samples. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method and system for inferring the risk of rainstorm and flood disaster chains based on weather classification and Bayesian networks, which solves the problem of characterizing the conditional transmission relationship of disaster chains based on the event scale of Bayesian networks in the context of atmospheric circulation and multi-hazard chain structure.

[0005] Technical Solution: The method for inferring the risk of rainstorm and flood disaster chains based on weather classification and Bayesian networks described in this invention includes the following steps:

[0006] Step 1: Collect atmospheric reanalysis data, daily rainfall data from rain gauges, flow data from hydrological stations, historical disaster records, and design flow data for the return period of sub-basins in the study area; the atmospheric reanalysis data includes at least the daily sea-level pressure field and the 500 hPa geopotential height field; the historical disaster records include disaster records at the administrative region scale and the sub-basin scale.

[0007] Step 2: Perform daily climatological anomaly, latitudinal cosine area weighting, and grid-by-grid standardization on the sea level pressure field and the 500hPa geopotential height field. After splicing the sea level pressure field and the 500hPa geopotential height field along the grid dimension, perform dimensionality reduction using empirical orthogonal functions. Retain the principal component time series whose cumulative variance contribution rate reaches the preset threshold. Perform K-means clustering on the principal component time series and determine the optimal number of subtypes to obtain the daily weather type label sequence.

[0008] Step 3: At the administrative scale, calculate the average rainfall from the daily rainfall data of rain gauges, segment precipitation events based on wet day thresholds and dry period thresholds, and calculate the duration, total rainfall, and maximum daily rainfall of each precipitation event. At the sub-basin scale, perform baseflow segmentation on the hydrological station flow data, and automatically identify flood events based on the flood peak threshold, flood initiation slope threshold, and flood recession slope threshold determined adaptively according to the flow quantile, to obtain the start and end times of the flood events and the peak flow.

[0009] Step 4: Based on the start and end times of the flood event, improve the disaster event record; for the improved disaster event record, match the precipitation event with the nearest start date within a preset number of days; the disaster event record includes flood disasters, landslide disasters, debris flow disasters, and urban flooding disasters;

[0010] Step 5: Classify the disaster event records matched with precipitation events into different levels and discretize the disaster-causing factors to form a fully discrete event sample table;

[0011] Step 6: Obtain the Bayesian network for probability inference based on the daily weather label sequence and event sample table;

[0012] Step 7: Perform forward inference based on a fixed weather pattern using a Bayesian network to obtain the posterior probability distribution of the disaster-causing factors, disaster level, and disaster damage level under a given weather pattern; perform backward inference based on a fixed disaster scenario to obtain the posterior probability distribution of the weather pattern, disaster-causing factors, and disaster damage level under a given disaster level.

[0013] Furthermore, in step 2, the daily climatological anomaly, latitudinal cosine area weighting, and grid-point standardization processing of the sea level pressure field and the 500 hPa geopotential height field includes:

[0014] For each grid point, subtracting its multi-year average over the same time period yields the anomaly field. :

[0015]

[0016] In the formula: For grid points At any moment The original field value, For a moment The corresponding month-day sequence number, For grid points On calendar days The multi-year average;

[0017] Weighted by taking the square root of the area of ​​the cosine of latitude for the anomaly field:

[0018]

[0019] In the formula: For grid points At any moment The anomaly field value after latitude and longitude area weighting, For grid points Latitude;

[0020] Standardization is achieved by dividing each grid point of the open-weighted anomaly field by the standard deviation of the anomaly field.

[0021] Further, in step 2, the sea level pressure field and the 500 hPa geopotential height field are concatenated along the grid dimension and then subjected to dimensionality reduction using an empirical orthogonal function. The principal component time series whose cumulative variance contribution rate reaches a preset threshold are retained. This includes: concatenating the sea level pressure field and the 500 hPa geopotential height field along the grid dimension into a standardized wide matrix. For standardized wide matrices Calculate the time covariance matrix , The dimension is time step × time step; extract Eigenvalues ​​arranged in descending order of eigenvalues and corresponding feature vectors The principal component time series is taken until the cumulative variance contribution rate first reaches the preset threshold. The cumulative variance contribution rate is calculated by summing the extracted eigenvalues ​​and... Calculation of the ratio of traces.

[0022] Furthermore, in step 2, performing K-means clustering on the principal component time series and determining the optimal number of clusters includes:

[0023] Within the preset genotype number scanning range, K-means clustering based on Euclidean distance is performed on each principal component time series, and the genotype number is recorded. The corresponding sum of squared errors within groups The range normalization of the within-group sum of squared errors sequence is performed using the following formula:

[0024]

[0025] And calculate the second-order difference:

[0026]

[0027] In the formula: , These are the minimum and maximum values ​​of the sum of squared errors within the group within the scanning range, respectively. The optimal number of morphologies is the number of morphologies corresponding to the point with the maximum absolute value of the second difference at the inner points of the scanning range. Then, a formal K-means clustering with multiple random starting points is performed using the optimal number of morphologies. The clustering result with the minimum sum of squared errors within the group is taken, and each day is assigned to the nearest cluster center to obtain the daily weather label sequence.

[0028] Further, in step 3, the average rainfall is calculated from the daily rainfall data of the rain gauge stations. Precipitation events are segmented based on wet day thresholds and dry period thresholds, and the duration, total rainfall, and maximum daily rainfall for each precipitation event are calculated, including:

[0029] The arithmetic mean of the daily rainfall at each rain gauge station within the administrative region is taken as the area average rainfall; dates with area average rainfall greater than or equal to the wet day threshold are recorded as wet days; when the interval between adjacent wet days exceeds the dry period threshold, they are divided into different precipitation events.

[0030] For each precipitation event, the duration is the number of days between the start and end dates of the precipitation event, the total rainfall is the sum of the average rainfall within the precipitation event, and the maximum daily rainfall is the maximum value of the average rainfall within the precipitation event.

[0031] Furthermore, in step 3, the flood peak threshold, the flood initiation slope threshold, and the flood receding slope threshold, adaptively determined according to the flow quantile, include:

[0032] The formula for the flood peak threshold is as follows:

[0033]

[0034] In the formula: For the peak flood threshold, , These are the 90th and 99th percentiles of the flow sequence for this sub-basin. To preset the minimum threshold lower limit, , For preset coefficients, Take a value of 0.5 to 0.7. Take values ​​of 0.15 to 0.25, with flow rates measured in cubic meters per second;

[0035] The formula for the slope threshold is as follows:

[0036]

[0037] In the formula: The threshold for the initial slope. The slope benchmark value is selected based on the preset area interval to which the sub-basin area belongs. This represents the maximum flow rate in the flow sequence of this sub-basin. This serves as a reference value for normalized flow rate. and All values ​​are in cubic meters per second; the receding slope threshold is the rising slope threshold multiplied by a preset reduction factor that is greater than 0 and less than 1.

[0038] Furthermore, step 5 includes:

[0039] Flood disasters are classified into levels based on a comparison of peak flow and design flow data with return period. Landslide disasters, debris flow disasters, and urban flooding disasters are classified into levels based on disaster loss indicators. These disaster loss indicators include the number of affected people, collapsed houses, affected crop area, and direct economic losses. The disaster loss indicators are divided into four levels: none, mild, moderate, and severe. The duration, total rainfall, maximum daily rainfall, and previous rainfall of precipitation events are discretized into a preset number of levels according to preset classification thresholds. Weather tags are attached to each disaster event record from the daily weather tag sequence based on the representative date of the disaster event record, forming a fully discrete event sample table.

[0040] Further, step 6 includes: using the daily weather model label sequence as the root node, duration, total rainfall, maximum daily rainfall and previous rainfall as the second-level nodes, disaster types such as floods, landslides, debris flows and waterlogging as the third-level nodes, and disaster loss indicators as the fourth-level nodes, to construct a Bayesian network with a directed acyclic graph structure.

[0041] Based on the event sample table, the conditional probability table is calculated node-by-node using an equivalent sample size based on Dirichlet prior, as shown in the following formula:

[0042]

[0043] In the formula: For the current node, for Parent node combination, The combined value of the parent node is And the current node's value is The number of samples, The combined value of the parent node is The total number of samples, The number of states of the current node. The equivalent sample size smoothing coefficient is greater than 0.

[0044] Furthermore, in step 7, forward inference and backward inference are achieved through joint tree inference: the directed acyclic graph of the Bayesian network is moralized and triangulated to construct a clique tree, and evidence propagation is completed through message passing based on the set evidence node states, and the posterior probability distribution of the target node is queried.

[0045] The rainstorm and flood disaster chain risk inference system based on weather classification and Bayesian networks described in this invention includes the following modules:

[0046] The data acquisition module collects atmospheric reanalysis data, daily rainfall data from rain gauges, flow data from hydrological stations, historical disaster records, and design flow data for the return period of sub-basins in the study area. The atmospheric reanalysis data includes at least the daily sea-level pressure field and the 500 hPa geopotential height field. The historical disaster records include disaster records at the administrative region scale and the sub-basin scale.

[0047] The weather classification module is used to perform daily climatological anomalies, latitudinal cosine area weighting, and grid-by-grid standardization on the sea level pressure field and the 500hPa geopotential height field. After splicing the sea level pressure field and the 500hPa geopotential height field along the grid dimension, the module performs dimensionality reduction using empirical orthogonal functions, retains the principal component time series with a cumulative variance contribution rate reaching a preset threshold, performs K-means clustering on the principal component time series, determines the optimal number of classifications, and obtains the daily weather type label sequence.

[0048] The event extraction module is used to calculate the average rainfall from daily rainfall data of rain gauges at the administrative scale, segment precipitation events based on wet day thresholds and dry period thresholds, and calculate the duration, total rainfall, and maximum daily rainfall of precipitation events. At the sub-basin scale, the module performs baseflow segmentation on hydrological station flow data, automatically identifies flood events based on flood peak thresholds, flood initiation slope thresholds, and flood receding slope thresholds determined adaptively according to flow quantiles, and obtains the start and end times of flood events and peak flow rates.

[0049] The event matching module is used to improve the disaster event record based on the start and end times of the flood event; for the improved disaster event record, it matches the precipitation event with the nearest start date within a preset number of days; the disaster event record includes flood disasters, landslide disasters, debris flow disasters and waterlogging disasters;

[0050] The tagging module is used to classify disaster event records matched with precipitation events into different levels and discretize the disaster-causing factors to form a fully discrete event sample table.

[0051] The network construction module is used to obtain a Bayesian network for probabilistic inference based on the daily weather-type label sequence and event sample table;

[0052] The risk inference module is used to perform forward inference based on a fixed weather pattern using a Bayesian network to obtain the posterior probability distribution of disaster-causing factors, disaster level, and disaster damage level under a given weather pattern; and to perform backward inference based on a fixed disaster scenario to obtain the posterior probability distribution of weather pattern, disaster-causing factors, and disaster damage level under a given disaster level.

[0053] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. It replaces manual circulation classification with data-driven objective weather classification, and the classification process is repeatable and can be applied in batches to historical sequences spanning decades; 2. It uses the time modality transpose method to achieve dimensionality reduction of empirical orthogonal functions, transforming the eigenvalue decomposition of the spatial covariance matrix into the eigenvalue decomposition of the temporal covariance matrix, allowing ordinary servers to complete the classification calculation of large-scale high-resolution reanalysis data; 3. By constructing a Bayesian network with a directed acyclic graph structure, it supports both positive risk prediction for a given weather pattern and reverse diagnosis for a given disaster scenario, with inference results possessing clear physical and probabilistic semantics, directly serving the specific technical application of flood control scheduling; 4. The entire process of precipitation event segmentation, flood event identification, and disaster matching is regularized and automated, with flood event identification parameters adaptively determined by flow quantiles and sub-basin areas, avoiding the inefficiency and inconsistency of manual compilation; 5. Flood levels are aligned with engineering hydrological design standards, objectively classified according to the design flow of the return period, and other disasters are classified according to a comprehensive multi-index disaster loss score, with unified and standardized node status definitions. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0055] Figure 2 This is a schematic diagram of the Bayesian network structure of the present invention;

[0056] Figure 3 The following is a weather classification result diagram in an embodiment of the present invention, wherein (a) is a result diagram of the elbow rule curve of weather classification, and (b) is a result diagram of the classification frequency of weather classification;

[0057] Figure 4 This is a schematic diagram of the posterior probability distribution of each node in the disaster chain under a given weather condition in an embodiment of the present invention. Detailed Implementation

[0058] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0059] Taking the Haihe River Basin and its Daqing River system as the study area, this project is implemented based on the R language platform. In the embodiments, the English identifiers such as weather_type, duration, total_prcp, max_prcp, Ant3, flood, landslide, defris_flow, and waterlogging are only program variable names.

[0060] Example 1

[0061] like Figure 1 As shown, the method for inferring the risk of rainstorm and flood disaster chains based on weather classification and Bayesian networks described in this invention includes:

[0062] Step S1: Collect atmospheric reanalysis data, daily rainfall data from rain gauges, flow data from hydrological stations, historical disaster records, and design flow data for the return period of sub-basins in the study area; the atmospheric reanalysis data includes at least the daily sea level pressure field and the 500 hPa geopotential height field; the historical disaster records include disaster records at the administrative region scale and disaster records at the sub-basin scale, and each disaster record includes at least the disaster type and the time of occurrence;

[0063] Step S2: Perform daily climatological anomaly, latitudinal cosine area weighting, and grid-by-grid standardization on the sea level pressure field and the 500hPa geopotential height field. After splicing the sea level pressure field and the 500hPa geopotential height field along the grid dimension, perform dimensionality reduction using empirical orthogonal functions. Retain the principal component time series whose cumulative variance contribution rate reaches the preset threshold. Perform K-means clustering on the principal component time series and determine the optimal number of subtypes to obtain the daily weather type label sequence.

[0064] The daily climatological anomalies, latitudinal cosine area weighting, and grid-point standardization of the sea level pressure field and the 500hPa geopotential height field include:

[0065] For each grid point, subtracting the multi-year average value of that grid point over the same "month-day" period yields the anomaly field A(g,t):

[0066]

[0067] In the formula: For grid points At any moment The original field value, For a moment The corresponding month-day sequence number, For grid points On calendar days The multi-year average;

[0068] Weighted by taking the square root of the area of ​​the cosine of latitude for the anomaly field:

[0069]

[0070] In the formula: For grid points At any moment The anomaly field value after latitude and longitude area weighting, For grid points Latitude;

[0071] Standardize the open-weighted anomaly field by dividing each grid point by the standard deviation of the anomaly field. After eliminating the dimensional differences between different variables, the two variable fields are spliced ​​together along the grid point dimension to form a standardized wide matrix S.

[0072] After concatenating the sea level pressure field and the 500 hPa geopotential height field along the grid dimension, dimensionality reduction was performed using an empirical orthogonal function. The principal component time series whose cumulative variance contribution rate reached a preset threshold were retained, including:

[0073] The sea level pressure field and the 500 hPa geopotential height field are stitched together along the grid dimension to form a normalized wide matrix. For standardized wide matrices Calculate the time covariance matrix , The dimension is time step × time step; extract Eigenvalues ​​arranged in descending order of eigenvalues and corresponding feature vectors The principal component time series is taken until the cumulative variance contribution rate first reaches the preset threshold. The cumulative variance contribution rate is calculated by summing the extracted eigenvalues ​​and... Calculation of the ratio of traces.

[0074] Performing K-means clustering on principal component time series and determining the optimal number of clusters includes:

[0075] Within the preset number of genotypes, K-means clustering based on Euclidean distance is performed on each principal component time series, and the sum of squared within-group errors corresponding to each genotype number K is recorded. The range normalization of the within-group sum of squared errors sequence is performed using the following formula:

[0076]

[0077] And calculate the second-order difference:

[0078]

[0079] In the formula: , These are the minimum and maximum values ​​of the sum of squared errors within the group within the scanning range, respectively. The optimal number of morphologies is the number of morphologies corresponding to the point with the maximum absolute value of the second difference at the inner points of the scanning range. Then, a formal K-means clustering with multiple random starting points is performed using the optimal number of morphologies. The clustering result with the minimum sum of squared errors within the group is taken, and each day is assigned to the nearest cluster center to obtain the daily weather label sequence.

[0080] Step S3: At the administrative scale, calculate the average rainfall from the daily rainfall data of rain gauges, segment precipitation events based on wet day thresholds and dry period thresholds, and calculate the duration, total rainfall, and maximum daily rainfall of each precipitation event; at the sub-basin scale, for disaster records, perform baseflow segmentation on the hydrological station flow data, and automatically identify flood events based on the flood peak threshold, flood initiation slope threshold, and flood receding slope threshold determined adaptively according to the flow quantile, and obtain the start and end times of the flood events and the peak flow.

[0081] The areal average rainfall is calculated from daily rainfall data from rain gauges. Precipitation events are segmented based on wet day and dry period thresholds. The duration, total rainfall, and maximum daily rainfall for each precipitation event are calculated, including:

[0082] The arithmetic mean of the daily rainfall at each rain gauge station within the administrative region is taken as the area average rainfall; dates with area average rainfall greater than or equal to the wet day threshold are recorded as wet days; when the interval between adjacent wet days exceeds the dry period threshold, they are divided into different precipitation events.

[0083] For each precipitation event, the duration is the number of days between the start and end dates of the precipitation event, the total rainfall is the sum of the average rainfall within the precipitation event, and the maximum daily rainfall is the maximum value of the average rainfall within the precipitation event.

[0084] The flood peak threshold, flood initiation slope threshold, and flood recession slope threshold, adaptively determined based on flow quantiles, include:

[0085] The formula for the flood peak threshold is as follows:

[0086]

[0087] In the formula: For the peak flood threshold, , These are the 90th and 99th percentiles of the flow sequence for this sub-basin. To preset the minimum threshold lower limit, , For preset coefficients, Take a value of 0.5 to 0.7. Take values ​​of 0.15 to 0.25, with flow rates measured in cubic meters per second;

[0088] The formula for the slope threshold is as follows:

[0089]

[0090] In the formula: The threshold for the initial slope. The slope benchmark value is selected based on the preset area interval to which the sub-basin area belongs. This represents the maximum flow rate in the flow sequence of this sub-basin. This serves as a reference value for normalized flow rate. and All values ​​are in cubic meters per second; the receding slope threshold is the rising slope threshold multiplied by a preset reduction factor that is greater than 0 and less than 1.

[0091] Step S4: Based on the start and end times of the flood event, refine the disaster event record; for the refined disaster event record, match the precipitation event with the nearest start date within a preset number of days, including:

[0092] Step S4.1 Based on the sub-basin-county association table, match the flood events that occurred within the sub-basin obtained in step S3 with the disaster events at the administrative region level: if the start and end times of the flood include the occurrence time of the disaster event, then add the flood disaster to the corresponding disaster event; if there are no disaster events within the start and end times of the flood, then add a new flood disaster event record.

[0093] Step S4.2 matches precipitation events to the improved disaster events: For disaster events that include flood disasters, if the flood rise date falls within the start and end time range of a certain precipitation event, the precipitation event whose start date is closest to the flood start time is selected; for disaster events that do not include flood disasters, if the disaster occurs within the start and end time range of a precipitation event, then the precipitation event is matched.

[0094] The disaster event records include flood disasters, landslide disasters, debris flow disasters, and urban flooding disasters;

[0095] Step S5: Classify the disaster event records matched with precipitation events into different levels and discretize the disaster-causing factors to form a fully discrete event sample table, specifically including:

[0096] Flood disasters are classified into levels based on the comparison between peak flow and design flow for the return period: peak flow less than the 20-year return period design flow is classified as mild; peak flow greater than or equal to the 20-year return period but less than the 50-year return period design flow is classified as moderate; peak flow greater than or equal to the 50-year return period design flow is classified as severe; and no flood occurs is classified as none.

[0097] Landslide disasters, debris flow disasters, and urban flooding disasters are classified into levels based on disaster damage indicators: the number of people affected, collapsed houses, affected crop area, and direct economic losses are mapped to scores according to a preset index range-score comparison table. The average of each score is taken as the comprehensive score. If the comprehensive score is less than or equal to the first score threshold, it is recorded as mild; if it is greater than the first score threshold but less than or equal to the second score threshold, it is recorded as moderate; if it is greater than the second score threshold, it is recorded as severe; if no corresponding disaster occurs, it is recorded as none.

[0098] The disaster loss indicators include the number of people affected, collapsed houses, affected area of ​​crops, and direct economic losses. The disaster loss indicators are divided into four levels: none, minor, moderate, and severe.

[0099] The duration, total rainfall, maximum daily rainfall, and previous rainfall of precipitation events are discretized into a preset number of levels according to preset classification thresholds: the maximum daily rainfall and total rainfall are discretized into four levels: below heavy rain, heavy rain, torrential rain, and extremely heavy rain; the duration is discretized into four levels: short duration, medium duration, long duration, and relatively long duration; and the previous rainfall is discretized into three levels: less, moderate, and more.

[0100] Based on the representative date of each disaster event record, a weather-type tag is attached to each disaster event record from the daily weather-type tag sequence, forming a fully discrete event sample table.

[0101] Step S6: Obtain the Bayesian network for probability inference based on the daily weather-type label sequence and event sample table;

[0102] A Bayesian network with a directed acyclic graph structure is constructed: the first layer consists of daily weather pattern label sequences, serving as the root node; the second layer consists of disaster-causing factor nodes, including duration, total rainfall, maximum daily rainfall, and previous rainfall; the parent node of each disaster-causing factor node is the daily weather pattern label sequence node; the third layer consists of disaster nodes, including disaster types such as floods, landslides, debris flows, and urban flooding; the parent node of each disaster node is all disaster-causing factor nodes; the fourth layer consists of disaster-damaged nodes, including the number of affected people, collapsed houses, affected crop area, and direct economic losses; the parent node of each disaster-damaged node is all disaster nodes.

[0103] Based on the event sample table, the conditional probability table is calculated node-by-node using an equivalent sample size based on Dirichlet prior, as shown in the following formula:

[0104]

[0105] In the formula: For the current node, for Parent node combination, The combined value of the parent node is And the current node's value is The number of samples, The combined value of the parent node is The total number of samples, The number of states of the current node. The equivalent sample size smoothing coefficient is greater than 0 when... equal When this is the case, the formula is the classic Laplace smoothing.

[0106] In addition to constructing a Bayesian network, an independent chain Bayesian network is also constructed for each type of disaster: the structure of the chain Bayesian network is that the daily weather model label sequence node points to each disaster-causing factor node, each disaster-causing factor node points to the disaster node, and the disaster node points to each disaster-damaged node; when constructing a chain Bayesian network for a certain disaster, only the event samples of the actual occurrence of the disaster are used to calculate the conditional probability table, and the stateless state is removed from the state space of the disaster node.

[0107] By using Bayesian networks for probabilistic inference, the problems of zero probability and ill-conditioned estimation under sparse samples can be solved.

[0108] Bayesian networks are used for panoramic risk queries across disasters, and the chained Bayesian networks described above are used for fine-grained inference of the level corresponding to the occurrence conditions of disasters.

[0109] Step S7: Perform forward inference based on a fixed weather pattern using a Bayesian network to obtain the posterior probability distribution of the disaster-causing factor, disaster level, and disaster damage level under a given weather pattern; perform backward inference based on a fixed disaster scenario to obtain the posterior probability distribution of the weather pattern, disaster-causing factor, and disaster damage level under a given disaster level.

[0110] Forward and backward inference are achieved through joint tree inference: the directed acyclic graph of the Bayesian network is moralized and triangulated to construct a clique tree, and evidence propagation is completed through message passing based on the set evidence node states, and the posterior probability distribution of the target node is queried.

[0111] The rainstorm and flood disaster chain risk inference system based on weather classification and Bayesian networks described in this invention includes the following modules:

[0112] The data acquisition module is used to collect atmospheric reanalysis data, daily rainfall data from rain gauges, flow data from hydrological stations, historical disaster records, and return period design flow data for each sub-basin in the study area.

[0113] The weather classification module is used to perform daily climatological anomalies, latitudinal cosine area weighting, and grid-by-grid standardization on the sea level pressure field and the 500hPa geopotential height field. After dimensionality reduction by empirical orthogonal functions, the principal component time series with the cumulative variance contribution rate reaching the preset threshold are retained. K-means clustering is performed on the principal component time series to determine the optimal number of classifications, and the daily weather type label sequence is obtained.

[0114] The event extraction module is used to segment independent precipitation events at the administrative scale based on wet day threshold and dry period threshold and calculate the duration, total rainfall and daily maximum rainfall characteristics. At the sub-basin scale, the flow process line is divided into baseflow segments and flood events are automatically identified based on adaptively determined peak flood threshold, flood initiation slope threshold and receding slope threshold.

[0115] The event matching module is used to match historical disaster records with precipitation events and flood events according to time window rules, determine the representative dates of each event, and calculate the characteristics of previous rainfall.

[0116] The tagging module is used to classify the disasters obtained from the matching, discretize and classify the disaster-causing factors, and attach weather-type tags to each event according to the representative date to form a fully discrete event sample table.

[0117] The network construction module is used to construct a four-layer Bayesian network with weather type as the root node, consisting of "weather type - disaster-causing factor - disaster - disaster loss". It uses an equivalent sample size based on Dirichlet prior to smooth the conditional probability table for each node.

[0118] The risk inference module is used to perform forward inference based on weather conditions and backward inference based on disaster scenarios.

[0119] The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for inferring the risk of rainstorm and flood disaster chains based on weather classification and Bayesian networks.

[0120] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned method for inferring the risk chain of rainstorm and flood disasters based on weather classification and Bayesian networks.

[0121] Example 2

[0122] Step 1: Data preparation. Collect the following five types of data:

[0123] (1) Atmospheric reanalysis data: ERA5 daily reanalysis data, covering the period from 1955 to 2023, with a spatial range of 60° to 140° east longitude and 0° to 60° north latitude, and a spatial resolution of 0.25° × 0.25°. Variables include sea level pressure and 500 hPa geopotential height, totaling 321 × 241 = 77,361 grid points and 25,202 days. Geometric potential units are converted to geopotential decimeters, and air pressure units are converted to hectopascals.

[0124] (2) Daily rainfall data from rain gauge stations: The daily rainfall sequence of corresponding rain gauge stations in each district and county within the study area was obtained, and a district-county-rain gauge mapping table was established.

[0125] (3) Hydrological station flow data: daily flow process sequence of hydrological stations at the outlet of each sub-basin and sub-basin area; at the same time, based on the latitude and longitude coordinates of the hydrological stations at the outlet of each sub-basin and the vector data of the administrative boundaries of districts and counties, spatial overlay analysis is performed to determine the districts and counties where the hydrological stations are located, and a sub-basin-district association table is established for the association of disaster records at the administrative scale and event samples at the sub-basin scale.

[0126] (4) Historical disaster record data: Historical disaster ledgers of each district and county, with fields including disaster type, time of occurrence, number of people affected, number of houses collapsed, area of ​​crops affected, and direct economic loss; the disaster type is standardized and mapped, and the quantity field is normalized; based on the sub-basin-district association table, the disaster records are assigned to the corresponding sub-basin according to the location of occurrence, forming sub-basin scale disaster sub-events.

[0127] (5) Design flow data for the return period: The 20-year return period design flow Q20 and the 50-year return period design flow Q50 of the outlet section of each sub-basin are taken from the hydrological design results or flood control planning data of the basin; for sub-basins without design results, the annual maximum flood peak series fitting curve can be used to deduce the data.

[0128] Step 2, Weather Classification;

[0129] Step 2.1, Daily Climatic Anomaly: For each grid point, subtract the multi-year average from 1955 to 2023 for that grid point on the same "month-day", removing seasonal cycles:

[0130]

[0131] In the formula: For grid points At any moment The original field value, For a moment The corresponding month-day sequence number, For grid points On calendar days The multi-year average.

[0132] Step 2.2, Latitude Cosine Area Weighting: Multiply the anomaly field by the square root of the latitude cosine of the grid points, i.e. This reduces the contribution weight of higher-dimensional grid points accordingly.

[0133] Step 2.3, Standardization and Stitching: Divide the weighted anomaly field by its standard deviation point by point to eliminate the dimensional difference between sea level pressure and geopotential height; stitch the two variable fields along the grid dimension into a standardized wide matrix S with 154722 rows × 25202 columns.

[0134] Step 2.4, Dimensionality Reduction using Empirical Orthogonal Functions: Calculating the Time Covariance Matrix using the time modality transpose method. The dimension is 25202×25202 (this embodiment uses a cross-product convention without dividing by the number of samples, which does not affect the direction of the feature vector and the variance proportion of each principal component); a large-scale matrix partial eigenvalue decomposition iterative algorithm (such as the Lanczos method or the implicit restart Arnoldi method) is used to extract... The top 50 eigenvalues, sorted from largest to smallest. and eigenvectors The pre-extraction number of 50 is a fixed value greater than the expected number of principal components to be retained; the principal component time series is taken as follows: The cumulative variance contribution rate is calculated by summing the extracted eigenvalues ​​and... The ratio of traces is calculated. In this embodiment, the cumulative variance contribution rate threshold is preset to 85%, and the first 23 principal components are retained, with a cumulative variance contribution rate of 85.5%. If eigenvalue decomposition is performed directly on the spatial dimension, a covariance matrix of 154722×154722 needs to be constructed. Storing this matrix alone would require approximately 191GB of memory, which cannot be completed by a regular server. After using time modality transpose, the matrix storage size is approximately 5.1GB, which can be completed by a regular server.

[0135] Step 2.5: Determine the optimal number of fractions using the elbow rule: For the 23-dimensional principal component time series, perform K-means clustering based on Euclidean distance (Hartigan-Wong algorithm, maximum number of iterations 100, each K is selected from 20 random starting points with the smallest sum of squared errors within the group) one by one within the range of fraction number K=2 to 12, and record the sum of squared errors within the group WSS(K) corresponding to each K; perform range normalization and calculate the second difference according to the following formula:

[0136]

[0137]

[0138] The optimal number of fractals is determined by taking the fractal number corresponding to the maximum absolute value of the second-order difference at the inlier points K=3 to 11 within the scanning range. In this embodiment, the optimal number of fractals is determined to be K=4. Figure 3 As shown in (a).

[0139] Step 2.6, Formal Clustering: Perform formal K-means clustering with K=4 and 100 random starting points (taking the clustering result with the smallest sum of squared errors within each group, with a maximum of 500 iterations), classifying the 25202 days into four weather types: Type 1 Upper-level trough (27.53%), Type 2 Low-level vortex (22.31%), Type 3 Northeast cold vortex + Southwest vortex (23.66%), and Type 4 Northward typhoon (26.51%). The total frequency of each type may not be 100% due to rounding. Figure 3 As shown in (b), the daily weather type label sequence is output. The composite field of sea level pressure and 500hPa geopotential height for each weather type is subjected to synoptic verification, and the type is named by combining the comprehensive weather map of typical historical rainstorm cases.

[0140] Step 3: Extracting precipitation events;

[0141] Step 3.1, Extraction of precipitation events in districts and counties: The arithmetic mean of the daily rainfall at each rain gauge station within a district or county is taken as the average rainfall of that district or county; dates with an average rainfall of 1 mm or more are recorded as wet days; when the interval between adjacent wet days exceeds 1 day, they are divided into different precipitation events; for each precipitation event, the duration is the number of days between the start and end dates of the event (unit: days), the total rainfall is the sum of the average rainfall within the event (unit: mm), and the maximum daily rainfall is the maximum value of the average rainfall within the event (unit: mm).

[0142] Step 3.2, Sub-basin flood event extraction: Perform Lyne-Hollick single-parameter digital filtering baseflow segmentation on the daily flow process of each sub-basin. The recursive formula is as follows:

[0143]

[0144] In the formula: for Real-time flow This represents the separated direct runoff component (taken as non-negative). For the filter coefficients, in this embodiment Take 0.995; the base current component is and The difference is obtained by filtering the baseflow process four times in alternating forward and reverse directions, which is used to determine the regression baseflow criterion for the start and end times of flood events.

[0145] The flood peak threshold is adaptively determined by the following formula:

[0146]

[0147] In the formula: For the peak flood threshold, , These are the 90th and 99th percentiles of the flow sequence for this sub-basin. To preset the minimum threshold, this embodiment uses 5 m³ / s; the coefficients 0.6 and 0.2 are determined by joint debugging in the watershed of the embodiment based on the number of identified events and the false negative rate, and can be taken in the range of 0.5-0.7 and 0.15-0.25 respectively.

[0148] The threshold for the initial slope is determined by the following formula:

[0149]

[0150] In the formula: The threshold for the initial rise slope (unit: cubic meters per second per day); The slope benchmark value is selected based on the sub-basin area, which is less than 1000 km². 2 Take 2, 1000 - 3000km 2 Take 5, greater than 3000km 2 Take 10 (unit: cubic meters per second per day); The maximum flow rate (in cubic meters per second) of the flow sequence for this sub-basin; the constant 100 is a normalized reference value for the flow rate (in cubic meters per second). Recession slope threshold. Pick .

[0151] The start time, end time, peak flow, and duration of each flood are obtained based on the flood event identification rules. The flood event identification rules are as follows: the flow rate increase at adjacent time steps exceeds the threshold of the flood inflection slope. The point is determined as the starting point; the maximum flow rate after the starting point is greater than or equal to the peak flow threshold. The peak of a flood is confirmed at that time; the receding water level after the peak is defined as the water level drop in adjacent time steps falling below the threshold of the receding slope. The system combines a receding water determination ratio (0.15 in this embodiment, meaning the flow rate drops to below 15% of the peak flow) to jointly determine the end point of each flood event. A dynamic slope coefficient (0.02 in this embodiment, meaning 2% of the flow rate variation is used as the threshold for determining a flat section) is used to identify the smoothness of the receding water phase. The shortest event duration threshold is set to 0 in this embodiment (i.e., no minimum duration limit is set). The flow sequence time resolution is daily. This yields the start time, end time, peak flow rate, and duration of each flood event.

[0152] Step 4: Matching events with disaster situations;

[0153] Step 4.1, County / District Scale Matching: For disaster event records, including floods, mudslides, landslides, and urban waterlogging, if the date of a disaster falls within the start and end interval of a certain precipitation event, then the disaster is linked to that precipitation event.

[0154] Step 4.2, Sub-basin scale matching: For sub-events of sub-basin disaster formed in step 1 (4), if the time of occurrence falls within the start and end interval of a flood event in that sub-basin, the disaster record is horizontally associated with that flood event; disaster records without matching are removed. For flood events that only occur as floods without accompanying disasters, supplement them as samples of the "rainstorm-flood" two-stage process.

[0155] Step 4.3, Dual-scale event merging: Based on the sub-basin-county association table, merge the sub-basin flood events into the corresponding county-level precipitation event table: If the flood rise date falls within the start and end interval of a precipitation event (with the end date extended by 1 day), take the precipitation event whose start date is closest to the flood rise date and link it, and write the flood start and end times and peak flow; if the precipitation event has been linked to other disaster records, copy the newly added flood record in that row to achieve one-to-many expansion; for precipitation events without corresponding flood events, their flood field is recorded as none.

[0156] Step 4.4, Representative Date and Previous Rainfall. Determine a representative date for each event: for events matching a disaster, use the date the disaster occurred; for events not matching a disaster, use the event's start date. Calculate the sum of the average rainfall over the three days preceding the representative date as the previous rainfall Ant3 (unit: mm).

[0157] Step 4.5: Labeling of Disaster Chain Parent-Child Relationships. Based on the order of disaster chain occurrence (first, second, and third order) in the disaster log, label each disaster record with upstream and downstream disaster tags. This labeling is used for statistical analysis of disaster chain transmission paths and to verify the structural assumption in Step 6 that "no directed edges are set between disaster nodes" (statistics show that the direct upstream of each disaster is a rainstorm or flood, and there are no stable lateral transmission records between disasters).

[0158] Step 5: Tagging and Discretization;

[0159] Step 5.1, Flood Classification. Compare the peak flow of each flood with the 20-year return period design flow Q20 and the 50-year return period design flow Q50 of the sub-basin: if the peak flow is less than Q20, it is classified as mild; if it is greater than or equal to Q20 and less than Q50, it is classified as moderate; if it is greater than or equal to Q50, it is classified as severe; if no flood occurs, it is classified as none.

[0160] Step 5.2: Classification of landslides, debris flows, and urban flooding. The four indicators—number of affected people, collapsed houses, affected crop area, and direct economic losses—are mapped to scores using the following indicator interval-score comparison table (interval endpoints are assigned using the convention of "greater than the lower limit and less than or equal to the upper limit," and the thresholds for each segment can be determined based on disaster statistics for the study area; this example is as follows):

[0161] Number of people affected: ≤10: 1 point; 10-100: 2 points; 100-1000: 3 points; 1000-1000: 3 points. 4 4 points total, 10 4 ~10 5 5 points total, 10 5 ~10 6 6 points total, 10 6 ~10 7 7 points total, 10 7 ~10 8 8 points, greater than 10 8 Total score: 9 points;

[0162] House collapses (number of rooms): ≤10 rooms: 2 points; 10-50 rooms: 3 points; 50-100 rooms: 4 points; 100-500 rooms: 5 points; 500-1000 rooms: 6 points; 1000-5000 rooms: 7 points; 5000-1000 rooms: 7 points. 4 Total 8 points, 10 4 ~2×10 4 Counted as 9 points, greater than 2×10 4 10 points;

[0163] Affected crop area (mu, 1 mu is approximately 0.0667 hectares): ≤100 fen, 100-500 fen, 500-1000 fen, 1000-5000 fen, 5000-10 ... 4 6 points total, 10 4 ~5×10 4 7 points total, 5 x 10 4 ~10 5 Total 8 points, 10 5 ~5×10 5 Counted as 9 points, greater than 5 × 10 5 10 points;

[0164] Direct economic loss (ten thousand yuan): ≤100 points = 1 point, 100-1000 points = 2 points, 1000-1000 points = 1 point. 4 3 points total, 10 4 ~5×10 4 4 points total, 5 x 10 4 ~10 5 5 points total, 10 5 ~5×10 5 Total 6 points, 5 x 10 5 ~10 6 Total 8 points, 10 6 ~10 7 Count 9 points, greater than 10 7 10 points.

[0165] The average of all non-missing scores is taken as the comprehensive disaster loss score; a comprehensive score of 3.33 or less is recorded as mild, a score greater than 3.33 and less than or equal to 6.67 is recorded as moderate, and a score greater than 6.67 is recorded as severe; disasters with records but no scoring data are recorded as mild; and disasters that did not occur are recorded as none.

[0166] Step 5.3, Disaster Loss Index Classification (interval endpoints are all "greater than the lower limit and less than or equal to the upper limit"): 10 or fewer affected people are classified as mild, 10 or more but less than or equal to 100 as moderate, and more than 100 as severe; 100 or fewer houses collapsed are classified as mild, 100 or more but less than or equal to 1000 as moderate, and more than 1000 as severe; 3300 or fewer hectares (approximately 50,000 mu) of affected crops are classified as mild, 3300 or more but less than or equal to 13300 hectares (approximately 200,000 mu) as moderate, and more than 13300 hectares as severe; direct economic losses of 1 million yuan or less are classified as mild, 1 million yuan or more but less than or equal to 10 million yuan as moderate, and more than 10 million yuan as severe; no disaster records are uniformly recorded as none. The above classification thresholds can be determined based on historical disaster data of the study area.

[0167] Step 5.4, Disaster-causing factors discretization. The maximum daily rainfall and total rainfall are divided into four levels according to 50 mm, 100 mm, and 250 mm, respectively denoted as below torrential rain, torrential rain, heavy torrential rain, and extremely heavy torrential rain (using the 24-hour precipitation level names as a naming convention); the duration is divided into four levels according to 1 day, 3 days, and 5 days, namely short duration, medium duration, long duration, and relatively long duration; the previous rainfall is divided into three levels according to 10 mm and 25 mm, namely less, moderate, and more.

[0168] Step 5.5: Weather Type Linking. Search for the corresponding weather type in the daily weather type label sequence based on the representative date of the event, and write it into the weather type field of the event sample table. This ultimately forms an event sample table containing 13 fully discrete fields: weather type, 4 disaster-causing factors, 4 disaster nodes, and 4 disaster loss nodes.

[0169] Step 6: Bayesian network construction.

[0170] Step 6.1, Structural Assumptions: No directed edges are set between disaster nodes. For example... Figure 2As shown, a four-layer directed acyclic graph is manually defined using domain knowledge: The first layer is the weather-type root node weather_type; the second layer consists of disaster-causing factor nodes duration, total_prcp, max_prcp, and Ant3, all with weather_type as their parent node; the third layer consists of disaster nodes flood, landslide, debris flow, and waterlogging, with all four disaster-causing factor nodes as their parent nodes; the fourth layer consists of disaster-damaged nodes affected_people, collapsed_houses, crop_damage_area, and direct_econ_loss, with all four disaster nodes as their parent nodes.

[0171] Step 6.2, Calculation of Conditional Probability Table. Based on the event sample table, the conditional probability table is calculated node-by-node using an equivalent sample size based on Dirichlet priors:

[0172]

[0173] The symbols in the formula are defined as before; the equivalent sample size smoothing coefficient in this embodiment. Take 1, It can take values ​​in the range of 0.5 to 5. equal This is equivalent to classic Laplace smoothing. In this embodiment, the number of parent combinations for each disaster node in the third layer is 4×4×3×4=192, and the number of parent combinations for each disaster node in the fourth layer is 4×4×4×4=256. The number of parent combinations is on the same order of magnitude as or even greater than the number of event samples. Under maximum likelihood estimation, the conditional probability of a large number of parent combinations without samples cannot be defined. However, all conditional probabilities calculated according to the above formula are strictly positive and normalized, avoiding zero probability and ill-conditioned estimation under sparse samples. The conditional probability tables of each node are assembled into a complete discrete Bayesian network.

[0174] Step 6.3: Construct independent chain-based Bayesian networks for each disaster. In addition to the directed acyclic graph (DAG) Bayesian network, construct independent chain-based Bayesian networks for four disasters: flood, landslide, urban flooding, and debris flow: Weather type → {duration, total rainfall, maximum daily rainfall, previous 3-day rainfall} → Disaster type → {four disaster-damaged nodes}. During construction, only event samples of actual disaster occurrence (not zero level) are used, and the state space of the disaster nodes is cleaned of "zero" states, allowing the conditional probability table to focus on characterizing the level distribution "under the condition of this disaster occurrence". The DAG Bayesian network is used for panoramic risk queries across disasters, while the independent chain-based Bayesian networks are used for fine-grained level inference under single disaster occurrence conditions. When comparing the inference results of the two types of networks, the posterior states of the disaster nodes in the DAG Bayesian network must first be cleaned of "zero" states and re-normalized to unify the probabilistic semantics.

[0175] Step 7: Risk Assessment;

[0176] Step 7.1: Compiling the Joint Tree. The directed acyclic graph of the Bayesian network is moralized and triangulated to construct a clique tree. Message passing is used to implement evidence propagation and posterior query, supporting accurate posterior inference under arbitrary evidence. This embodiment uses the compile, setEvidence, and querygrain functions of the gRain package in R language.

[0177] Step 7.2, Forward Inference Based on Fixed Weather Pattern. Using a Bayesian network with a directed acyclic graph structure, and taking weather pattern nodes as evidence (e.g., setting "weather pattern = low-level vortex pattern"), the posterior probability distributions of each disaster-causing factor, disaster, and damage node are queried to obtain a risk profile of each link in the disaster chain under this weather pattern. Figure 4 As shown, forward inference was performed on the four weather types to compare the risk differences of each link in the disaster chain under different circulation backgrounds.

[0178] Step 7.3: Reverse Inference Based on Fixed Disaster Scenarios. Using an independent chain-based Bayesian network and taking the disaster node level as evidence (e.g., setting flood = severe, i.e., a flood with a return period greater than 50 years), reverse-engineer the posterior probability distributions of weather patterns, causative factors, and disaster-damaged nodes. This identifies the most probable combination of weather background and causative factors that triggers the scenario, along with their corresponding disaster damage levels, providing a basis for scenario extrapolation and disaster prevention resource pre-positioning.

Claims

1. A method for inferring the risk of rainstorm and flood disaster chains based on weather classification and Bayesian networks, characterized in that, Includes the following steps: Step 1: Collect atmospheric reanalysis data, daily rainfall data from rain gauges, flow data from hydrological stations, historical disaster records, and design flow data for the return period of sub-basins in the study area; the atmospheric reanalysis data includes at least the daily sea-level pressure field and the 500 hPa geopotential height field; the historical disaster records include disaster records at the administrative region scale and the sub-basin scale. Step 2: Perform daily climatological anomaly, latitudinal cosine area weighting, and grid-by-grid standardization on the sea level pressure field and the 500hPa geopotential height field. After splicing the sea level pressure field and the 500hPa geopotential height field along the grid dimension, perform dimensionality reduction using empirical orthogonal functions. Retain the principal component time series whose cumulative variance contribution rate reaches the preset threshold. Perform K-means clustering on the principal component time series and determine the optimal number of subtypes to obtain the daily weather type label sequence. Step 3: At the administrative scale, calculate the average rainfall from the daily rainfall data of the rain gauge station, segment the precipitation events according to the wet day threshold and dry period threshold, and calculate the duration, total rainfall and maximum daily rainfall of the precipitation events. At the sub-basin scale, the baseflow of hydrological station flow data is segmented, and flood events are automatically identified based on the flood peak threshold, flood initiation slope threshold and flood recession slope threshold determined adaptively according to the flow quantile, so as to obtain the start and end times of flood events and the flood peak flow. Step 4: Based on the start and end times of the flood event, improve the disaster event record; for the improved disaster event record, match the precipitation event with the nearest start date within a preset number of days; the disaster event record includes flood disasters, landslide disasters, debris flow disasters, and urban flooding disasters; Step 5: Classify the disaster event records matched with precipitation events into different levels and discretize the disaster-causing factors to form a fully discrete event sample table; Step 6: Obtain the Bayesian network for probability inference based on the daily weather label sequence and event sample table; Step 7: Perform forward inference based on a fixed weather pattern using a Bayesian network to obtain the posterior probability distribution of the disaster-causing factors, disaster level, and disaster damage level under a given weather pattern; perform backward inference based on a fixed disaster scenario to obtain the posterior probability distribution of the weather pattern, disaster-causing factors, and disaster damage level under a given disaster level.

2. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, Step 2 involves performing daily climatological anomalies, latitudinal cosine area weighting, and grid-by-grid standardization on the sea level pressure field and the 500hPa geopotential height field, including: For each grid point, subtracting its multi-year average over the same time period yields the anomaly field. : In the formula: For grid points At any moment The original field value, For a moment The corresponding month-day sequence number, For grid points On calendar days The multi-year average; Weighted by taking the square root of the area of ​​the cosine of latitude for the anomaly field: In the formula: For grid points At any moment The anomaly field value after latitude and longitude area weighting, For grid points Latitude; Standardization is achieved by dividing each grid point of the open-weighted anomaly field by the standard deviation of the anomaly field.

3. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, In step 2, the sea level pressure field and the 500 hPa geopotential height field are concatenated along the grid dimension and then subjected to dimensionality reduction using an empirical orthogonal function. The principal component time series whose cumulative variance contribution rate reaches a preset threshold are retained. This includes: concatenating the sea level pressure field and the 500 hPa geopotential height field along the grid dimension into a standardized wide matrix. For standardized wide matrices Calculate the time covariance matrix , The dimension is time step × time step; extract Eigenvalues ​​arranged in descending order of eigenvalues and corresponding feature vectors The principal component time series is taken until the cumulative variance contribution rate first reaches the preset threshold. The cumulative variance contribution rate is calculated by summing the extracted eigenvalues ​​and... Calculation of the ratio of traces.

4. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, Step 2, performing K-means clustering on the principal component time series and determining the optimal number of clusters, includes: Within the preset genotype number scanning range, K-means clustering based on Euclidean distance is performed on each principal component time series, and the genotype number is recorded. The corresponding sum of squared errors within groups The range normalization of the within-group sum of squared errors sequence is performed using the following formula: And calculate the second-order difference: In the formula: , These are the minimum and maximum values ​​of the sum of squared errors within the group within the scanning range, respectively. The optimal number of morphologies is the number of morphologies corresponding to the point with the maximum absolute value of the second difference at the inner points of the scanning range. Then, a formal K-means clustering with multiple random starting points is performed using the optimal number of morphologies. The clustering result with the minimum sum of squared errors within the group is taken, and each day is assigned to the nearest cluster center to obtain the daily weather label sequence.

5. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, In step 3, the average rainfall is calculated from the daily rainfall data of the rain gauges. Precipitation events are segmented based on wet day and dry period thresholds, and the duration, total rainfall, and maximum daily rainfall for each event are calculated, including: The arithmetic mean of the daily rainfall at each rain gauge station within the administrative region is taken as the area average rainfall; dates with area average rainfall greater than or equal to the wet day threshold are recorded as wet days; when the interval between adjacent wet days exceeds the dry period threshold, they are divided into different precipitation events. For each precipitation event, the duration is the number of days between the start and end dates of the precipitation event, the total rainfall is the sum of the average rainfall within the precipitation event, and the maximum daily rainfall is the maximum value of the average rainfall within the precipitation event.

6. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, In step 3, the flood peak threshold, flood initiation slope threshold, and flood receding slope threshold, adaptively determined according to the flow quantile, include: The formula for the flood peak threshold is as follows: In the formula: For the peak flood threshold, , These are the 90th and 99th percentiles of the flow sequence for this sub-basin. To preset the minimum threshold lower limit, , For preset coefficients, Take a value of 0.5 to 0.

7. Take values ​​of 0.15 to 0.25, with flow rates measured in cubic meters per second; The formula for the slope threshold is as follows: In the formula: The threshold for the initial slope. The slope benchmark value is selected based on the preset area interval to which the sub-basin area belongs. This represents the maximum flow rate in the flow sequence of this sub-basin. This serves as a reference value for normalized flow rate. and All values ​​are in cubic meters per second; the receding slope threshold is the rising slope threshold multiplied by a preset reduction factor that is greater than 0 and less than 1.

7. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, Step 5 includes: Flood disasters are classified into levels based on a comparison of peak flow and design flow data with return period. Landslide disasters, debris flow disasters, and urban flooding disasters are classified into levels based on disaster loss indicators. These disaster loss indicators include the number of affected people, collapsed houses, affected crop area, and direct economic losses. The disaster loss indicators are divided into four levels: none, mild, moderate, and severe. The duration, total rainfall, maximum daily rainfall, and previous rainfall of precipitation events are discretized into a preset number of levels according to preset classification thresholds. Weather tags are attached to each disaster event record from the daily weather tag sequence based on the representative date of the disaster event record, forming a fully discrete event sample table.

8. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, Step 6 includes: using the daily weather model label sequence as the root node, duration, total rainfall, maximum daily rainfall and previous rainfall as the second-level nodes, disaster types such as floods, landslides, debris flows and waterlogging as the third-level nodes, and disaster loss indicators as the fourth-level nodes, to construct a Bayesian network with a directed acyclic graph structure. Based on the event sample table, the conditional probability table is calculated node-by-node using an equivalent sample size based on Dirichlet prior, as shown in the following formula: In the formula: For the current node, for Parent node combination, The combined value of the parent node is And the current node's value is The number of samples, The combined value of the parent node is The total number of samples, The number of states of the current node. The equivalent sample size smoothing coefficient is greater than 0.

9. The method for inferring the risk of rainstorm and flood disaster chains according to claim 1, characterized in that, In step 7, forward inference and backward inference are achieved through joint tree inference: the directed acyclic graph of the Bayesian network is moralized and triangulated to construct a clique tree, and evidence propagation is completed through message passing based on the set evidence node states, and the posterior probability distribution of the target node is queried.

10. A system for inferring the risk of rainstorm and flood disaster chains based on weather classification and Bayesian networks, characterized in that, Includes the following modules: The data acquisition module collects atmospheric reanalysis data, daily rainfall data from rain gauges, flow data from hydrological stations, historical disaster records, and design flow data for the return period of sub-basins in the study area. The atmospheric reanalysis data includes at least the daily sea-level pressure field and the 500 hPa geopotential height field. The historical disaster records include disaster records at the administrative region scale and the sub-basin scale. The weather classification module is used to perform daily climatological anomalies, latitudinal cosine area weighting, and grid-by-grid standardization on the sea level pressure field and the 500hPa geopotential height field. After splicing the sea level pressure field and the 500hPa geopotential height field along the grid dimension, the module performs dimensionality reduction using empirical orthogonal functions, retains the principal component time series with a cumulative variance contribution rate reaching a preset threshold, performs K-means clustering on the principal component time series, determines the optimal number of classifications, and obtains the daily weather type label sequence. The event extraction module is used to calculate the average rainfall on the surface from daily rainfall data from rain gauges at the administrative scale, segment precipitation events based on wet day thresholds and dry period thresholds, and calculate the duration, total rainfall, and maximum daily rainfall of precipitation events. At the sub-basin scale, the baseflow of hydrological station flow data is segmented, and flood events are automatically identified based on the flood peak threshold, flood initiation slope threshold and flood recession slope threshold determined adaptively according to the flow quantile, so as to obtain the start and end times of flood events and the flood peak flow. The event matching module is used to improve the disaster event record based on the start and end times of the flood event; for the improved disaster event record, it matches the precipitation event with the nearest start date within a preset number of days; the disaster event record includes flood disasters, landslide disasters, debris flow disasters and waterlogging disasters; The tagging module is used to classify disaster event records matched with precipitation events into different levels and discretize the disaster-causing factors to form a fully discrete event sample table. The network construction module is used to obtain a Bayesian network for probabilistic inference based on the daily weather-type label sequence and event sample table; The risk inference module is used to perform forward inference based on a fixed weather pattern using a Bayesian network to obtain the posterior probability distribution of disaster-causing factors, disaster level, and disaster damage level under a given weather pattern; and to perform backward inference based on a fixed disaster scenario to obtain the posterior probability distribution of weather pattern, disaster-causing factors, and disaster damage level under a given disaster level.

Images