Cross-domain atmospheric pollution control trend evaluation and analysis method based on DPSIR-DFCM-LLM

Abstract

The invention discloses a cross-domain atmospheric pollution abatement trend assessment and analysis method based on DPSIR-DFCM-LLM. The method comprises the following steps: building an atmospheric pollution abatement assessment system based on a DPSIR framework; fusing a dynamic fuzzy cognitive map (DFCM) and a large language model (LLM); determining the weight of the evaluation index; and performing evaluation process and simulation analysis. According to the invention, the DPSIR model is adopted to build an atmospheric pollution cross-regional governance evaluation system, a dynamic fuzzy cognitive map is constructed to calculate the weight of each index of the evaluation system, and scientific and feasible policy suggestions are provided; meanwhile, a large language model (LLM) is fused, the efficiency and accuracy of policy text analysis and public opinion mining are improved, and an evaluation system is further optimized; the method is beneficial for promoting cross-regional treatment of atmospheric pollution to step towards a higher level, and also provides experience support for regional collaborative treatment in a larger range and different fields.

Description

Technical Field

[0001] The present invention relates to the field of evaluation and analysis of air pollution control, and particularly to a method for evaluating and analyzing the cross-domain air pollution control trend based on DPSIR-DFCM-LLM. Background Art

[0002] Regarding the necessity of collaborative governance and the realistic challenges faced, most are demonstrated at the theoretical level. However, it is not easy to evaluate the air pollution control effect of cross-regional governance for heterogeneity problems. Although many studies have explored the heterogeneous policy effects through traditional methods, such as adding interaction terms or using grouped regression analysis and obtained valuable insights, this method faces some challenges: when the heterogeneity variables are highly collinear with the policy variables, the interaction terms between them may not provide meaningful explanations, and it is difficult to handle multiple heterogeneities. Additionally, the strict requirements of traditional models for the functional form (such as only linear or piecewise) often do not match the complex structure of actual data, easily causing model misspecification, and the characterization of multiple heterogeneities cannot be effectively carried out. Therefore, adding interactive items to the traditional analysis framework, these interactive items cannot be fully utilized when facing the complex situation of multiple heterogeneities, although it is for simplifying the interpretation of the real world. This limits the value of the conclusions of traditional methods for heterogeneous policy effects, especially under the condition of big data, where the key features of data such as non-linearity, heterogeneity, and dynamics are restricted by the increase in sample size, variable dimension, and noise. Summary of the Invention

[0003] Object of the Invention: The object of the present invention is to provide a method for evaluating and analyzing the cross-domain air pollution control trend based on DPSIR-DFCM-LLM.

[0004] Technical Solution: The present invention includes the following steps:

[0005] (1) Build an air pollution control evaluation system based on the DPSIR framework;

[0006] (2) Fusion of the Dynamic Fuzzy Cognitive Map (DFCM) and the Large Language Model (LLM);

[0007] (3) Determine the weights of evaluation indicators;

[0008] (4) Evaluation process and simulation analysis.

[0009] Furthermore, the step (1) based on the DPSIR framework includes 16 indicators.

[0010] Furthermore, the step (2) includes determining the causal relationship of the dynamic fuzzy cognitive map and the fusion of DFCM and LLM.

[0011] Furthermore, the causal relationship of the dynamic fuzzy cognitive map is represented in the form of a graph at the structural level. The nodes of the fuzzy cognitive map represent the concepts of the system under study, and the edges represent the causal relationships between these concepts.

[0012] Furthermore, the fuzzy cognitive map is used to describe the relationship between input concepts and outputs. Using "if-then" fuzzy rules, the weight W of the causal relationship between two nodes ij is described by a fuzzy inference system. represents the degree of causal influence between concept nodes C i , C j . is a mapping. In the dynamic fuzzy cognitive map, a rule has a single antecedent related to a concept or a change in a concept and outputs the change intensity V representing the perturbation in the target concept ij :

[0013] If C i is A THEN Variation is V ij on C j

[0014] If Variation is V ij IN C i THEN Variation is V ij on C j .

[0015] Furthermore, the fuzzy cognitive map is represented by a triple {C, W, f}, where C is the set of concept nodes, W is the set of causal relationships between two nodes, and f is the inference transformation function of the node states of the fuzzy cognitive map. ij

[0016] Furthermore, it is characterized in that the fusion of the DFCM and the LLM intuitively represents the causal relationships between different factors in the form of a graph, integrates data from different sources through fuzzy logic and processes them uniformly, and can dynamically select and adjust concept nodes according to the actual situation.

[0017] Furthermore, it is characterized in that step (3) includes calculating the air quality index based on P1, P2, P3, P4, P5, and these pollutants are considered the main indicators of air pollution.

[0018] Further, the evaluation process in step (4) includes identifying and extracting descriptive information related to each concept node in the DFCM through natural language processing techniques, converting the extracted key information into concept nodes and causal relationships in the DFCM, monitoring and analyzing newly released policy documents, media reports, and social media comments in real time through the LLM to identify new information that may affect the weights of the DFCM, using the sentiment analysis function of the LLM to evaluate the public's views on a certain policy or technology, and using the LLM to analyze the impact of major events on each node.

[0019] Further, the simulation analysis in step (4) tests the air pollution control evaluation system constructed based on DPSIR, LLM, and the dynamic fuzzy cognitive map by running different scenarios.

[0020] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: The present invention uses the DPSIR model to build an air pollution cross-regional governance evaluation system, constructs a dynamic fuzzy cognitive map to calculate the weights of various indicators of the evaluation system, and provides scientific and feasible policy recommendations; at the same time, it integrates the large language model (LLM), improves the efficiency and accuracy of policy text analysis and public opinion mining, and further optimizes the evaluation system; it is conducive to promoting the air pollution cross-regional governance to a higher level, and also provides experience support for regional collaborative governance in a larger scope and different fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 are the air pollution control evaluation indicators corresponding to regions A and B;

[0022] Figure 2 is the air pollution control evaluation model diagram based on the dynamic fuzzy cognitive map;

[0023] Figure 3 are the indicators and their thresholds described in fuzzy logic language, as well as the membership function curves;

[0024] Figure 4 is the set of variant fuzzy variables;

[0025] Figure 5 are the changing trends of the air pollution control evaluation indicators corresponding to three scenarios, where (a) is the full-coverage governance scenario; (b) is the zero-governance scenario; (c) is the random governance. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The technical solutions of the present invention will be further described below with reference to the accompanying drawings.

[0027] The specific steps of this method are as follows:

[0028] 1. Build an air pollution control evaluation system based on the DPSIR model

[0029] 1.1 Indicators and Connotations

[0030] The European Environment Agency adopted the DPSIR framework in 1995 and it is described as "a causal framework for clarifying the interactions between society and the environment". Under this framework, socioeconomic activities are considered as the driving forces exerting pressure on the environment, thus affecting its biological, chemical or physical state. In past research, scholars have widely applied the DPSIR framework to analyze various environmental problems. Through the application of the DPSIR framework, researchers can more comprehensively understand the nature of environmental problems, identify key influencing factors, and propose effective environmental policies and management measures. In addition, the DPSIR framework has also been widely used to evaluate the implementation effects of environmental policies, helping governments and policymakers better understand the effectiveness and deficiencies of environmental policies.

[0031] Table 1 Performance Evaluation Index System for Air Pollution Control

[0032]

[0033]

[0034] Based on the relevant research of previous scholars, index screening was carried out, and finally 16 representative indicators were determined. Among them, in the driving force dimension, the driving force generally refers to the root causes of changes in the atmospheric environment caused by factors such as population density, energy consumption, vehicle exhaust, petrochemical industry, etc.; in the pressure dimension, the pressure mainly refers to the consumption of natural resources caused by various human social activities and emissions of various pollutants; in the state dimension, the state refers to the environmental state formed under the action of pressure, with air quality indicators as the assessment indicators; in the Impact dimension, Impact reflects the impact of the system on human health in the current environmental state; in the response dimension, the response refers to the corresponding actions taken after the ecological environment deteriorates and brings adverse effects to life, including formulating and implementing environmental supervision policies, promoting the use of clean energy, promoting the green coverage rate, promoting new energy vehicles, and investing special funds for pollution control, with the ultimate goal of improving air quality and weakening the adverse effects on the ecological environment.

[0035] 1.2 Measured Data

[0036] The measured data of the above variables all come from the release of authoritative institutions.

[0037] 2. Dynamic Fuzzy Cognitive Map (DFCM) Integrated with LLM

[0038] 2.1 Causal Relationship of Dynamic Fuzzy Cognitive Map

[0039] The Dynamic Fuzzy Cognitive Map (DFCM) is a system model, which is represented in the form of a graph at the structural level. Its nodes represent the concepts of the system under study, and the edges represent the causal relationships between these concepts.

[0040] A general fuzzy cognitive map is represented by a triple {C, W, f}, where C is a set of concept nodes, and W is the weight of the causal relationship between two nodes ij the set of causal relationships between representing the concept node C i , C j the degree of causal influence between is a mapping, and f is the node state inference transformation function of the fuzzy cognitive map. There are many implementations of fuzzy cognitive maps in the literature, but they rely on using monotonic and symmetric weights to represent the causal relationships between concepts, which is not very effective considering the dynamic characteristics of most environmental systems. Therefore, in this paper, a dynamic rule-based fuzzy cognitive mapping method, namely the dynamic fuzzy cognitive map, is created by eliciting knowledge from experts and stakeholders using the relevant metrics required for the evaluation created previously. The linguistic term set of the dynamic fuzzy cognitive map concept is used to describe the relationship between the input concept and the output, using "if-then" fuzzy rules. The weight W of the causal relationship between two nodes ij is described using a fuzzy inference system (FIS), representing the concept node C i , C j the degree of causal influence between is a mapping. In the dynamic fuzzy cognitive map, the rule has a single antecedent related to a concept or a change in a concept and outputs the change intensity V representing the perturbation in the target concept ij :

[0041] If C i is A THEN Variation is V ij on C j

[0042] If Variation is V ij IN C i THEN Variation is V ij on C j

[0043] A single fuzzy variable is used here to represent the changes in causal nodes and impact nodes, without the need to duplicate for different metrics. The defined values of the change variable are ("NVH: Negative Very High", "NH: Negative High", "NM: Negative Medium", "NL: Negative Low", "NEGL: Negligible", "PL: Positive Low", "PM: Positive Medium", "PH: Positive High", "PVH: Positive Very High"). Additionally, although there is one variable to handle all changes, the syntax allows for a user-friendly knowledge acquisition dialogue with experts, where the causal nodes or effect nodes in specific rules can be clearly seen.

[0044] 2.2 Integrating DFCM and LLM

[0045] Through literature review and analysis, most environmental assessment methods focus on the mechanism of air pollution control policies on macroeconomics or industrial efficiency from a static perspective, and few literatures discuss the effectiveness of dynamic policy adjustments. Secondly, when integrating expert information, a small number of literatures involve the uncertainty and ambiguity problems caused by quite subjective viewpoints. To address these two issues, this paper uses the Dynamic Fuzzy Cognitive Map (DFCM) to calculate the weights of various indicators in the assessment system. At the same time, to further improve the accuracy and real-time performance of the assessment, this paper introduces the Large Language Model (LLM). The LLM can efficiently process and analyze large amounts of text data, extract key information from multi-source texts such as policy documents, news reports, and social media, and assist in constructing and optimizing the DFCM model.

[0046] Although large language models (LLMs) excel in processing and analyzing text data, extracting information, predicting trends, etc., in the evaluation of the effectiveness of air pollution control, the dynamic fuzzy cognitive map (DFCM) still plays an irreplaceable and important role: DFCM can handle complex non-linear causal relationships, and the weights in DFCM can be dynamically adjusted according to changes in the system state. While it is difficult for LLMs to capture complex non-linear relationships when modeling complex causal relationships and have fixed weights or parameters that are difficult to adjust dynamically. When the system is affected by multiple factors, DFCM can better capture the dynamic evolution of the system and provide a more robust model when facing uncertainty and ambiguity. LLMs are difficult to capture the dynamic changes of the system and are difficult to directly handle uncertainty and ambiguity. DFCM intuitively represents the causal relationships between different factors in the form of a graph, making the model highly interpretable and facilitating policymakers to understand and trust. LLMs are usually black-box models and are difficult to interpret. DFCM can integrate data from different sources through fuzzy logic and uniformly process them, and can dynamically select and adjust concept nodes according to the actual situation. It is difficult for LLMs to comprehensively integrate multi-source data such as numerical data, and their processing ability is limited. These characteristics endow DFCM with unique advantages in complementing LLMs and jointly improving the accuracy and scientificity of the evaluation in the assessment of the effectiveness of air pollution control. Therefore, combining LLMs with DFCM can make up for their respective deficiencies, give full play to their respective strengths at the same time, and provide more comprehensive and accurate support for air pollution control.

[0047] 3. Determination of the weights of evaluation indicators

[0048] As can be seen from Table 1, air pollution is mainly caused by human driving forces. The DPSIR loop of air pollution in the model has driving forces such as "energy consumption per unit of GDP", "number of heavy industries", "population density", and "vehicle ownership", which will affect air quality. The air quality index is calculated based on P1, P2, P3, P4, P5, and these pollutants are considered the main indicators of air pollution; health risks are mainly related to the mortality and morbidity caused by asthma, bronchitis, conjunctivitis, and respiratory infections; in order to reduce the impact of air pollution on health and quality of life. Five emission reduction strategies are listed: improving the impact of "public transportation" in the city, increasing the proportion of "pollution control investment", using "clean fuels" in the city, implementing existing "pollution control policies and regulations", and promoting the use of "new energy vehicles". The final dynamic fuzzy cognitive map model with different causal relationships between indicators in the evaluation model is as Figure 2 shown:

[0049] 4. Evaluation process and simulation analysis

[0050] 4.1 Evaluation process

[0051] The integrated LLM can extract valuable information from a large amount of text data, including but not limited to policy documents, news reports, academic literature, and social media comments, to enhance the accuracy and richness of the data. It can also identify and extract descriptive information related to each concept node in the DFCM through natural language processing techniques. Convert the extracted key information into concept nodes and causal relationships in the DFCM. Monitor and analyze newly released policy documents, media reports, social media comments, etc. in real time through the LLM to identify new information that may affect the weights of the DFCM. Use the sentiment analysis function of the LLM to evaluate the public's views on a certain policy or technology. Use the LLM to analyze the impact of major events on each node. Finally, use the indicators described in fuzzy logic language and their thresholds, as well as membership functions such as Figure 3 shown, taking population density as an example:

[0052] For the weights between indicators, experts use if-then rules representing the causal relationships between indicators to determine the degree of influence exerted by one indicator on another. The fuzzy rules for evaluating indicator weights are shown in Table 3. The degree of fluctuation caused by one indicator to another is described by the fuzzy variable "mutation", and its fuzzy set is as Figure 4 shown:

[0053] Table 3: Fuzzy Rules for Evaluating Indicator Weights

[0054]

[0055] 4.2 Simulation Analysis of Air Pollution Control

[0056] Simulation Analysis of Air Pollution Control

[0057] It can be seen from the air pollution control evaluation system built based on the DPSIR model that the trend of the driving force is uncertain. Here, in view of the nature of air pollution control performance evaluation, by setting three different scenarios, different air pollution control policies or measures are analyzed, and the dynamic fuzzy cognitive map model is run to generate the quantitative trend of air pollution. For the convenience of the comparative analysis of the control results under the three scenarios below, by running simulations of different scenarios, the air pollution control evaluation system constructed based on DPSIR, LLM, and dynamic fuzzy cognitive map is tested. Here, the three scenarios are named "zero control", "random control", and "full coverage control" respectively: among them, zero control represents the worst trend scenario without considering any control policies and measures; random control represents the context baseline scenario with appropriate coordinated implementation of the document measures deployed by the regulatory agency; full coverage control represents the super strong trend scenario with increased implementation of air pollution control measures. Initial stimuli are assigned to the indicators of the three strategy scenarios as shown in Table 4.

[0058] Table 4 Initial Incentive Settings for Three Scenarios

[0059] D1 D2 D3 D4 P1 P2 P3 P4 P5 S I R1 R2 R3 R4 R5 Zero governance 5000 35 35 5 120 60 120 120 60 5 0.8 0 0 0 0 0 Random governance 5000 35 35 3 120 60 120 120 60 5 0.8 0.3 0.3 0.3 0.3 0.3 Full-coverage governance 5000 35 35 3 120 60 120 120 60 5 0.8 1 1 1 1 1

[0060] By analyzing the changing trends of the air pollution control evaluation indicators corresponding to the three scenarios as Figure 5 shown, it can be found that: First, the levels of different pollutants in different scenarios have all increased to a certain extent; Second, in the zero-control scenario, all the policy measures for air pollution control are fully implemented. From the results, it can be seen that this scenario has reduced the overall impact on air pollution and health risks; after the operation process of the random-control scenario tends to converge, the pollution indicators continue to rise and the environmental situation deteriorates sharply, and the overall air pollution and health risks are at the highest values; at the same time, it also reflects that even without the support of sufficient policy measures, the air pollution indicators are predictable, but the health risks can reach a shocking level. The random-control scenario aims to analyze the current situation, while the full-coverage control scenario model focuses on the implementation of more powerful policies. The results of this scenario show that compared with the results of the zero-control scenario, the air quality has improved and the health risks have decreased. The results of the full-coverage control scenario show that the air pollution has only decreased by 15%, and the health risks have slightly decreased by 1%. By restricting public transportation, law enforcement, and maximizing the introduction of clean fuels, the reduction of health risks can be achieved.

[0061] 4.3 Empirical Analysis of Regional Heterogeneity

[0062] In this paper, by using the DPSIR and dynamic fuzzy cognitive map methods, the overall effects of air pollution control policies are analyzed. However, the main pollutants, natural environment, economic development level, and policy implementation intensity vary among different regions, and there may be differences in the governance effects of policies in practice. Therefore, in order to further test whether the air pollution control policies have regional heterogeneity characteristics, in this paper, based on the evaluation index samples of Regions A and B, they are respectively embedded into the established air pollution control evaluation system. As Figure 1 shown, it shows the changing trends of the air pollution control evaluation indicators corresponding to Regions A and B, and a comparative analysis of the overall effects of the air pollution control policies in the two regions is carried out. It can be seen from the figure that the overall air quality in both Regions A and B has been significantly improved, and Region A is more prominent in terms of the promoting effect. Specifically for different types of air pollutants, the policy has an obvious inhibitory effect on the emissions of PM2.5, SO2, and NO2 in Region B, but has little effect on reducing the concentrations of PM10 and O3; in the treatment of the three major types of air pollutants, PM2.5, PM10, and SO2, the absolute values and significance levels of the emission reduction coefficients of various pollutants in Region A are higher than those in Region B.

Claims

1. A cross-domain air pollution control trend evaluation and analysis method based on DPSIR-DFCM-LLM, characterized in that, It includes the following steps: (1) Build an air pollution control evaluation system based on the DPSIR framework; (2) Integration of the Dynamic Fuzzy Cognitive Map (DFCM) and the Large Language Model (LLM); (3) Determination of the weights of evaluation indicators; (4) Evaluation process and simulation analysis.

2. The cross-domain air pollution control trend evaluation and analysis method based on DPSIR-DFCM-LLM according to claim 1, characterized in that The step (1) includes 16 indicators based on the DPSIR framework.

3. The cross-domain air pollution governance trend evaluation and analysis method based on DPSIR-DFCM-LLM according to claim 1, characterized in that, The step (2) includes the determination of the causal relationship of the dynamic fuzzy cognitive map and the integration of the DFCM and the LLM.

4. The cross-domain air pollution control trend evaluation and analysis method based on DPSIR-DFCM-LLM according to claim 3, wherein The causal relationship of the dynamic fuzzy cognitive map is represented in the form of a graph at the structural level. The nodes of the fuzzy cognitive map represent the concepts of the system under study, and the edges represent the causal relationships between these concepts.

5. The cross-domain air pollution control trend evaluation and analysis method based on DPSIR-DFCM-LLM according to claim 4, characterized in that, The fuzzy cognitive map is used to describe the relationship between input concepts and outputs, using "if-then" fuzzy rules, and the weight W of the causal relationship between two nodes ij is described by an analog inference system represents the degree of causal influence between concept nodes C i , C j and is a mapping. In a dynamic fuzzy cognitive map, a rule has a single antecedent related to a concept or a change in a concept and outputs the change intensity V representing the perturbation in the target concept ij : If C i is A THEN Variation is V ij on C j If Variation is V ij IN C i THEN Variation is V ij on C j 。 6. The cross-domain air pollution governance trend evaluation and analysis method based on DPSIR-DFCM-LLM according to claim 4, characterized in that, The fuzzy cognitive map is represented by a triple {C, W, f}, where C is a set of concept nodes, and W is a set of causal relationships between two nodes ij and f is an inference transformation function for the node states of the fuzzy cognitive map.

7. The cross-domain air pollution governance trend evaluation and analysis method based on DPSIR-DFCM-LLM according to claim 3, characterized in that, The integration of the DFCM and the LLM intuitively represents the causal relationship between different factors in the form of a graph, integrates data from different sources through fuzzy logic and processes them uniformly, and can dynamically select and adjust the concept nodes according to the actual situation.

8. The method for cross - domain air pollution control trend evaluation and analysis based on DPSIR - DFCM - LLM according to claim 1, characterized in that, The step (3) includes that the air quality indicators are calculated according to P1, P2, P3, P4, P5, and these pollutants are considered the main indicators of air pollution.

9. The cross-domain air pollution control trend evaluation and analysis method based on DPSIR-DFCM-LLM according to claim 1, characterized in that The evaluation process of the step (4) includes identifying and extracting the descriptive information related to each concept node in the DFCM through natural language processing technology, converting the extracted key information into the concept nodes and causal relationships in the DFCM, monitoring and analyzing newly released policy documents, media reports, and social media comments in real time through the LLM, identifying new information that may affect the weights of the DFCM, using the sentiment analysis function of the LLM to evaluate the public's views on a certain policy or technology, and using the LLM to analyze the impact of major events on each node.

10. The method for cross - domain air pollution control trend assessment and analysis based on DPSIR - DFCM - LLM according to claim 1, wherein, The simulation analysis of the step (4) tests the air pollution control evaluation system constructed based on the DPSIR, the LLM, and the dynamic fuzzy cognitive map by running and simulating different scenarios.

Images