Method for constructing enterprise pollution state model based on spatial driving factor analysis

Abstract

The embodiment of the application discloses a kind of enterprise pollution state model construction methods based on spatial driving factor analysis, comprising: obtaining the geographical position and first pollution state index of each enterprise in the region to be analyzed;According to the geographical position of each enterprise, weight matrix is constructed;Eigenvalue decomposition is carried out to weight matrix, and at least one eigenvector is extracted;With each eigenvector as independent variable, with the first pollution state index and each first non-space characteristic of each enterprise as dependent variable, linear regression is carried out, and the second pollution state index after each enterprise is stripped global spatial effect and each second non-space characteristic is obtained;With each second non-space characteristic as independent variable, with second pollution state index as dependent variable, enterprise pollution state model is fitted;Through spatial interpretation analysis, the pollution state model is decomposed into spatial background effect and non-space background effect.This embodiment accurately analyzes the local area characteristics of pollution formation, and clearly defines the driving mechanism of enterprise pollution.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of environmental data analysis and environmental pollution assessment technology, and in particular to a method for constructing an enterprise pollution status model based on spatial driving factor analysis. Background Technology

[0002] Environmental pollution assessment is a process that uses scientific methods to qualitatively and quantitatively analyze the properties, concentrations, or impacts of pollutants on the environment and human health. Through environmental pollution assessment, pollution status and risks can be identified, providing data support for decision-making regarding environmental governance and ecological restoration.

[0003] Current environmental pollution assessment methods include geostatistical methods and global machine learning methods.

[0004] Geostatistical methods are commonly used pollution assessment techniques in environmental science, primarily employing techniques such as semivariogram analysis and kriging interpolation to study the spatial distribution characteristics of pollution. These methods are based on the theory of regionalized variables, assuming that pollution phenomena exhibit spatial autocorrelation, meaning that geographically proximate sample points often share similar pollution characteristics. In practical applications, pollution monitoring data from enterprise sites are typically collected first, then the semivariogram is calculated to quantify spatial autocorrelation, and finally, interpolation methods such as ordinary kriging, pan-kriging, or co-kriging are used to generate pollution distribution maps. For example, in assessing heavy metal pollution in soil, the heavy metal concentration data from sampling points, combined with covariates such as topography and soil type, are used to predict the pollution distribution across the entire region using pan-kriging methods. These methods can reflect the spatial continuity of pollution to some extent, providing spatial distribution references for environmental management.

[0005] With the development of artificial intelligence technology, global machine learning methods are increasingly being applied to enterprise pollution assessment. These methods primarily employ algorithms such as random forests, support vector machines, and neural networks, using enterprise characteristics (e.g., industry category, production scale, process type) as input features and pollution status as the prediction target to train a predictive model. In typical applications, a large amount of enterprise attribute data and pollution monitoring data is first collected. After data preprocessing and feature engineering, the dataset is randomly divided into training and test sets. Then, a machine learning model is trained and its performance is evaluated. These methods can effectively capture the non-linear relationship between enterprise characteristics and pollution status, identify key pollution driving factors, and demonstrate good performance in pollution prediction.

[0006] However, traditional methods often employ global statistical models or ordinary machine learning algorithms, neglecting geospatial autocorrelation and spatial heterogeneity, resulting in models failing to accurately capture the regional characteristics of pollution formation; and failing to clarify the driving mechanisms of corporate pollution. Summary of the Invention

[0007] This invention provides a method for constructing an enterprise pollution status model based on spatial driving factor analysis to solve at least one of the above problems.

[0008] In a first aspect, embodiments of the present invention provide a method for constructing an enterprise pollution state model based on spatial driving factor analysis, including: Obtain the geographical location and primary pollution status indicators of each enterprise within the area to be analyzed; Based on the geographical location of each enterprise, a weight matrix is ​​constructed to represent the spatial relationship between enterprises; and the weight matrix is ​​decomposed into eigenvalues ​​to extract at least one feature vector to represent the global spatial effect within the region. Using each feature vector as the independent variable and the first pollution state index and each first non-spatial feature of each enterprise as the dependent variable, linear regression was performed to obtain the second pollution state index and each second non-spatial feature of each enterprise after stripping the global spatial effect. Using each of the second non-spatial features as independent variables and the second pollution state index as the dependent variable, a pollution state model for enterprises is fitted. Through Geoshapley spatial interpretation analysis, the pollution state model is decomposed into spatial background effects and non-spatial background effects to obtain the final pollution state model.

[0009] In a second aspect, embodiments of the present invention provide an electronic device, the electronic device comprising: One or more processors; Memory, used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the enterprise pollution state model construction method based on spatial driving factor analysis as described in any embodiment.

[0010] Thirdly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the enterprise pollution state model construction method based on spatial driving factor analysis as described in any embodiment.

[0011] In summary, this embodiment provides a method for constructing an enterprise pollution status model based on spatial driving factor analysis. It fully considers the geospatial autocorrelation and spatial heterogeneity of enterprise pollution, and can accurately analyze the regional characteristics of pollution formation. Specifically, this method can achieve the following beneficial effects: 1. Spatial patterns are extracted and identified through MEM, and then spatial influences of pollution status indicators and non-spatial features are removed through FWL analysis to generate clean data without significant spatial autocorrelation, thus solving the core problem of spatial pseudo-correlation interfering with the judgment of non-spatial features in traditional models.

[0012] 2. Multi-dimensional analysis: By integrating the global net effect of FWL coefficients, AutoML feature importance values, and GeoShapley values ​​(spatial heterogeneity effects), a three-dimensional effect system of "global-prediction-regional" is constructed. First, the global spatial effect is removed to avoid the global spatial effect from masking the influence of other driving factors when it is too significant. Based on the data after removal, the system can accurately distinguish whether the enterprise's pollution is due to its own characteristics (such as industry type and production duration) or the influence of spatial environmental factors (such as regional ecological environment). This breaks through the limitations of traditional single-dimensional analysis and further clarifies the driving mechanism of enterprise pollution.

[0013] 3. Differentiating the role and interaction of space with other factors: Geo-Shapley analysis of spatial interaction effects can guide regional governance (such as developing differentiated measures for areas with high interaction effects); at the same time, based on the spatial variation coefficient of GeoShapley, the impact differences of driving factors (such as precipitation and wind speed) in different regions can be clarified, providing regionally differentiated management suggestions for environmental pollution improvement; spatial driving factor heat maps can guide the allocation of governance resources to high-impact areas, improving resource utilization efficiency. Attached Figure Description

[0014] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a method for constructing an enterprise pollution status model based on spatial driving factor analysis, provided by an embodiment of the present invention. Figure 2 This is a visualization result output by GeoShapley analysis provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0017] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0018] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0019] Figure 1 This is a flowchart illustrating a method for constructing an enterprise pollution status model based on spatial driving factor analysis, provided by an embodiment of the present invention. This method is suitable for modeling and analyzing the pollution status of enterprises across regions and is executed by electronic equipment. Figure 1 As shown, the method specifically includes: S110. Obtain the geographical location and pollution status indicators of each enterprise within the area to be analyzed.

[0020] The region to be analyzed here refers to a large area spanning multiple regions, such as the entire country. The scale of each region is not limited and can be a city, province, or even a larger area. This step first collects enterprise point data and environmental data within this large area as the data source for the entire method.

[0021] In one specific implementation, enterprise point data may include the enterprise's latitude and longitude coordinates, industry category, and pollution status indicators (such as pollutant content, whether pollution exceeds standards, etc.), start of production, end of production, and production duration. Multi-source environmental raster data may include ecological environment quality data and natural environment data. Among them, ecological environment quality data may adopt the CHEQ-500 dataset, and natural environment data may include, for example, annual precipitation data, annual average wind speed data, annual average ground temperature data, and annual evaporation data.

[0022] After collecting the raw data, the consistency of the spatial data coordinates and range is first checked, and the data is resampled to ensure consistent raster accuracy. Specifically, the company's Excel data is read and converted into GeoDataFrame objects. Then, based on the company's latitude and longitude coordinates, the raster values ​​of the corresponding locations are extracted from each environmental raster data. One-hot encoding is performed on the industry categories to generate an industry feature matrix. Continuous variables are standardized, missing values ​​are handled, and spatial interpolation methods are used to fill in the null values ​​in the environmental raster data.

[0023] S120. Based on the geographical location of each enterprise, construct a weight matrix to represent the spatial relationship between enterprises.

[0024] This step constructs a spatial weight matrix, with each enterprise as a row and column. Assume... To represent the number of effective firms, the spatial weight matrix is: The matrix represents the strength of the spatial relationship between companies in a row and companies in a column.

[0025] In one specific implementation, the spatial weight matrix can be calculated based on the K-nearest neighbor method. Specifically, for each enterprise, the spatial weights between the current enterprise and its K nearest neighbor enterprises are calculated, and these spatial weights together form the spatial weight matrix. Here, K is a natural number that can take a value between 10 and 20. Optionally, spatial weights considering geographical distance are as follows:

[0026] in, Indicates enterprise With enterprises The spatial weights between, i.e., the first space weight matrix. Line 1 Column elements; Indicates enterprise With enterprises The geographical distance between them (calculated using latitude and longitude coordinates). The closer the geographical distance, the greater the distance between them. The larger the value, the more it conforms to the spatial correlation law of "near is strong and far is weak". This weight initially quantifies the spatial correlation between enterprises from the perspective of geographical distance.

[0027] The above method optimizes the sparsity of the spatial weight matrix using the K-nearest neighbor method. That is, for each enterprise point, only the weights of the K nearest neighbors are retained, and the weights of the rest are set to 0, thus transforming the spatial weight matrix into a sparse matrix and significantly improving the computational efficiency of subsequent operations.

[0028] In another specific embodiment, the environmental data acquired in S110, in addition to the aforementioned ecological environment data and natural environment data, may also include enterprise environmental data. Enterprise environmental data is used to reflect the enterprise's development environment, such as GDP per capita. Correspondingly, the construction of the spatial weight matrix may also include the following steps: Step 1: Based on the ecological quality and environmental data of each enterprise's geographical location and the enterprise's environmental data, calculate the coupling coordination degree for each enterprise. The coupling coordination degree characterizes the degree of compatibility between enterprise development and the ecological environment. Optionally, its calculation method is as follows:

[0029] Wherein, U1 represents the comprehensive index of the ecological environment subsystem composed of ecological environment quality data, and U2 represents the comprehensive index of the enterprise environmental subsystem composed of enterprise environmental data. Each indicator in the ecological environment quality data is an indicator of the ecological subsystem, and each indicator in the enterprise environmental data (such as GDP per capita) is an indicator of the enterprise environmental subsystem.

[0030] Step 2: Adjust the weight matrix based on the differences in coupling coordination between enterprises to obtain the final weight matrix. This embodiment introduces differences in coupling coordination to adjust the geographical proximity weights; the formula is:

[0031] in, Indicates enterprise With enterprises Inter-adjusted weights Indicates enterprise The degree of coupling coordination Indicates enterprise The degree of coupling coordination This is used to characterize the difference in the degree of adaptation of the two enterprises to the regional environment; This is a correction factor for the difference in coupling degree, with a value ranging from 0.3 to 0.6. The larger the value, the stronger the impact of coupling differences on spatial correlation.

[0032] Then, the corrected weight matrix is ​​standardized to ensure that the sum of the weights in each row is 1, which facilitates the interpretation of "the proportion of a company's spatial influence from other companies". The formula is as follows:

[0033] in, The standardized spatial weights are determined by each The matrix formed That is the final spatial weight matrix.

[0034] Of course, the above standardization steps can also be performed on a spatial weight matrix that only considers geographical distance. Furthermore, the subsequent steps of the entire method are applicable to both the uncorrected and corrected spatial weight matrices, with the latter being superior.

[0035] S130. Perform eigenvalue decomposition on the weight matrix to extract at least one eigenvector characterizing the global spatial effect within the region.

[0036] Let the spatial weight matrix to be processed be denoted as... This step involves eigenvalue decomposition of the matrix to extract globally fixed patterns representing spatial relationships. (Spatial feature vector).

[0037] In one specific implementation, the Moran matrix is ​​first constructed. Specifically, a centralization operator is used to eliminate the interference of global trends on spatial correlations, retaining only the relative correlations between regions, as shown in the formula:

[0038] in, Represents the Moran matrix ( Matrix), matrix elements Indicates enterprise and The relative spatial correlation strength; Represents the centralization operator, ,in for The identity matrix; 1 is An all-1 vector is used to remove the global mean and focus on the relative differences in the region.

[0039] Then, eigenvalue decomposition is performed on the Moran matrix M to extract MEM, as shown in the formula:

[0040] in, express ( (A vector, where each element corresponds to a spatial feature of an enterprise). for The corresponding eigenvalues. Each This represents a unique spatial pattern, such as "high in the east and low in the west," "high in the center and low in the periphery," or "dispersed distribution." The larger the absolute value of the eigenvalue, the stronger the corresponding spatial pattern. The stronger its spatial explanatory power.

[0041] S140. Using each feature vector as the independent variable and each enterprise's pollution status index and each non-spatial feature as the dependent variable, perform linear regression to obtain each enterprise's pollution status index and each non-spatial feature after removing the global spatial effect.

[0042] The non-spatial characteristics of enterprises include coded industry type, standardized precipitation (i.e., precipitation in the grid where the enterprise is located), production duration, and output value. This embodiment aims to model the relationship between enterprise pollution status indicators and non-spatial characteristics. This step is used to address the global spatial effect among these variables, preventing the overemphasis on the enterprise's own characteristics and local spatial characteristics when the global spatial effect is too significant. For ease of distinction and description, this embodiment refers to the pollution status indicators before and after removing the global spatial effect as the first pollution status indicator and the second pollution status indicator, respectively, and the non-spatial characteristics before and after removing the global spatial effect as the first non-spatial characteristic and the second non-spatial characteristic, respectively.

[0043] In one specific implementation, the input data for this step includes: enterprise pollution status indicators. ( (vectors, 1 / 0 variables, representing contaminated and uncontaminated respectively), non-spatial feature matrix ( matrix, (the number of non-spatial features), and the filtered gather( matrix, After filtering The quantity, for example, can be selected from the 8 features with the largest values. , respectively denoted as - ).

[0044] Based on the above data, a linear regression was first performed using each feature vector as the independent variable and the pollution status index of each enterprise as the dependent variable to obtain the pollution status index of each enterprise after removing the global spatial effect. This step uses... right By performing linear regression to isolate the influence of spatial patterns on pollution status, we obtain the "pure dependent variable". The formula is:

[0045] in, express right The regression coefficients, for ; Represents the residual, as The vector, This eliminates the interference of spatial autocorrelation on pollution status and can serve as a pollution status indicator for each enterprise after stripping away global spatial effects. As the primary pollution status indicator, This is the second pollution status indicator.

[0046] Similarly, using each eigenvector as the independent variable and each non-spatial characteristic of each enterprise as the dependent variable, linear regression is performed to obtain the non-spatial characteristics of each enterprise after removing the global spatial effect. This step uses... For each non-spatial feature ( Linear regression was performed to isolate the influence of spatial patterns on non-spatial features, yielding the "pure independent variables". The formula is:

[0047] in, for right The regression coefficients, for The vector, For residuals, for Vectors; integrating all The pure independent variable matrix can be obtained. ,for The matrix. Similarly, That is, the first non-spatial feature. This is the second non-spatial feature.

[0048] After completing the above peeling process, the peeling effect can be verified to check its effectiveness. and Whether to completely strip away the global space influence. In one specific implementation, the process may include the following steps: Step 1: Using the secondary spatial characteristics of each enterprise as independent variables and the secondary pollution state indicators of each enterprise as dependent variables, perform a multiple linear regression and calculate the global average net effect based on the regression model. Specifically, use the pure independent variables... For pure dependent variable Perform multiple linear regression to obtain the FWL coefficient (global average net effect) and significance, using the following formula:

[0049] in, FWL coefficient ( The vector represents "the global average net impact of non-spatial features on the pollution state after excluding spatial interference" (e.g., When the industry type is "heavy industry", the net increase in pollution level is 0.42 units. For regression residuals ( (vector) Expectation matrix Sum of standard deviation matrix Each satisfies , The standard deviation is represented by t-test. If the p-value is less than 0.05, the global net effect is significant.

[0050] Simultaneously, spatial autocorrelation tests were performed on the second pollution state indicators of each enterprise to obtain spatial autocorrelation. Optionally, calculations can be performed... Moran's I index (a spatial autocorrelation test index) is used to verify whether spatial influences have been completely removed. The formula is:

[0051] in, for The mean (i.e., the mean of the companies).

[0052] Step 2: Verify the spatial stripping effect based on the global average net effect and spatial autocorrelation. Optionally, if both the global average net effect and the spatial autocorrelation are less than the corresponding thresholds, the global spatial effect stripping is considered successful. For example, if Moran's I index is close to 0 (i.e., the absolute value of the index is <0.1) and p>0.05, it indicates... If there is no significant spatial autocorrelation, the spatial effect removal is successful. Otherwise, if the global average net effect is greater than or equal to the corresponding threshold, or if the spatial autocorrelation is greater than or equal to the corresponding threshold, the global spatial effect removal is considered to be ineffective. In this case, the K value can be increased, and the operation in S120 of "calculating the spatial weights between the current enterprise and the K nearest other enterprises for each enterprise" can be returned until the global spatial effect removal is successful.

[0053] S150. Using each of the second non-spatial characteristics as independent variables and the second pollution state index as the dependent variable, fit the enterprise pollution state model.

[0054] This embodiment models data stripped of spatial effects in order to clarify data relationships beyond the global spatial pattern.

[0055] In one specific implementation, the final model can be determined through an automatic model selection framework, and an optimized model can be established in the model selection process as follows:

[0056] in, This represents the hyperparameter configuration in a machine learning model. This represents a collection of various hyperparameter configuration schemes. Indicates the hyperparameters of the machine learning model The verification loss below, For the first The computational cost of various hyperparameter configurations, The number of hyperparameters configured; For cost budgeting, CPU / GPU usage time and / or memory consumption can be used as indicators. This model addresses the time-constrained hyperparameter search problem, aiming to achieve cost-effectiveness. Inside, find the validation loss. Minimal model and hyperparameter configuration .

[0057] Optionally, the automatic model selection framework can be FLAML (Fast and Lightweight AutoML), a high-efficiency automated machine learning framework that can automatically select the optimal machine learning algorithm (such as XGBoost, LightGBM, Random Forest, etc.) and hyperparameter configuration to optimize model performance within a limited time budget. This embodiment uses a four-step progressive strategy (initialization → exploration → fusion → convergence) within this framework to efficiently identify the optimal hyperparameter configuration within a limited time budget, while ensuring the generalization ability of spatial data.

[0058] In one specific implementation, data preparation is performed first. A stratified sampling method can be used to divide the dataset into a training set and a test set in an 8:2 ratio. and Split into training set ( , ) and test set ( , ), and maintain a consistent distribution ratio of contaminated states in the training and test sets.

[0059] Then, integrate the FLAML AutoML framework to achieve intelligent model selection. Optionally, a cost-aware search strategy can be employed, which includes the following steps: Step 1: Low-cost initialization. This involves starting the search with the hyperparameter configuration that minimizes computational overhead, ensuring rapid acquisition of baseline performance. Quickly obtaining a "baseline model" (that can run and has basic performance) provides a "starting point" for subsequent searches, avoiding wasting time with high-cost configurations from the outset.

[0060] Step 2: Conduct incremental exploration. A cost-saving optimization (CFO) algorithm can be used to achieve a cost-performance trade-off by optimizing a composite objective function. g(c) = ·f(c) + (1 )·[cost(c) / cost max ] Where g(c) represents the comprehensive evaluation score, and the optimization objective is to minimize it; The coefficient represents the trade-off, α∈[0,1], which can be determined as needed. The larger α is, the more emphasis is placed on "model performance". The smaller the size, the more emphasis is placed on "cost savings". max This represents the maximum possible computational cost in the search space.

[0061] Step 3: Perform global-local fusion. Optionally, the BlendSearch algorithm can be used to combine the global exploration capabilities of Bayesian optimization with the local search efficiency of CFO. The priority of the two search methods is determined as follows:

[0062] in, Represents a global search. Represents local search, priority This indicates that global search is prioritized. This indicates a preference for local searches. It can be seen that the priority is based on the ratio of performance improvement to cost. Dynamic adjustment, if The global / local switching threshold has not yet been reached. Using a global search, the range of parameters to be searched is very large; if Reaching the threshold It performs a fine-grained search in the vicinity of the current optimal region.

[0063] Meanwhile, the search space employs a hierarchical design. XGBoost's parameter space includes the number of trees. Maximum depth Learning rate The parameter space of LightGBM includes the number of leaf nodes. Feature sampling ratio When the performance improvement is significant Early stopping is triggered in a timely manner to avoid over-searching the converged parameter region. The coefficient of determination represents the model performance of the historical best solution. This represents the model performance (coefficient of determination) corresponding to the current solution. To set a threshold.

[0064] Meanwhile, during model training, considering spatial autocorrelation, spatial block cross-validation can be used to verify model performance, thus avoiding data leakage problems caused by traditional random cross-validation. Optionally, spatial clustering is first performed. Based on the spatial coordinates of enterprise points, the K-means clustering algorithm is used to divide the study area into n spatial blocks (n is the number of cross-validation folds). Then, spatial independent folds are generated. In each round of cross-validation, it is ensured that the training set and validation set are spatially isolated from each other. At the same time, to enhance the spatial isolation effect, a buffer distance parameter can be set to ensure that there is a certain distance interval between adjacent spatial blocks and avoid boundary effects.

[0065] By combining spatial block cross-validation, unbiased R² estimation can be performed to ensure that the hyperparameter optimization process takes into account the autocorrelation characteristics of spatial data and prevents overfitting.

[0066] Finally, the optimal trained model and its evaluation results are output and passed to the subsequent GeoShapley spatial interpretation and analysis module. This model is the enterprise's pollution state model, which can save model objects and hyperparameter configurations, record model performance indicators (R², RMSE, MAE), and generate feature importance ranking and SHAP value interpretation.

[0067] S160. Through Geoshapley spatial interpretation analysis, the pollution state model is decomposed into spatial background effects and non-spatial background effects to obtain the final pollution state model.

[0068] After obtaining the optimal model through S150, this embodiment performs Geoshapley analysis on the model to further identify and extract spatial driving factors of pollution status that are related to the enterprise itself and its local environment, in addition to global spatial effects. Compared to global spatial effects (such as "high pollution in the eastern region and low pollution in the western region"), the impact of these local factors on pollution exhibits significant spatial heterogeneity (i.e., the impact mechanisms of each enterprise are different). More importantly, these factors (such as the enterprise's production duration and specific processes) are usually more controllable than macro-geographical factors (for example, improving macro-factors in the entire eastern and western regions is not easy, but improving factors of individual enterprises is relatively easy). Therefore, separating global effects and focusing on these local driving factors provides a key basis for achieving "precise and differentiated" enterprise pollution control in this application. This is also the reason why this embodiment separates global spatial effects and further extracts these factors.

[0069] Specifically, GeoShapley is a spatial interpretation method that extends the classic Shapley value theory, incorporating geospatial interactions into the feature contribution calculation framework. By considering the spatial neighborhood relationships of samples, it decomposes the model's prediction results into spatial background effects and non-spatial feature effects, enabling interpretability analysis of spatial machine learning models. Optionally, a third-party GeoShapley library can be used to implement GeoShapley analysis. The main tasks in this embodiment are input data preparation (i.e., the aforementioned optimal model and its independent and dependent variable data), parameter configuration, result interpretation, and visualization. The internal principles of GeoShapley are explained below.

[0070] GeoShapley extends the classic Shapley value theory by incorporating spatial interactions into the feature contribution calculation framework, thus enhancing model predictions. Decomposed into spatial background effect and non-spatial characteristic effect:

[0071] in, This refers to the spatial background, which is the portion that is only related to the company's geographical location (latitude and longitude). This represents the contribution of non-spatial features. The above formula is the final pollution state model, which clarifies the mechanism of action of each driving factor of the pollution state and provides a further explanation of the optimal model obtained in S150.

[0072] Based on this model, the influence of each driving factor can be visualized. Figure 2 This is the visualization result output by GeoShapley analysis in this embodiment. The left figure shows the spatial background effect, and the right figure shows the effect of the most important non-spatial feature. Each point in the figure represents a company, and the color of each point represents the value of the local effect or the non-spatial effect of the feature. The larger the value, the more significant the impact.

[0073] Furthermore, to demonstrate the advantages of this embodiment, different experiments were designed as shown in the table below:

[0074] As shown in the table, `experiment_id` represents different batches of experiments. In the baseline experiment, the data does not require processing (i.e., no spatial effect stripping). The geographic matrix refers to data fitting experiments using only a spatial weight matrix constructed from geographic coordinates for spatial effect stripping. The hybrid matrix refers to data fitting experiments using a spatial weight matrix that integrates geographic coordinates and economic-ecological coupling for spatial effect stripping (i.e., the method of this application). The first column, `w_matrix`, indicates the matrix calculation method: `N / A` indicates no spatial weight matrix, `Geo` indicates a spatial weight matrix constructed using geographic coordinates, and `Fused` indicates a spatial weight matrix that integrates geographic coordinates and economic-ecological coupling. The third column, `spatial_model`, indicates the spatial model processing method; the fourth column, `final_r2`, indicates the model fit; the fifth column, `residual_moran_i`, indicates the residual Moran index; and the sixth column, `residual_moran_p`, indicates the significance of the residual Moran index. Columns 5 and 6 are used to indicate the effectiveness of the model; a higher fit in column 4 indicates better model performance. It can be seen that the fit of the method in this embodiment is significantly improved. More specifically, GeoShapley decomposes the model prediction for each sample into two independent components:

[0075]

[0076] in, Indicates only with enterprises The spatial background effect related to geographical location; Representing non-spatial features In enterprises The spatial variation coefficient is used to reflect the spatial heterogeneity of characteristic effects; This represents the number of non-spatial features. It can be smaller than This is equivalent to extracting a portion of important non-spatial features from all non-spatial features; This represents the set of neighboring firms obtained in the K-nearest neighbor algorithm; Indicates non-empty component features The baseline contribution, For the neighbors (i.e., the kth closest enterprise to enterprise i) for features Spatial modulation term of contribution.

[0077] Based on the above data, suggestions can be provided for improving the pollution status of enterprises or regions. Since the spatial background effect is limited by geographical location and difficult to change, suggestions can be made based on the specific circumstances of each enterprise. The size of the pollution level determines the priority for improving the pollution status of each enterprise; for example, prioritizing improvements to those... Smaller businesses will Overly large companies should be given lower priority. Unlike spatial characteristics that are only related to the enterprise itself or its surrounding environment and are relatively easy to change, the spatial variability coefficients calculated by GeoShapley can reflect the differences in the impact of these specific driving factors (such as annual precipitation, wind speed, etc.) on the enterprise's pollution status at different spatial locations. For a single enterprise, this can be... Greater than the set threshold The recommendations are to optimize the enterprise's pollution situation. For example, if production time has a significant impact, it is recommended to adjust production time; if rainfall has a significant impact, it is recommended to increase waterproofing measures. For each region, the common impacts within the region can be addressed. It is recommended to optimize the targets for improving the pollution status of the region.

[0078] The above analysis can identify key pollution drivers and their spatial distribution patterns, providing environmental management departments with precise regional or enterprise-specific control recommendations. Furthermore, based on this data, it can also provide suggestions for enterprise site selection (e.g., for enterprises in specific macro-regions where certain industry types are unsuitable due to ecological conditions, wind speed, and temperature) and environmental risk assessment. Additionally, it can generate spatial driver heat maps to guide the allocation of environmental governance resources.

[0079] In summary, this embodiment provides a method for constructing an enterprise pollution status model based on spatial driving factor analysis. It fully considers the geospatial autocorrelation and spatial heterogeneity of enterprise pollution, and can accurately analyze the regional characteristics of pollution formation. Specifically, this method can achieve the following beneficial effects: 1. Quantification of Regional Collaborative Characteristics: By collecting data on economic development (e.g., GDP per capita) and ecological environment quality, calculating the coupling coordination degree and incorporating it into the data system, this approach, for the first time, integrates the collaborative relationship between "enterprise development and ecological environment" into enterprise pollution assessment. This addresses the shortcomings of traditional techniques that neglect regional sustainable development factors, making the data characteristics more closely aligned with the actual mechanisms of pollution formation. Furthermore, in constructing the spatial weight matrix, the integration of "geographical distance + coupling coordination degree difference" more accurately quantifies the spatial relationships between enterprises than the traditional "single geographic distance weight," improving the accuracy of subsequent MEM extraction and GeoShapley analysis. 2. Spatial patterns are extracted and identified through MEM, and then spatial influences of pollution status indicators and non-spatial features are removed through FWL analysis to generate clean data without significant spatial autocorrelation, thus solving the core problem of spatial pseudo-correlation interfering with the judgment of non-spatial features in traditional models.

[0080] 3. By integrating the FLAML AutoML framework, it automatically adapts to algorithms such as XGBoost, LightGBM, and Random Forest, selecting the optimal model with R² as the objective. It matches data features without manual intervention, improving model prediction accuracy and adaptability, and avoiding the limitations of traditional single-algorithm trial and error. It achieves a performance-efficiency balance within a limited time budget, solving the problem of excessively time-consuming hyperparameter search in traditional AutoML, and adapting to rapid parameter space optimization using XGBoost, LightGBM, and other algorithms.

[0081] 4. Multi-dimensional analysis: By integrating the global net effect of FWL coefficients, AutoML feature importance values, and GeoShapley values ​​(spatial heterogeneity effects), a three-dimensional effect system of "global-prediction-regional" is constructed. First, the global spatial effect is removed to avoid the global spatial effect from masking the influence of other driving factors when it is too significant. Based on the data after removal, the system can accurately distinguish whether the enterprise's pollution is due to its own characteristics (such as industry type and production duration) or the influence of spatial environmental factors (such as regional ecological environment). This breaks through the limitations of traditional single-dimensional analysis and further clarifies the driving mechanism of enterprise pollution.

[0082] 5. Differentiating the role and interaction of space with other factors: Geo-Shapley analysis of spatial interaction effects can guide regional governance (such as developing differentiated measures for areas with high interaction effects); at the same time, based on the spatial variation coefficient of GeoShapley, the impact differences of driving factors (such as precipitation and wind speed) in different regions can be clarified, providing regionally differentiated management suggestions for environmental pollution improvement; spatial driving factor heat maps can guide the allocation of governance resources to high-impact areas, improving resource utilization efficiency.

[0083] It should be noted that all data involved in this application are information and data that have been fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0084] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 60, a memory 61, an input device 62, and an output device 63; the number of processors 60 in the device can be one or more. Figure 3 Taking a processor 60 as an example; the processor 60, memory 61, input device 62, and output device 63 in the device can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0085] The memory 61, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the enterprise pollution state model construction method based on spatial driving factor analysis in this embodiment of the invention. The processor 60 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 61, thereby realizing the aforementioned enterprise pollution state model construction method based on spatial driving factor analysis.

[0086] The memory 61 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 61 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 61 may further include memory remotely located relative to the processor 60, which can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0087] Input device 62 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Output device 63 may include display devices such as a display screen.

[0088] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the enterprise pollution state model construction method based on spatial driving factor analysis of any embodiment.

[0089] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0090] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0091] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0092] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as C or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing an enterprise pollution status model based on spatial driving factor analysis, characterized in that, include: Obtain the geographical location and primary pollution status indicators of each enterprise within the area to be analyzed; Based on the geographical location of each enterprise, a weight matrix is ​​constructed to represent the spatial relationship between enterprises; and the weight matrix is ​​decomposed into eigenvalues ​​to extract at least one feature vector to represent the global spatial effect within the region. Using each eigenvector as the independent variable and the first pollution state index and each first non-spatial feature of each enterprise as the dependent variable, a linear regression is performed to obtain the second pollution state index and each second non-spatial feature of each enterprise after removing the global spatial effect; specifically, using each eigenvector... The independent variable is the first pollution state index of each enterprise. Performing linear regression on the dependent variable yields the following model: ,in, express right The regression coefficients, Represents the residual; As a secondary pollution status indicator after enterprises have stripped away the global spatial effects; based on each characteristic vector As independent variables, the second non-spatial characteristics of each enterprise are used respectively. Performing linear regression on the dependent variable yields the following model: ,in, express right The regression coefficients, Represents the residual; As the second non-spatial characteristic of each enterprise after stripping away the global spatial effects; Using each of the second non-spatial features as independent variables and the second pollution state index as dependent variables, a pollution state model for enterprises is fitted; and through Geoshapley spatial interpretation analysis, the pollution state model is decomposed into spatial background effects and non-spatial background effects to obtain the final pollution state model. Specifically, the decomposition of the pollution state model into spatial background effects and non-spatial background effects through Geoshapley spatial interpretation analysis includes: Through GeoShapley space interpretation analysis, the enterprise Pollution status prediction Decomposed into: in, Indicates only with enterprises The spatial background effect related to geographical location; Representing non-spatial features In enterprises The spatial variation coefficient is used to reflect the spatial heterogeneity of characteristic effects; Indicates the number of non-spatial features; This represents the set of neighboring firms obtained in the K-nearest neighbor algorithm; Representing non-spatial features The baseline contribution, For the neighbors Features Spatial modulation term of contribution.

2. The method according to claim 1, characterized in that, The step of obtaining the geographical location and first pollution state index of each enterprise in the area to be analyzed includes: obtaining the geographical location and first pollution state index of each enterprise in the area to be analyzed, as well as the ecological environment data and enterprise environmental data at each geographical location; Accordingly, after constructing a weight matrix to characterize the spatial relationship between enterprises based on their geographical locations, the method further includes: calculating the coupling coordination degree of each enterprise to characterize the degree of adaptation between enterprise development and the ecological environment based on the ecological environment quality data and enterprise environmental data at each geographical location; and correcting the weight matrix based on the differences in coupling coordination degree between enterprises to obtain the final weight matrix.

3. The method according to claim 2, characterized in that, The step of correcting the weight matrix based on the differences in coupling coordination between enterprises to obtain the final weight matrix includes: The weight matrix is ​​corrected according to the following formula: in, Indicates enterprise With enterprises Inter-adjusted weights Indicates enterprise With enterprises Geographical distance, and Different companies and enterprises The degree of coupling coordination This is a correction factor for the difference in coupling degree.

4. The method according to claim 1, characterized in that, The step of constructing a weight matrix to represent the spatial relationship between enterprises based on their geographical locations includes: calculating the spatial weight between the current enterprise and its K nearest neighbor enterprises for each enterprise, where K is a natural number; and constructing a weight matrix to represent the relationship between enterprises based on the spatial weights. Accordingly, after performing linear regression with each feature vector as independent variable and each enterprise's first pollution state index and each first non-spatial feature as dependent variables to obtain each enterprise's second pollution state index and each second non-spatial feature after removing the global spatial effect, the process further includes: performing multiple linear regression with each enterprise's second spatial feature as independent variable and each enterprise's second pollution state index as dependent variable to determine the global average net effect; performing spatial autocorrelation tests on each enterprise's second pollution state index to determine the spatial autocorrelation; if both the global average net effect and the spatial autocorrelation are less than the corresponding thresholds, the global spatial effect is considered successfully removed; otherwise, the K value is increased, and the process returns to calculating the spatial weights between the current enterprise and its K nearest enterprises for each enterprise, until the global spatial effect is successfully removed.

5. The method according to claim 1, characterized in that, The process of fitting an enterprise pollution status model using each second non-spatial feature as an independent variable and the second pollution status index as a dependent variable includes: By using an automatic model selection framework, the optimal machine learning model is determined with each enterprise's second non-spatial feature as the independent variable and each enterprise's second pollution state index as the dependent variable. The following optimization objectives were followed during the determination process: in, This represents the hyperparameter configuration in a machine learning model. This represents the validation loss of the machine learning model. For the first The computational cost of various hyperparameter configurations, The number of hyperparameters configured. For cost budgeting, This represents a collection of various hyperparameter configuration schemes. This indicates that within the cost budget B, we are looking for a way to minimize the verification loss. Minimal machine learning model and hyperparameter configuration.

6. The method according to claim 1, characterized in that, After decomposing the pollution state model into spatial background effects and non-spatial background effects, the method further includes: According to each company The size of the pollution level determines the priority for improving the pollution status of each enterprise; and / or Will Greater than the set threshold Recommendations for improving the enterprise The optimization target for the pollution state.

7. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the enterprise pollution state model construction method based on spatial driving factor analysis as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the enterprise pollution state model construction method based on spatial driving factor analysis as described in any one of claims 1-6.

Images