Urban haze disaster risk assessment method based on information entropy improved PCA-ISSA-CatBoost model

Abstract

The invention relates to the technical field of haze disaster risk assessment, in particular to an urban haze disaster risk assessment method based on an information entropy improved PCA-ISSA-CatBoost model, and the method comprises the steps: S1, obtaining a sample data set based on an index feature system; s2, dividing the sample data set into a plurality of initial pellets through pellet calculation, and refining pellet levels of the initial pellets step by step according to the granularity to form a pellet set with a plurality of levels; weighting each particle ball in each level of particle balls, and performing principal component analysis to obtain a corresponding principal component; s3, fusing the principal components of the pellets to obtain a global principal component; screening out a preset number of features to form an initial feature set Fs; s4, a CatBoost model is constructed; s5, searching an optimal feature subset Fopt of the initial feature set Fs and an optimal hyper-parameter theta opt of the CatBoost model by using a sparrow search algorithm; s6, carrying out the training of the CatBoost model; and S7, carrying out haze disaster risk assessment by using the trained CatBoost model. According to the method, accurate, reliable and low-dependency risk prediction can be carried out on haze disasters.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of haze disaster risk assessment, and particularly relates to a city haze disaster risk assessment method based on an information entropy improved PCA-ISSA-CatBoost model BACKGROUND

[0002] At present, air pollution has become one of the most serious environmental problems in the world. Among them, haze will have serious adverse effects on human health. Many previous studies have found that high concentrations of PM will increase the risk of death and morbidity. Urban haze pollution is becoming more and more serious, and is increasingly affecting all aspects of social development.

[0003] Research on city haze risk assessment is of great significance to the governance of haze disasters. In order to reduce the harm of haze disasters to society, an effective haze disaster risk prediction model is urgently needed. Clarifying the causal relationship between the influencing factors of haze disasters is the premise of establishing a haze prediction model.

[0004] However, the current research does not evaluate haze pollution from the perspective of multiple risks, and the model constructed by the index system evaluation method has a certain subjectivity in determining the index weight, thereby affecting the accuracy of risk assessment. The evaluation process based on the historical disaster mathematical statistics evaluation method is relatively simple, but it is highly dependent on historical data, and new disaster data needs to be continuously introduced to ensure the reliability of the evaluation results.

[0005] Specifically, in the existing method, the index system-based method relies on expert experience weighting (such as AHP hierarchical analysis method), which is highly subjective and prone to human bias (for example: the index weight setting is highly arbitrary, affecting the importance ranking of features). The method based on historical disaster statistics requires continuous updating of high-quality disaster data support, but in practice, historical data is sparse, noisy, and unevenly covered in space and time, and the model has weak generalization ability (for example: the evaluation results are distorted when data is missing).

[0006] This is because both methods are limited by the way data is processed - the former relies on artificial prior knowledge, and the latter is subject to data acquisition costs and quality. The essence is that the scientificity and adaptability of the feature extraction and optimization process are insufficient.

[0007] Therefore, how to accurately, reliably and low-dependently predict the risk of haze disasters has become a problem to be solved at present. SUMMARY

[0008] In view of the deficiencies of the prior art, the present application provides a city haze disaster risk assessment method based on an information entropy improved PCA-ISSA-CatBoost model, which can accurately, reliably and low-dependently predict the risk of haze disasters.

[0009] To solve the above technical problems, the present application adopts the following technical solutions:

[0010] The city haze disaster risk assessment method based on information entropy improved PCA-ISSA-CatBoost model comprises the following steps:

[0011] S1, based on the index characteristic system for assessing haze disaster risk, the corresponding sample data set of the target city is obtained; each sample in the sample data set includes the corresponding data of each feature in the index characteristic system;

[0012] S2, the sample data set is divided into multiple initial particle balls by particle ball calculation, and the initial particle balls are gradually refined in particle size to form multiple levels of particle ball sets; the entropy of each feature of the samples in each level of particle ball set is calculated as the corresponding weight to obtain the feature weight matrix of each level; for each particle ball in each level of particle ball, the corresponding level of feature weight matrix is used for weighting and principal component analysis to obtain the corresponding principal component;

[0013] S3, the principal components of each particle ball in each level are fused by weighted fusion to obtain global principal components; the load values of each feature in the index characteristic system are analyzed based on the global principal components to select a preset number of features to form an initial feature set Fs;

[0014] S4, a CatBoost model is constructed for haze disaster risk assessment according to the feature set data of haze;

[0015] S5, the sparrow search algorithm is used to initialize the feature subset of the initial feature set Fs and the hyperparameters of CatBoost as sparrow individuals X i , and a fitness function f(X i ) based on the CatBoost model is defined to measure the advantages and disadvantages of individuals, search for the optimal feature subset F opt of the initial feature set Fs and the optimal hyperparameters θ opt of the CatBoost model;

[0016] S6, the optimal feature subset F opt corresponding to the feature set data and the optimal hyperparameters θ opt are used to train the CatBoost model;

[0017] S7, the optimal feature subset F opt corresponding to the feature set data in the actual data of the target city is obtained, and the trained CatBoost model is used for haze disaster risk assessment.

[0018] Compared with the prior art, the present application has the following beneficial effects:

[0019] 1. Realize objective adaptive feature weight allocation and reduce subjective dependence. Through information entropy weighted PCA and particle ball calculation, the weight of each level feature is dynamically calculated, and based on this, the key features are screened. Compared with the index system evaluation method in the prior art (such as expert weighting or analytic hierarchy process), this method avoids the subjective bias caused by manual weight setting (such as misjudgment of the importance of features), and significantly improves the objectivity of feature screening. This ensures that the risk assessment model can automatically extract the core disaster-causing factors from multiple factors (such as weather, pollutant concentration), and improves the accuracy of the assessment.

[0020] 2. Joint optimization of features and hyperparameters is realized through sparrow search algorithm to reduce error accumulation risk. The algorithm is used to optimize the feature subset and CatBoost hyperparameters (such as learning rate or tree depth) at the same time, and the model performance is used as the fitness function to drive the search process. Compared with the step-by-step optimization method in the prior art (such as manually screening features first and then adjusting parameters, or relying on grid search), this scheme completes double objective optimization in one iteration process, reducing the intermediate error accumulation in the optimization phase. This greatly improves the model training efficiency and overall stability, effectively alleviating the overfitting or local optimum problem when facing high-dimensional data.

[0021] 3. Enhance the robustness and generalization ability of the model, and reduce the dependence on historical data. The combination of CatBoost model and information entropy improved feature dimension reduction automatically processes high-dimensional, heterogeneous data (such as missing values or noise) during the training phase, and retains the core information through global principal component. Compared with the historical disaster mathematical statistics method, this method reduces the excessive dependence of the model on new disaster data (such as the need to continuously update historical samples to maintain accuracy), and improves the reliability of the model in the case of sparse or incomplete data. This makes the risk assessment more adaptable to the dynamic changes of actual urban haze problems.

[0022] In summary, the method can accurately, reliably and low-dependently predict the risk of haze disasters. By integrating objective weighting, efficient optimization algorithm and advanced integrated model, the scheme effectively overcomes the shortcomings of high subjectivity, low optimization efficiency and strong data dependence of the prior art, and provides more scientific and adaptive decision support for urban haze governance.

[0023] Preferably, in S1, the first-level indexes of the index feature system include disaster-causing factor risk, disaster environment sensitivity and disaster-bearing body; the second-level indexes of the disaster-causing factor risk include economic density, number of motor vehicles, proportion of secondary industry, housing construction area and proportion of coal consumption; the second-level indexes of the disaster environment sensitivity include annual average concentrations of VOCs, PM2.5, PM10, SO2 and NO2; and the second-level indexes of the disaster-bearing body include population density, number of health institutions and built-up area green area.

[0024] Such a setting, 1, to build a multi-dimensional risk assessment framework, to solve the index coverage one-sidedness. Through the division of the first index (disaster-causing factor, disaster-prone environment, disaster-bearing body), it first introduces the disaster formation mechanism (hazard), environmental background conditions (sensitivity) and social disaster resistance ability (vulnerability) into a unified system, breaking through the limitation of traditional methods which only focus on a single dimension (such as pollutant concentration). Compared with the past research which only analyzes the haze pollutant itself (such as PM2.5 concentration), this scheme clearly introduces social and economic driving factors (such as economic density, proportion of secondary industry) and disaster-bearing capacity indicators (such as the number of health institutions), and more comprehensively reveals the complex causes of haze risk.

[0025] 2, the index design has scientificity and feasibility, reduces the threshold of practice. The secondary indicators all use quantifiable and easily accessible public data (such as "number of motor vehicles", "coal consumption proportion", "green area of built-up area"), which do not need to rely on professional equipment or high-cost monitoring, greatly improving the actual operability of the model. Traditional methods often fail to obtain index data (such as real-time monitoring of specific pollutants) or have ambiguous definitions (such as "environmental sensitivity" lacks quantitative standards), resulting in evaluation remaining at the theoretical level; this scheme selects common indicators from statistical yearbooks and environmental bulletins to ensure that the evaluation results can be directly used for policy making.

[0026] 3, accurately anchor key risk driving factors, improve the pertinence of evaluation. The secondary index design directly points to the core mechanism of haze disaster, for example, "VOCs, SO2, NO2 concentration" is related to photochemical pollution generation, "population density" and "green area" reflect the exposure and mitigation capacity of disaster, so that the model can distinguish the differentiated risk sources of different cities. Existing index system often has index redundancy (such as including irrelevant environmental parameters) or omission of key factors (such as ignoring the influence of industrial structure), while this scheme selects indicators based on disaster chain logic (such as economic density → industrial emission → pollution aggravation), which significantly improves the causal explanatory power of the evaluation results.

[0027] Preferably, in S2, when dividing the initial particle ball, for each initial particle ball G j , the particle ball center is:

[0028]

[0029] where |. | represents the number of samples in the initial particle ball G j , and x i represents the i-th sample in the initial particle ball G j ;

[0030] The particle ball radius is:

[0031] Such a setting, through the mean and maximum distance quantization design, provides a high robustness and adaptability solution for the initial division of granules. It not only solves the pain points of strong subjectivity, incomplete coverage and insufficient flexibility in the prior art, but also improves the scientific rigor of the initialization stage through statistical principles.

[0032] Preferably, S2 comprises:

[0033] For a sample data set X = {x1, x2,..., x n}, wherein each sample is divided into m initial granules G1, G2,..., G m by granule calculation; and the m initial granules are refined in granule size step by step to form a plurality of granule sets , wherein k represents the number of granule levels after step-by-step refinement, represents the initial granule G m after step-by-step refinement of the kth level granule size;

[0034] For the pth level granule set, 1≤p≤k, the entropy of the jth feature of the sample in the set is calculated:

[0035]

[0036] In the formula, p ji represents the normalized probability of feature j in sample i, and n represents the number of samples in the granule set;

[0037] The pth level granule set entropy matrix W = diag(H1, H2,..., H d ) is obtained;

[0038] For the feature matrix Xq of the sample in the qth granule of the pth level, use the entropy matrix W of this layer to weight X to obtain the corresponding weighted matrix Then, the principal component analysis is performed on the weighted matrix , and the eigenvectors corresponding to the largest y eigenvalues are selected to form the principal component matrix

[0039] Such a setting, 1, realizes hierarchical feature objective weighting, and eliminates subjective deviation in weight setting. By calculating the entropy (H j ) of each feature in each granule set and generating an entropy matrix (W), the importance of the feature is completely determined by the dynamic data distribution. Traditional PCA or fixed weight dimension reduction method (such as equal weight weighting) is easily disturbed by expert experience or ignores the local importance of the feature, while the present scheme uses the information value of the entropy quantification feature in the hierarchical granule to significantly improve the adaptability and scientificity of weight allocation.

[0040] 2. Fusion of local and global feature expressions to enhance data representation capabilities. Multi-level refinement of granules (from coarse to fine granularity) allows different levels to focus on data structures of different scales. Entropy-weighted PCA acts on each layer of granules separately, preserving both macroscopic distribution patterns and extracting microscopic local patterns. In existing technologies, global PCA is prone to losing details due to data heterogeneity (e.g., sparse regional features are submerged). Hierarchical processing, while retaining the global principal components, captures regional specificity in urban haze data (e.g., differences in pollution between industrial and residential areas), avoiding information loss from "one-size-fits-all" dimensionality reduction.

[0041] 3. Principal component analysis for directional dimensionality reduction, improving feature noise resistance and discriminability. Entropy matrix weighting High-information features are strengthened, while low-information features are weakened, allowing subsequent PCA to focus on key dimensions. Selecting the principal component corresponding to the largest eigenvalue further filters out redundant noise. Traditional PCA treats features equally and is susceptible to interference from highly correlated features or noise (such as collinearity between meteorological and emission sources). This approach, however, utilizes a dual mechanism of entropy-weighted pre-screening and hierarchical principal component extraction to improve the discriminative efficiency and robustness of features after dimensionality reduction.

[0042] Preferably, S3 includes:

[0043] The principal component matrices of each sphere in each layer are weighted and fused to obtain the global principal component:

[0044]

[0045] Where, α p represents the weight of the p-th layer granularity, is the qth particle in the pth layer The weight of

[0046] Perform dimensionality reduction analysis on the global principal components and select b features with the largest load value in each principal component to form the initial feature set Fs.

[0047] This setting can 1. resolve the conflict of multi-scale data structures and achieve unified expression of global features. p Adjust the importance of different granularity levels, This approach distinguishes data density differences between particles in the same layer, fusing the dispersed principal component information into a consistent global representation. Traditional principal component fusion methods (such as simple concatenation or mean pooling) ignore the hierarchical and local variations in data structure, leading to interference between macro-trends and micro-patterns. This approach, through differentiated weight allocation, allows heterogeneous features, such as high-pollution details in industrial areas and urban background values, to coexist harmoniously in a unified space.

[0048] 2. Feature screening driven by load value, breaking the dependence on artificial experience. Key features are screened based on the absolute value of the principal component load value (the contribution weight of the feature to the principal component), making the construction of Fs completely driven by data statistical rules. Existing methods rely on artificial threshold setting or feature importance ranking of models such as random forest, which is easily disturbed by subjective preset or trapped in local optimum; this scheme uses a mathematically interpretable statistical quantity (load value) as the screening standard to eliminate experience bias from the root and ensure that the features are strongly related to the risk cause.

[0049] Preferably, in S5, the sparrow search algorithm is used for the optimized sparrow search algorithm, wherein the improvement includes: when updating the position of the discoverer, a nonlinear control parameter is introduced to coordinate the global search and local development ability of the algorithm; when updating the position of the joiner, a Levy flight strategy is introduced to disturb the current optimal solution and enhance the local exploitation ability.

[0050] Such a setting, 1, balances global search and local development, breaking through the convergence dilemma of traditional algorithms. The nonlinear control parameter dynamically adjusts the exploration step of the discoverer (such as large-scale global exploration in the early stage and fine local development in the later stage), avoiding the algorithm from falling into premature convergence or ineffective oscillation. The original sparrow search algorithm (SSA) uses a fixed linear parameter, which is easy to miss the global optimal solution in complex optimization problems (such as converging to a local suboptimal feature subset of haze risk assessment); ISSA (this algorithm) significantly improves the solution space coverage ability in high-dimensional and multi-peak scenarios through adaptive nonlinear mechanism.

[0051] 2. Strengthen the disturbance of local optimal solution to improve optimization accuracy and stability. The Levy flight strategy gives the joiner the ability to jump long distances and search short distances, and through intelligent disturbance of the current optimal solution, it breaks through the local extreme value trap. Traditional random disturbance (such as Gaussian noise) is easy to cause the optimization direction to deviate or converge slowly, while the heavy-tailed distribution characteristics of Levy flight (short-distance intensive search + long-distance jump) balance development and exploration, and accurately locate the optimal configuration in the simultaneous optimization of CatBoost hyperparameters and feature subsets.

[0052] 3. Improve the efficiency of solving complex optimization problems and reduce the dependence on parameter tuning. The double-strategy cooperation greatly reduces the number of iterations required for convergence (such as quickly locking the optimal feature subset F opt and the optimal hyperparameter θ opt ), while reducing the sensitivity to initial parameter settings. Standard SSA and genetic algorithm require multiple parameter tuning and trial and error, and are prone to fluctuations when faced with haze risk multivariate optimization; ISSA relies on adaptive parameter control and directional disturbance to significantly reduce computational overhead while ensuring the reliability of the results.

[0053] Preferably, in S5, when the improved sparrow search algorithm is used, the discoverer position is updated as follows:

[0054]

[0055] In the formula, represents the position of the i-th sparrow in the t+1 generation; T max is the maximum number of iterations; a is a uniform random number in the interval (0, 1]; Q is a random number conforming to the standard normal distribution; L represents a matrix; R2 is an alert value; ST is a security value; c is a nonlinear control parameter;

[0056] The calculation formula of the nonlinear control parameter c is:

[0057]

[0058] In the formula, c max , c min is the maximum and minimum value of the parameter c; t is the current number of iterations; δ is a uniform random number between [-1, 1] and the value range of δ is between 0-1.

[0059] Such settings, 1, dynamically coordinate global exploration and local development, break through the convergence stalemate. Through the nonlinear control parameter c(t) (combined with logarithmic decay and exponential regulation) dynamically scaling the discoverer's moving step, enhancing the global exploration ability in the early iteration (t is small), and automatically turning to fine exploitation in the late iteration (t is close to Tmax). The traditional sparrow algorithm (SSA) uses a fixed linear parameter, which is difficult to adapt to the stage characteristics of the optimization process (such as extensive exploration in the early stage and accurate tuning in the late stage), and is easy to fall into local optimum; The c(t) of ISSA balances the search strategy in a mathematically driven way, significantly improving the coverage ability of complex solution space.

[0060] 2, reduce the dependence of algorithm parameter adjustment, improve the engineering practicability. The adaptive design of parameter c(t) (the random item δ enhances the robustness, and c max , c min control boundary) greatly reduces the need for manual parameter adjustment, and is not sensitive to the initial setting. In the prior art, genetic algorithm, particle swarm optimization and other parameters such as crossover rate and inertia weight need to be manually set, and the practice threshold is high; ISSA uses endogenous nonlinear control and environmental response mechanism, so that the algorithm can be plug and play in various optimization scenarios (such as high-dimensional feature selection), reducing the landing cost.

[0061] Preferably, in S5, when the improved sparrow search algorithm is used, the position of the joiner is updated as follows:

[0062]

[0063] In the formula, represents point-to-point multiplication; Levy(λ) represents a path conforming to the Levy distribution, and satisfies 1<λ≤3, λ is the power number; Xi,j(t+1) represents the position of the ith sparrow in the jth dimension at the (t+1)th iteration; n represents the total number of sparrow populations; Xi,j(t+1) represents the position of the ith sparrow in the jth dimension at the (t+1)th iteration; n represents the total number of sparrow populations; Xi,j(t+1) represents the position of the ith sparrow in the jth dimension at the (t+1)th iteration; n represents the total number of sparrow populations; Xi,j(t+1) represents the position of the ith sparrow in the jth dimension at the (t+1)th iteration; n represents the total number of sparrow populations;

[0064] Such a setting enhances the global exploration ability through the Levy flight mechanism and resolves the local optimal trap. When i≤n / 2, the random disturbance of the Levy flight path (Levy(λ)) is introduced, so that the joiner can perform long-distance jumping and short-distance fine search in the solution space, effectively avoiding the algorithm from stagnating in a local suboptimal solution. The traditional sparrow algorithm (SSA) or standard particle swarm optimization (PSO) relies on fixed random walk (such as Gaussian noise), which is easy to fall into early convergence in high-dimensional problems; the Levy flight (satisfying 1<λ≤3) of the present scheme significantly improves the coverage efficiency of the unexplored area, and is especially suitable for the multi-peak optimization challenge (such as feature subset and hyperparameter space) in the haze risk assessment. The creativity lies in the combination of non-Gaussian random process and biological heuristic algorithm, which breaks through the exploration limitation of static disturbance.

[0065] Preferably, when calculating the search path Levy(λ) of the Levy flight, the calculation formula of simulating the Levy flight path is adopted:

[0066] s=μ / |v| 1 / β ;

[0067] In the formula, s is the Levy flight path Levy(λ); β is a preset constant, and β∈(0, 2); the parameters μ and v are normal distribution random numbers, which are subject to the normal distribution shown in the following formula:

[0068]

[0069] And the standard deviation corresponding to the normal distribution satisfies the following formula:

[0070]

[0071] Such a setting strictly approximates the theoretical characteristics of the Levy distribution, and overcomes the mathematical distortion of the traditional random step. Through the Gamma function expression of σ u , the distribution form of the random number is accurately controlled, and it is ensured that the generated step s strictly satisfies the heavy-tailed distribution characteristics (i.e. small step length is dense, and long step length is sparse) of the Levy flight in mathematics. The traditional method (such as simple normal distribution or uniform distribution random step) cannot generate the theoretical Levy path, which leads to the deviation of the search process from the expected long-tail exploration mode; the present scheme constrains the random process by an analytical expression, and for the first time realizes the mathematical verifiability of the step length distribution.

[0072] Preferably, in the sparrow search algorithm, the fitness function f(X i ) is the cross-validation accuracy of the CatBoost model. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to make the purpose, technical solutions and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0074] Figure 1 Flowchart of this method;

[0075] Figure 2 It is a research flow chart in the embodiment;

[0076] Figure 3 Schematic diagram of PCA for improving information entropy of sphere calculation in the embodiment. DETAILED DESCRIPTION

[0077] The following is a further detailed description through specific implementation methods:

[0078] Example:

[0079] like Figure 1 、 Figure 2 As shown, this embodiment discloses a method for assessing urban haze disaster risk based on an information entropy-improved PCA-ISSA-CatBoost model, including the following steps:

[0080] S1. Based on the indicator characteristic system for assessing haze disaster risks, obtain the corresponding sample data set for the target city; each sample in the sample data set includes the corresponding data for each feature in the indicator characteristic system.

[0081] Among them, the first-level indicators of the indicator characteristic system include the danger of disaster-causing factors, sensitivity of disaster-prone environments and disaster-prone bodies; the second-level indicators of the danger of disaster-causing factors include economic density, the number of motor vehicles, the proportion of secondary industry, the area of ​​housing construction and the proportion of coal consumption; the second-level indicators of sensitivity of disaster-prone environments include the annual average concentrations of VOCs, PM2.5, PM10, SO2 and NO2; the second-level indicators of disaster-prone bodies include population density, the number of health institutions and the green area of ​​built-up areas.

[0082] Such, by the first index (disaster factor, disaster environment, disaster body) division, the disaster formation mechanism (risk) is first introduced into the unified system, and the environmental background conditions (sensitivity) and social disaster resistance ability (fragility), breaking through the traditional method only pays attention to the single dimension (such as pollutant concentration) limitation. Compared with the past research only analyzes the haze pollutant itself (such as PM2.5 concentration), the scheme clearly introduces the social and economic driving factors (such as economic density, the proportion of the second industry) and the disaster bearing capacity index (such as the number of health institutions), and more comprehensively reveals the complex causes of haze risk. In addition, the secondary index adopts quantifiable, easy-to-obtain public data (such as "number of motor vehicles", "coal consumption proportion", "built-up area green area"), without relying on professional equipment or high-cost monitoring, which greatly improves the practical operability of the model. Traditional methods often cause index data difficult to obtain (such as real-time monitoring value of specific pollutants) or definition ambiguous (such as "environmental sensitivity" lacks quantitative standard), resulting in evaluation staying at the theoretical level; the scheme selects the commonly used indicators in the statistical yearbook and environmental protection bulletin, ensuring that the evaluation results can be directly used for policy making. Moreover, the secondary index design directly points to the core mechanism of haze disaster, for example, "VOCs, SO2, NO2 concentration" is related to photochemical pollution generation, "population density" and "green area" reflect the disaster exposure and mitigation capacity, so that the model can distinguish the different sources of risk in different cities. The existing index system often has index redundancy (such as the inclusion of irrelevant environmental parameters) or key factor omission (such as ignoring the influence of industrial structure), while the scheme selects the index based on the disaster chain logic (such as economic density→industrial emission→pollution aggravation), which significantly improves the causal explanatory power of the evaluation results.

[0083] According to the regional disaster system theory, the disaster-causing factor, disaster-birth environment and disaster-bearing body are indispensable in the formation process of disaster. The regional disaster system is an abnormal system of the earth's surface composed of the three, and the disaster risk is affected by the common action of the three. In the formation process of disaster, the disaster-causing factor, disaster-birth environment and disaster-bearing body are indispensable. The regional disaster system is an abnormal system of the earth's surface composed of the three, and the disaster risk is affected by the common action of the three. As shown in Table 1, five indexes are selected from human factors to quantitatively evaluate the environmental sensitivity, including economic density representing the degree of economic development; population density representing the degree of urban development; the number of motor vehicles representing the amount of motor vehicle emissions; the proportion of the secondary industry representing the dependence of the regional economy on the secondary industry; coal consumption representing the amount of industrial pollution emissions of a city; and housing construction area representing the housing construction area of a city. The harmfulness of VOCs gathering into haze components refers to the damage of various components of haze to the urban economy and the health of residents. As shown in Table 1, five indexes, including the annual average concentration of PM2.5, PM10, O3, SO2 and NO2, are selected as the evaluation indexes of the harmfulness of VOCs gathering into haze in the Fenwei Plain. In the weather of VOCs gathering, the concentration of aerosols in the air rises, and the atmosphere is relatively stable, which is not conducive to reducing the risk of haze gathering and the vulnerability of the city to haze gathering. Urban greening can absorb harmful gases and dust, reduce air pollution, and reduce the vulnerability of the disaster-bearing body. Therefore, population density, the number of health institutions and the green area of built-up area are selected as the evaluation indexes of the vulnerability of the city to haze gathering.

[0084] Table 1 Haze disaster risk assessment index system

[0085]

[0086] S2, divide the sample data set into a plurality of initial granules by granule calculation, and then gradually refine the granule levels of the initial granules according to the granularity to form a plurality of levels of granule sets; calculate the entropy of each feature of the samples in each level of the granule set as the corresponding weight to obtain a feature weight matrix of each level; for each granule in each level of the granule, perform principal component analysis after weighting using the feature weight matrix of the corresponding level to obtain the corresponding principal component.

[0087] wherein, when dividing the initial granules, for each initial granule G j , the granule center is:

[0088]

[0089] wherein, |. | represents the number of samples in the initial granule G j , x i represents the i th sample in the initial granule G j .

[0090] The radius of the granule is:

[0091] In this way, by quantization design of mean and maximum distance, a high-robustness and self-adaptive solution is provided for initial division of the granule. It not only solves the pain points of strong subjectivity, incomplete coverage and insufficient flexibility in the prior art, but also improves the scientific rigor of the initialization stage through statistical principles.

[0092] In specific implementation, S2 includes:

[0093] For a sample data set X = {x1, x2,..., x n}, where each sample is divided into m initial granules G1, G2,..., G m by granule calculation; and the m initial granules are refined in granule levels to form a plurality of granule sets , where k represents the number of granule levels after the refinement, represents the initial granule G m after the refinement of the kth level of granule;

[0094] For the pth level of the granule set, 1≤p≤k, the entropy of the jth feature of the sample in the set is calculated:

[0095]

[0096] In the formula, p ji represents the normalized probability of feature j in sample i, and n represents the number of samples in the granule set;

[0097] The pth level of the granule set entropy value matrix W = diag(H1, H2,..., H d ) is obtained;

[0098] For the feature matrix Xq of the sample in the qth granule of the pth level, the entropy value matrix W of the layer is used to weight X to obtain the corresponding weighted matrix The weighted matrix is subjected to principal component analysis, and the eigenvectors corresponding to the largest y eigenvalues are selected to form the principal component matrix

[0099] In this way, by calculating the entropy value (H j) and generates an entropy matrix (W) to make the feature importance fully determined by the data distribution. Traditional PCA or fixed weight dimension reduction methods (such as equal weight) are easily disturbed by expert experience or ignore the local importance of features. The present scheme uses the information value of the entropy quantified feature in the hierarchical granule to significantly improve the adaptability and scientificity of weight allocation. In addition, the multi-level refinement of granules (from coarse granularity to fine granularity) allows different levels to focus on different scales of data structure. Entropy weighted PCA is applied to each layer of granules to retain macro distribution rules and extract micro local patterns. In the prior art, global PCA easily loses details (such as sparse area features being submerged) due to data heterogeneity, while hierarchical processing retains global principal components while capturing regional specificity (such as pollution differences between industrial and residential areas) in city haze data, avoiding information loss from "one-size-fits-all" dimension reduction. Moreover, entropy matrix weighting strengthens high information content features and weakens low information content features, allowing subsequent PCA to focus on key dimensions. Filtering the principal components corresponding to the largest eigenvalues further filters redundant noise. Traditional PCA treats features equally, which is easily disturbed by highly correlated features or noise (such as the collinearity of weather and emission sources), while the present scheme improves the discriminant efficiency and robustness of features after dimension reduction through the double mechanism of entropy weighting pre-screening and hierarchical principal component extraction.

[0100] S3, by weighted fusion, fuses the principal components of each granule in each level to obtain global principal components; based on the global principal components, analyzes the load values of each feature in the feature system of each index, and selects a preset number of features to form an initial feature set Fs.

[0101] In specific implementation, S3 includes:

[0102] The principal component matrices of each granule of each layer are weighted and fused to obtain global principal components:

[0103]

[0104] In the formula, α p represents the weight of the pth layer granularity (such as higher weight for coarse granularity), is the weight of the qth granule in the pth layer (such as sample number proportion);

[0105] The global principal components are subjected to dimension reduction analysis, and the b features with the largest load values in each principal component are selected to form an initial feature set Fs.

[0106] In this way, the double-layer weight mechanism (α p adjusts the importance of different granular levels, ​The data density difference of the same layer granulocyte is distinguished, and the dispersed principal component information is fused into a consistent global representation. The traditional principal component fusion method (such as simple splicing or mean pooling) ignores the hierarchy and local difference of the data structure, resulting in the interference between macro trends and micro patterns; the scheme makes the industrial area high-pollution details and the heterogeneous characteristics such as urban background values harmoniously coexist in a unified space through differentiated allocation of weights. In addition, based on the absolute value of the principal component load value (the contribution weight of the feature to the principal component), the key features are screened, so that the construction of Fs is completely driven by data statistical rules. The existing method relies on manual setting of threshold or feature importance ranking of random forest model, which is easy to be disturbed by subjective preset or fall into local optimum; the scheme takes the mathematically interpretable statistical quantity (load value) as the screening standard, eliminates the experience preference from the root, and ensures that the features are strongly related to the risk causes.

[0107] As shown in Figure 3 , wherein 3a shows the schematic diagram of granulocyte calculation. First, the granulocyte center and radius are defined. Second, as shown in Figure 3 b, the granularity hierarchy is divided, and the data set is divided into different granularity hierarchies through granular calculation. Then, entropy calculation is performed, and the information entropy is calculated on each granularity hierarchy to evaluate the distribution of data. Further, PCA dimension reduction is performed, and the principal components are extracted on each granularity hierarchy by PCA. Finally, the principal component selection is performed, and the optimal principal component is selected for dimension reduction in combination with the information entropy analysis.

[0108] As shown in Table 2, in specific implementation, the principal component analysis KMO sample adequacy quantity is 0.764, which is greater than the threshold value 0.5, indicating that there is correlation between the variables, which meets the requirements; the Sig value is 0.000, which is less than 0.05, indicating that the data can be subjected to principal component analysis and has scientificity and reference value. The principal component characteristic values and contribution rates are shown in Table 3. According to the principle that the characteristic value is greater than 1, the first three items are selected as the principal components, and the variance contribution rates of the three items are 40.518%, 20.501% and 16.604% respectively, and the cumulative contribution rate of the three items is 77.623%, which can basically reflect the information of the original indexes, so the first three items are taken as the principal component factors and represented by F1, F2 and F3. In this way, the original 12 indexes are replaced by the first three principal components, and the load condition of each factor to the original index can be calculated.

[0109] Table 2 KMO test and Bartlett's test

[0110]

[0111] Note: ***, **, * represent 1%, 5%, 10% significance levels respectively

[0112] Table 3 Principal component characteristic values and contribution rates

[0113]

[0114] From Figure 3 c, the F1 eigenvalue is 4.498, the contribution rate is 37.598%, and it is the first driving force of the formation of haze risk in the target city. As shown in Figure 3 c, the analysis of the load value of the principal component F1 found that the first principal component F1 was larger than 0.83 in the values of NO2, SO2, PM2.5, and PM10, which means that NO2, SO2, PM2.5, and PM10 are the first constituent factors of haze risk. The top four indexes of the load value of the principal component F2 are NHI, NMV, ED, and VOCs, which shows that the target city's haze hazard is mainly affected by the number of motor vehicles, the number of health institutions, the green area of built-up area, and the concentration of VOCs. The index with the highest load value of F3 is X6 (annual average concentration of PM10), followed by the proportion of the second industry, which reflects that PM10 and the second industry are also important environmental factors for the formation of haze risk. Figure 3 d is the factor loading quadrant analysis scatter plot.

[0115] Haze disaster risk principal component comprehensive score and ranking

[0116] Load represents the degree of linear relationship between each original variable and the formation of haze risk, which can be used to explain the meaning of principal component. By calculating the load matrix, the correlation between each principal component and the original index can be understood, and the meaning of the principal component and its explanatory ability to the original index can be further understood. Therefore, the index weight is expressed as the exponential load. The haze risk on each principal component is calculated as follows:

[0117]

[0118] Where: H ny is the risk value on the principal component, n is the unit number, m is the number of indexes, and the unit H 1y represents the haze risk index of the evaluation unit on the first principal component F1; Weight m is the index value; is the standardized value.

[0119] When using principal components for analysis, each principal component represents a direction in the original data, and its importance is represented by the cumulative contribution rate. The cumulative contribution rate is the sum of the proportion of the total variance occupied by the first n principal components. This means that the principal component with a higher contribution rate can retain more data information, and the haze comprehensive risk index is:

[0120]

[0121] In the formula: H i(i = 1, 2, 3) represent the operation scores of each principal component, respectively, P i (i = 1, 2, 3) represent the contribution rate of each principal component.

[0122] S4, constructing a CatBoost model for haze disaster risk assessment according to the feature set data of haze;

[0123] Catboost is an advanced gradient learning algorithm based on the GB-DT algorithm framework, and has better ability to process categorical features; Catboost uses more effective strategies to reduce overfitting, uses combined categorical features, greatly enriches the feature dimension, and uses the entire data set for training, CatBoost algorithm can well solve various data bias problems in the original GBDT, and has good robustness.

[0124] CatBoost has the flexibility to provide categorical column indices, which can be encoded as one-hot encoding using one_hot_max_size. Suppose the observation data set D = (X K ,y k )k = 1, 2, …, n, where is a vector containing m features, y k ∈R is the value of the label. An effective way to handle categorical features is to use a calculated value (TS: Target Statistic) instead of (i represents the i-th sub-type feature of the k-th training sample), first randomly sort the set of input observations to generate multiple random orders. Calculate one of the label value orders placed in the same category before a given value, and the calculation of this value TS is as follows: Because some categories appear less frequently, smoothing is needed to convert all categorical feature values to numerical values:

[0125]

[0126] where, is an indicator function: if [x jk = x ik ], the function value is 1, otherwise the function value is 0; p is the prior value, and the parameter α > 0 is the weight of the prior, Adding a prior value helps to reduce the noise obtained from low-frequency categories.

[0127] CatBoost uses the order principle, and the calculation of the TS value relies on the sample set that has been observed so far. Based on the greedy algorithm to select the structure of the tree, find all possible ways of splitting, calculate the penalty function of each way, select the smallest one, assign the result to the leaf node, and repeat this process for the subsequent leaf nodes. Before building a new tree, perform a random rearrangement, build a new tree in the direction of gradient descent, and CatBoost uses different arrangements in different gradient boosting steps.

[0128] S5, initializing the feature subset of the initial feature set Fs and the hyperparameters of CatBoost as sparrow individuals X i , and defining a fitness function f(X i ) based on the CatBoost model to measure the pros and cons of individuals, searching for the optimal feature subset F opt of the initial feature set Fs and the optimal hyperparameters θ opt of the CatBoost model.

[0129] Sparrow search algorithm is a new type of heuristic swarm intelligence algorithm. Through simulating the behavior of sparrows foraging and avoiding predators, the optimization search of model parameters is realized. Among them, the fitness function f(X i ) is the cross-validation accuracy of the CatBoost model. In specific implementation, other fitness functions can also be set by those skilled in the art, as long as they meet the purpose of the method, which will not be described here.

[0130] In specific implementation, the sparrow search algorithm is an optimized sparrow search algorithm, wherein the improvement includes: when updating the position of the discoverer, a nonlinear control parameter is introduced to coordinate the global search and local development ability of the algorithm; when updating the position of the joiner, the Levy flight strategy is introduced to disturb the current optimal solution and enhance the local exploitation ability. In this way, the nonlinear control parameter dynamically adjusts the exploration step of the discoverer (such as large-scale global exploration in the early stage and fine local development in the later stage), avoiding the algorithm from falling into premature convergence or ineffective oscillation. The original sparrow search algorithm (SSA) uses a fixed linear parameter, which is easy to miss the global optimal solution in complex optimization problems (such as converging to a local suboptimal feature subset of haze risk assessment); ISSA (the algorithm) significantly improves the solution space coverage ability in high-dimensional and multi-peak scenarios through an adaptive nonlinear mechanism. The Levy flight strategy gives the joiner the ability to jump long distances and search fine distances, and through intelligent disturbance of the current optimal solution, it breaks through the local extreme value trap. Traditional random disturbance (such as Gaussian noise) is easy to cause the optimization direction to deviate or converge slowly, while the heavy-tailed distribution characteristics of Levy flight (short-distance intensive search + long-distance jump) balance development and exploration, and accurately locate the optimal configuration in the synchronous optimization of CatBoost hyperparameters and feature subsets. The double-strategy cooperation greatly reduces the number of iterations required for convergence (such as quickly locking the optimal feature subset F optand the optimal hyperparameter θ opt ), while reducing the sensitivity to initial parameter settings. Standard SSA and genetic algorithms require multiple parameter tuning and trial-and-error, and are prone to fluctuations when faced with multi-variable optimization of haze risk; ISSA relies on adaptive parameter control and directional disturbance to significantly reduce computational overhead while ensuring the reliability of the results.

[0131] where, when using the improved sparrow search algorithm, the position of the discoverer is updated as follows:

[0132]

[0133] where, represents the position of the ith sparrow in the t+1 generation; T max is the maximum number of iterations; α is a uniform random number in the interval (0, 1]; Q is a random number conforming to the standard normal distribution; L represents the matrix; R2 is the alert value; ST is the safety value; c is the nonlinear control parameter;

[0134] The calculation formula of the nonlinear control parameter c is:

[0135]

[0136] where, c max , c min is the maximum and minimum value of the parameter c; t is the current iteration number; δ is a uniform distribution random number between [-1, 1] and the value range of δ is between 0-1.

[0137] In this way, the nonlinear control parameter c(t) (combined with logarithmic decay and exponential regulation) dynamically scales the discoverer's moving step, enhancing global exploration ability in the early iteration (when t is small) and automatically switching to fine exploitation in the late iteration (when t is close to Tmax). The traditional sparrow algorithm (SSA) uses a fixed linear parameter, which is difficult to adapt to the stage characteristics of the optimization process (such as extensive exploration in the early stage and accurate tuning in the late stage), and is prone to fall into local optimum; the c(t) of ISSA autonomously balances the search strategy in a mathematically driven manner, significantly improving the coverage ability of complex solution space. In addition, the adaptive design of the parameter c(t) (the δ random term enhances robustness, and the c max , c min control boundary) greatly reduces the need for manual parameter tuning and is not sensitive to initial settings. In existing technologies, genetic algorithms, particle swarm optimization, etc. require manual setting of parameters such as crossover rate and inertia weight, which has a high practical threshold; ISSA uses endogenous nonlinear control and environmental response mechanisms, making the algorithm plug-and-play in various optimization scenarios (such as high-dimensional feature selection), reducing the landing cost.

[0138] When using the improved sparrow search algorithm, the position of the joiner is updated as follows:

[0139]

[0140] wherein, denotes point-to-point multiplication; Levy(λ) denotes a path subject to a Levy distribution, and satisfies 1 < λ ≤ 3, λ is a power number; denotes the position of the ith sparrow in the jth dimension in the t+1th iteration; n denotes the total number of sparrow populations; denotes the worst position; denotes the optimal position; denotes the optimal position of the sparrow in the dth dimension in the t+1th iteration of the population.

[0141] In calculating the search path Levy(λ) of the Levy flight, the calculation formula of the Levy flight path is used:

[0142] s = μ / |v| 1 / β ;

[0143] wherein, s is the Levy flight path Levy(λ); β is a preset constant, and β ∈ (0, 2); the parameters μ and v are normal distribution random numbers, subject to a normal distribution shown in the following formula:

[0144]

[0145] and the standard deviation corresponding to the normal distribution takes a value satisfying the following formula:

[0146]

[0147] In this way, the global exploration ability is enhanced by the Levy flight mechanism, and the local optimal trap is resolved. When i ≤ n / 2, the random disturbance of the Levy flight path (Levy(λ)) is introduced, so that the joiner can perform long-distance jumping and short-distance fine search in the solution space, and effectively avoid the algorithm from stagnating in a local suboptimal solution. The traditional sparrow algorithm (SSA) or the standard particle swarm optimization (PSO) relies on fixed random walk (such as Gaussian noise), and is easy to fall into early convergence in high-dimensional problems; the Levy flight (satisfying 1 < λ ≤ 3) of the present scheme significantly improves the coverage efficiency of the unexplored area, and is especially suitable for the multi-peak optimization challenge (such as the feature subset and the hyperparameter space) in the haze risk assessment. The creativity lies in the combination of the non-Gaussian random process and the bio-inspired algorithm, which breaks through the exploration limitation of static disturbance. Moreover, the present method strictly approximates the theoretical characteristics of the Levy distribution, and overcomes the mathematical distortion of the traditional random step. Through σ uThe Gamma function expression accurately controls the random number distribution form, and ensures that the generated step s strictly satisfies the heavy-tailed distribution characteristics of the Levy flight in mathematics (i.e. small step length is dense, and long step length is sparse). The traditional method (such as simple normal distribution or uniform distribution random step) cannot generate a theoretical Levy path, resulting in deviation of the search process from the expected long-tail exploration mode; the scheme constrains the random process with an analytical expression, and for the first time realizes the mathematical verifiability of the step length distribution.

[0148] The scout position is updated as:

[0149]

[0150] In the formula: β is a step length control parameter; K is a random number in [-1, 1], representing the moving direction of the sparrow; f i and f W and f g The distribution represents the fitness, the worst fitness and the best fitness of the current sparrow.

[0151] In order to better understand the effect of the method, the following is explained.

[0152] In the iterative optimization process, the search stability of the ISSA algorithm (i.e. the improved sparrow search algorithm in the method) is high, and the optimal hyperparameter combination of the CatBoost model has been searched in about 300 iterations. The SSA algorithm has large fluctuation in the iterative optimization process, and still does not appear to converge in 600 iterations, and the fitness is further widened from the ISSA algorithm. Under the given termination iteration condition, the model parameters obtained by using ISSA optimization are Learing_rate=0.25, gamma=0.13, max_depth=7, min_child_weight=3, and lambda=1.

[0153] Among them, the model parameter optimization adopts 10-fold cross-validation to optimize the parameters, and the grid search is performed on the main 3 parameters of the CatBoost model, which are the number of weak classifiers, the maximum depth of the decision tree and the learning rate. The optimal parameters of 6 models and the training set accuracy are shown in Table 4.

[0154] Table 4 Optimal parameters of each model and training set accuracy

[0155]

[0156] In order to evaluate the accuracy of the model, four evaluation indexes are selected, including accuracy (Accuracy), precision (Precision), recall (Recall) and F value (F Measure ).

[0157] Accuracy=TP +T N / T P +T N +F P +F N ;

[0158] Precision=T P / T P +E P ;

[0159] Recall=T P / T P +F N ;

[0160]

[0161] In the formula: T P is a true example (a sample correctly predicted as a positive class), T N represents a true negative example (a data point classified as a negative sample), F P is a false positive example (a negative example incorrectly divided into a positive example), and F N is a false negative example (a sample incorrectly predicted as a negative example). The accuracy represents the overall accuracy of the prediction, and the precision and recall focus on the proportion of samples that the model predicts as positive and that are actually positive, respectively. The F value considers both precision and recall and is a comprehensive evaluation index, commonly used in unbalanced data sets or situations where precision and recall need to be balanced.

[0162] S6, using the optimal feature subset F opt corresponding feature set data and the optimal hyperparameters θ opt , train the CatBoost model;

[0163] S7, obtain the optimal feature subset F opt corresponding feature set data in the actual data of the target city, and use the trained CatBoost model to perform haze disaster risk assessment.

[0164] Compared with the prior art, the weight of each level feature is dynamically calculated by information entropy weighted PCA and granulometry, and key features are screened based on this. Compared with the index system evaluation method in the prior art (such as expert weighting or analytic hierarchy process), this method avoids the subjective bias caused by artificial weight setting (such as misjudgment of the importance of features), and significantly improves the objectivity of feature screening. This ensures that the risk assessment model can automatically extract core disaster-causing factors from multiple factors (such as weather and pollutant concentration), improving the accuracy of the assessment. In addition, the algorithm optimizes the feature subset and CatBoost hyperparameters (such as learning rate or tree depth) at the same time, and uses the model performance as the fitness function to drive the search process. Compared with the step-by-step optimization method in the prior art (such as manually screening features first and then adjusting parameters, or relying on grid search), this scheme completes double-objective optimization in one iteration process, reducing the accumulation of intermediate errors in the optimization phase. This greatly improves the model training efficiency and overall stability, effectively alleviating the overfitting or local optimum problem when facing high-dimensional data. Moreover, combining the CatBoost model with the information entropy improved feature dimension reduction, the high-dimensional and heterogeneous data (such as missing values or noise) are automatically processed in the training phase, and the core information is preserved through global principal component. Compared with the historical disaster mathematical statistics method, this method reduces the excessive dependence of the model on new disaster data (such as the need to continuously update historical samples to maintain accuracy), improving the reliability of the model in the case of sparse or incomplete data. This makes the risk assessment more suitable for the dynamic changes of actual urban haze problems.

[0165] The model was verified by the haze process data of multiple cities from November to December 2023, and the results showed that the model could effectively evaluate the risk of the population affected by major VOCs disaster events, the risk of transportation affected by disasters, the risk of crop area affected by disasters, the risk of direct economic loss, and the comprehensive disaster risk before the occurrence of the disaster. This method has important application value in the risk assessment and management of haze disasters. The main conclusions of the study are as follows:

[0166] (1) The haze disaster risk assessment model based on granulometry, information entropy PCA, ISSA and CatBoost algorithm shows good applicability. In actual case testing, the model performs well in evaluating the affected population, traffic impact, crop loss area, direct economic loss and comprehensive risk, and the key performance indicators such as accuracy, precision, recall and F1 value all reach more than 82%, proving its effectiveness and generalization ability in actual disaster risk assessment.

[0167] (2) The model of the method quantifies the contribution of each feature to the evaluation result through the importance evaluation index, revealing the characteristics of different risk assessment objectives being affected by different factors. For example, the risk of affected population is mainly affected by the characteristics of the disaster-bearing body, while the risks of traffic, crop loss, and economic loss are more affected by the disaster environment. Economic density is the main driving force of comprehensive risk. Feature selection and weight allocation using particle swarm calculation enhance the interpretability of the model, helping to understand the decision-making process of the machine learning model and revealing the internal relationship between each index and the evaluation result.

[0168] (3) The ensemble learning algorithm reduces the dependence on disaster mechanisms and simplifies the modeling process by processing a large amount of related data. The accuracy of the evaluation result depends largely on the diversity of the indicators and the sufficiency of the sample size. Studies have shown that considering a comprehensive set of indicators of disaster-causing factors, disaster-prone environment, and disaster-bearing body can significantly improve the accuracy of the assessment by 10% to 15% compared to using only disaster-causing factors.

[0169] (4) The haze disaster risk assessment model constructed by the particle swarm calculation-information entropy PCA-ISSA-CatBoost algorithm can provide more accurate and comprehensive risk assessment. Combined with the disaster system theory, the complexity of the disaster system and the mutual relationship between factors can be more comprehensively considered. In the modeling process, the use of comprehensive learning algorithm for in-depth analysis of historical disaster data can more accurately assess the risk of haze disasters. This method, which integrates comprehensive learning algorithm, disaster system theory, and particle swarm calculation technology, provides a more comprehensive perspective and more scientific basis for risk assessment.

[0170] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and not to limit the technical solutions. Those of ordinary skill in the art should understand that modifications or equivalent replacements to the technical solutions of the present application without departing from the spirit and scope of the technical solutions should be covered within the scope of the claims of the present application.

Claims

1. An urban haze disaster risk assessment method based on information entropy-improved PCA-ISSA-CatBoost model is characterized by: The following steps are involved: S1. Based on the indicator feature system for assessing haze disaster risk, obtain a sample data set corresponding to the target city; each sample in the sample data set includes the corresponding data for each feature in the indicator feature system; S2. Divide the sample data set into multiple initial spheres through sphere calculation, then refine the initial spheres into sphere levels according to granularity to form sphere sets at multiple levels; calculate the entropy of each feature of the sample in each level of sphere set as the corresponding weight to obtain the feature weight matrix of each level; perform principal component analysis on each sphere in each level of sphere set after weighting it using the feature weight matrix of the corresponding level to obtain the corresponding principal component; S3, by means of weighted fusion, the principal components of each sphere in each level are fused to obtain the global principal component; Based on the global principal component analysis, the load value of each feature in each indicator feature system is analyzed, and a preset number of features are screened out to form the initial feature set Fs; S4. Construct a CatBoost model to assess the risk of haze disasters based on haze feature set data; S5, use the sparrow search algorithm to initialize the feature subset of the initial feature set Fs and the hyper parameters of CatBoost to the sparrow individual X i , and define the fitness function f(X i ) measure the quality of individuals and search for the optimal feature subset F of the initial feature set Fs opt And the optimal hyperparameter θ of the CatBoost model opt ; S6. Use the optimal feature subset F opt Corresponding feature set data and optimal hyperparameters θ opt , train the CatBoost model; S7. Obtain the optimal feature subset F in the actual data of the target city. opt The corresponding feature set data is obtained and the trained CatBoost model is used to assess the risk of haze disasters.

2. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 1, characterized in that: In S1, the first-level indicators of the indicator characteristic system include the risk of disaster-causing factors, sensitivity of disaster-prone environments and disaster-prone bodies; the second-level indicators of the risk of disaster-causing factors include economic density, the number of motor vehicles, the proportion of secondary industry, the area of ​​housing construction and the proportion of coal consumption; the second-level indicators of sensitivity of disaster-prone environments include the annual average concentrations of VOCs, PM2.5, PM10, SO2 and NO2; the second-level indicators of disaster-prone bodies include population density, the number of health institutions and the green area of ​​built-up areas.

3. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 1, characterized in that: In S2, when dividing the initial spheres, for each initial sphere G j , the center of the sphere is: Among them, |.| represents the initial particle sphere G j The number of samples in x i Represents the initial ball G j The i-th sample in ; The radius of the sphere is:

4. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model as claimed in claim 1, characterized in that: S2 include: For the sample data set X={x1,x2,...,x n }, where each sample Through granular calculation, it is divided into m initial granular balls G1, G2, ..., G m ; The m initial spheres are gradually refined into sphere levels according to the granularity to form multiple sphere sets Among them, k represents the number of granular layers after gradual refinement. Represents the initial ball G m The granular ball of the kth level after gradual refinement; For the p-th layer of spheres, 1≤p≤k, calculate the entropy of the j-th feature of the sample in the set: Where p ji represents the normalized probability of feature j in sample i, and n represents the number of samples in the sphere set; Get the entropy matrix of the particle set of the pth layer W=diag(H1,H2,...,H d ); For the characteristic matrix of the sample in the qth sphere in the pth layer 1≤q≤mq, use the entropy matrix W of this layer to weight the X and get the corresponding weighted matrix Then add the weight matrix Perform principal component analysis and select the eigenvectors corresponding to the largest y eigenvalues ​​to form the principal component matrix after dimensionality reduction 5. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 4 is characterized in that: S3 includes: The principal component matrices of each sphere in each layer are weighted and fused to obtain the global principal component: Where, α p represents the weight of the p-th layer granularity, is the qth ball in the pth layer The weight of Perform dimensionality reduction analysis on the global principal components and select b features with the largest load value in each principal component to form the initial feature set Fs.

6. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 1, characterized in that: In S5, the sparrow search algorithm used is an optimized sparrow search algorithm, wherein the improvements include: when updating the discoverer position, introducing nonlinear control parameters to coordinate the global search and local development capabilities of the algorithm; when updating the joiner position, introducing the Levy flight strategy to perturb the current optimal solution and enhance the local mining capability.

7. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 6, characterized in that: In S5, when using the improved sparrow search algorithm, the finder position is updated as follows: Where, represents the position of the i-th sparrow in generation t+1; T max is the maximum number of iterations; α is a uniform random number in the interval (0,1]; Q is a random number that conforms to the standard normal distribution; L represents the matrix; R2 is the warning value; ST is the safety value; c is the nonlinear control parameter; The calculation formula of nonlinear control parameter c is: Where c max 、c min are the maximum and minimum values ​​of parameter c; t is the current iteration number; δ is a random number uniformly distributed between [-1, 1] and the value range of δ is between 0 and 1.

8. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 7, characterized in that: In S5, when using the improved sparrow search algorithm, the position of the joiner is updated as follows: Where, represents point-to-point multiplication; Levy(λ)) represents a path that obeys the Levy distribution and satisfies 1<λ≤3, where λ is the power; represents the joiner position of the i-th sparrow in the j-th dimension in the t+1-th iteration; n represents the total number of sparrows in the population; Indicates the worst position; Indicates the optimal position; It represents the optimal position of the sparrow in the d dimension at the t+1th iteration of the population.

9. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 8, characterized in that: When calculating the search path Levy(λ) of the Levy flight, the calculation formula for simulating the Levy flight path is used: s=μ / |v| 1 / β ; Where s is the Levy flight path Levy(λ); β is a preset constant, and β∈(0,2); the parameters μ and v are normally distributed random numbers, obeying the normal distribution shown in the following formula: And the standard deviation of the corresponding normal distribution satisfies the following formula:

10. The urban haze disaster risk assessment method based on the information entropy improved PCA-ISSA-CatBoost model according to claim 1, characterized in that: In the sparrow search algorithm, the fitness function f(X i ) is the cross-validation accuracy of the CatBoost model.

Images