Atmospheric pollution cause and traceability analysis method and system, electronic equipment and storage medium

Abstract

The invention provides an atmospheric pollution cause and traceability analysis method and system, electronic equipment and a storage medium, and the method comprises the steps: obtaining a pollution traceability request inputted in a natural language form, carrying out the semantic analysis of the pollution traceability request based on a large language model, and extracting each analysis element; executing a plurality of data acquisition tasks in parallel according to the analysis elements, and obtaining multi-source data including pollutant time sequence data, meteorological time sequence data and enterprise emission list data; performing multi-source heterogeneous data fusion processing on the multi-source data to generate a structured data set; based on a multi-expert collaborative analysis engine, dynamically scheduling multiple field experts to analyze the structured data set, the multiple field experts including meteorological experts and component analysis experts; and performing comprehensive reasoning and information fusion on analysis results of experts in each field to generate a comprehensive analysis conclusion containing pollution sources and causes. The system response efficiency is improved, and the analysis accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and artificial intelligence, and in particular to a method, system, electronic device and storage medium for analyzing the causes and sources of air pollution based on a large language model. Background Technology

[0002] Current efforts to trace the sources of air pollution have the following shortcomings in terms of technological application: On the one hand, existing analytical tools suffer from technical shortcomings such as high professional barriers and fragmented operational processes. Traditional source tracing analysis relies on specialized models such as HYSPLIT, PMF, and EKMA. Each model requires manual parameter configuration, data preprocessing, model operation, and result interpretation, necessitating operators with expertise in multiple disciplines, including environmental science and meteorology. More critically, the lack of a unified integration interface and collaboration mechanism between multiple models necessitates repeated switching between different software and manual data transfer, resulting in complex operations, high error rates, and long analysis cycles, making it difficult to meet the timeliness requirements of pollution emergency response.

[0003] On the other hand, existing technologies lack the ability to efficiently coordinate multi-source data and multiple models. Data sources such as air quality monitoring data and meteorological data differ significantly in format, spatiotemporal resolution, and storage methods, making it difficult for existing methods to achieve automated data fusion and precise spatiotemporal matching. Furthermore, transmission path analysis, source apportionment, and sensitivity analysis rely on multiple independent models, lacking a unified scheduling mechanism between models. This hinders the automatic integration and cross-validation of analysis results, leading to a single analytical dimension, insufficient information utilization, and difficulty in supporting accurate identification and rapid source tracing of pollution.

[0004] In summary, existing technologies in the field of air pollution source tracing suffer from problems such as complex operation, strong professional dependence, difficulty in multi-source data fusion, and low efficiency of multi-model collaboration, making it difficult to meet the actual needs of environmental management departments for rapid response, accurate source tracing, and scalability at the grassroots level. Summary of the Invention

[0005] This invention provides a method, system, electronic device, and storage medium for analyzing the causes and sources of air pollution based on a large language model, in order to solve the problems existing in related technologies. The technical solution is as follows: In a first aspect, embodiments of the present invention provide a method for analyzing the causes and sources of air pollution based on a large language model, including: The system acquires pollution source tracing requests in natural language form, performs semantic parsing on the pollution source tracing requests based on a large language model, and extracts analytical elements including target area, time range, and target pollutant. Based on the analysis elements, multiple data acquisition tasks are executed in parallel. These tasks include acquiring multi-source data related to the pollution source tracing request from multiple data sources, including pollutant time-series data, meteorological time-series data, and enterprise emission inventory data. Multi-source heterogeneous data fusion processing is performed on pollutant time-series data, meteorological time-series data, and enterprise emission inventory data to generate a structured dataset; Based on a multi-expert collaborative analysis engine, multiple domain experts are dynamically scheduled to analyze structured datasets. These experts include meteorological experts who perform transmission path analysis and component analysis experts who perform source resolution. By integrating and fusion the analytical results of experts from various fields using a large language model, a comprehensive analytical conclusion including the sources and causes of pollution is generated.

[0006] In one implementation, the multi-source heterogeneous data fusion processing of pollutant time-series data, meteorological time-series data, and enterprise emission inventory data includes: Anomaly detection using a multi-strategy fusion approach is performed on pollutant time-series data and meteorological time-series data to identify outliers or missing values. The multi-strategy fusion approach includes: statistical anomaly detection based on dynamic thresholds, anomaly determination based on multi-model voting, and rule verification based on national standard limits. Multiple interpolation strategies matching the missing data pattern are used to correct outliers or missing values, resulting in corrected time series data. These strategies include linear interpolation, historical mean imputation, and weighted moving average. The corrected time-series data is correlated and aligned with the enterprise emission inventory data in both time and spatial dimensions to generate a structured dataset; A comprehensive analytical view containing pollutant concentrations, meteorological conditions, and emission source information is constructed based on structured datasets.

[0007] In one implementation, executing multiple data acquisition tasks in parallel includes: Based on NL2SQL technology, the analysis elements are converted into structured query statements to obtain pollutant time series data from the environmental monitoring database; Based on NL2API technology, the analysis elements are parsed into API call parameters to obtain meteorological time series data from meteorological data sources; Extract the latitude and longitude coordinates, industry, and emission inventory data of enterprises within a preset range surrounding the target area from the enterprise information database; the enterprise emission inventory data includes the emission intensity and composition information of pollutants.

[0008] In one implementation, the meteorological expert's transmission path analysis method includes: The backward motion trajectory of the air mass was simulated using the Lagrange particle diffusion model based on meteorological time-series data. Cluster analysis was performed on multiple backward trajectories to identify the dominant transport pathways of pollutants; By combining the geographical coordinates of enterprises in the enterprise emission inventory data, potential contributing source enterprises located upwind of the dominant transmission path are screened out.

[0009] In one implementation, the source resolution method of the component analysis expert includes: When the target pollutant is ozone, the empirical kinetic simulation model is invoked to determine the sensitivity type of ozone formation based on the precursor concentration ratio in the pollutant time series data. Combined with the volatile organic compound component data in the enterprise emission inventory and the maximum incremental reactivity coefficient of each component, the ozone formation potential of each enterprise is calculated to obtain the ozone formation potential ranking results. When the target pollutant is particulate matter, the positive definite matrix factorization model is invoked to quantify and output the contribution rate of each pollution source based on the particulate matter component data in the pollutant time series data.

[0010] In one implementation, a comprehensive analytical conclusion is generated by integrating and reasoning the analysis results of experts from various fields using a large language model, including the sources and causes of pollution: Based on the dominant transmission path output by meteorological analysis experts, the transmission type of the pollution event is determined to be either regional transmission type or locally generated type. Based on the ozone formation potential ranking results or the contribution rate of each pollution source output by the component analysis experts, and combined with the spatial relationship between the geographical location of enterprises and the dominant transmission path, a list of key suspected enterprises and a relative contribution assessment are generated. The transmission type of pollution incidents, the list of key suspected enterprises, and the contribution rate of each pollution source are integrated into a structured comprehensive analysis output.

[0011] In one implementation, the multiple domain experts also include visualization experts and report generation experts; Visualization experts are used to generate visualization charts including time series plots, trajectory plots, ranking plots, and distribution plots. Report generation experts use this technology to integrate comprehensive analytical conclusions with visual charts into structured analytical reports.

[0012] Secondly, embodiments of the present invention provide an air pollution cause and source tracing analysis system, which performs the air pollution cause and source tracing analysis method as described above, including: The natural language interaction unit is used to obtain pollution source tracing requests input by users in natural language form; The core unit of the large language model is trained on domain corpus and used to perform semantic parsing of pollution source tracing requests to extract analytical elements. The parallel data acquisition unit is used to execute multiple data acquisition tasks in parallel according to the analysis elements. The multiple data acquisition tasks include at least obtaining pollutant time-series data from the environmental monitoring database, obtaining meteorological time-series data from the meteorological data source, and obtaining enterprise emission inventory data from the enterprise information database. The data fusion processing unit is used to perform multi-source heterogeneous data fusion processing on pollutant time series data, meteorological time series data and enterprise emission inventory data to generate structured datasets; A multi-expert collaborative analysis unit is used to dynamically call multiple domain expert modules to analyze structured datasets; the multiple domain expert modules include at least a meteorological analysis expert module for transmission path analysis and a component analysis expert module for source resolution; The results integration and output unit is used to input the analysis results of expert modules from various fields into the core unit of the large language model. The core unit of the large language model performs cross-result fusion and causal reasoning to generate a comprehensive analysis conclusion that includes the sources and causes of pollution.

[0013] Thirdly, embodiments of the present invention provide an electronic device comprising a memory and a processor. The memory and the processor communicate with each other via an internal connection path. The memory stores instructions, and the processor executes the instructions stored in the memory. When the processor executes the instructions stored in the memory, it causes the processor to perform the method described in any of the above embodiments.

[0014] Fourthly, embodiments of the present invention provide a computer-readable storage medium that stores a computer program, wherein when the computer program is run on a computer, the methods in any of the embodiments described above are executed.

[0015] The advantages or beneficial effects of the above technical solutions include at least the following: This invention allows users to input pollution source tracing requests in natural language. The large language model automatically performs semantic parsing, multi-source data querying, professional model invocation, and result integration on the requests. This transforms the traditional analysis process, which requires multiple manual operations, into a fully automated process, significantly reducing the number of human-computer interactions and operation time, and improving system response efficiency.

[0016] The ReAct framework initiates a multi-expert collaborative analysis process, with a large language model acting as the core controller to dynamically call multiple domain experts to analyze structured datasets. Expert modules such as meteorological analysis and component analysis work collaboratively under unified scheduling, avoiding the repetitive calculations and waiting time caused by the independent operation of multiple models and manual integration of results in traditional methods. This significantly shortens the entire process cycle from data input to conclusion output, meets the timeliness requirements of pollution emergency response, and improves the accuracy of atmospheric pollution causal source analysis.

[0017] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0018] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the various drawings denote the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings depict only some embodiments disclosed in the invention and should not be construed as limiting the scope of the invention.

[0019] Figure 1 This is a flowchart illustrating the method for analyzing the causes and sources of air pollution according to the present invention. Figure 2 This is a flowchart illustrating the dynamic scheduling process of the multi-expert collaborative analysis engine based on the ReAct framework of this invention. Figure 3 This is a schematic diagram of the five-layer vertical architecture of the atmospheric pollution cause and source tracing analysis system of the present invention. Figure 4 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0020] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0021] Example 1 This embodiment provides a method for analyzing the causes and sources of air pollution based on a large language model, referencing... Figure 1 As shown, the method specifically includes the following steps: Step S1: Obtain the pollution source tracing request input in natural language.

[0022] By constructing a natural language interaction interface, pollution source tracing requests input by users in a conversational and unstructured format can be received. This natural language interaction interface can be implemented in various forms, such as web page dialog boxes, instant messaging robots, voice input terminals, or API calls. Users do not need to follow strict instruction formats or technical jargon; they only need to describe their analytical intent in everyday language.

[0023] Pollution source tracing requests may include key analytical elements such as the target area (e.g., a city), time range (e.g., "the past three days"), and target pollutant type (e.g., "PM2.5" or "ozone"), but these are all freely expressed in natural language without structured processing. This embodiment converts the natural language into text form to facilitate semantic analysis of the user request by a large language model.

[0024] Step S2: Based on the large language model, perform semantic parsing on the pollution source tracing request to extract analytical elements including the target area, time range, and target pollutant.

[0025] The user request text obtained in step S1 is input into the large language model. This large language model has undergone specialized domain-adaptive training, and its training corpus covers professional fields such as environmental monitoring reports, meteorological texts, pollution source analysis cases, and environmental protection policies and regulations, enabling it to have accurate semantic understanding and entity recognition capabilities in the field of environmental science.

[0026] Specifically, the large language model, through its built-in attention mechanism and contextual understanding capabilities, performs lexical analysis, syntactic analysis, and semantic role labeling on user request text, identifying and extracting three core analytical elements: Target area: Identifies the geographical scope involved in the request, such as place names of a province or city, and supports multi-granularity positioning from the city level, district / county level to the monitoring station level; Time range: Parse the implicit time descriptions in the request, such as natural time expressions like "the past three days", "yesterday", "March 15th to 17th", etc., and convert them into standardized start and end timestamps; Target pollutant type: Identify the names of pollutants mentioned in the request, such as "PM2.5", "ozone", "PM10", etc., and support single pollutant or multi-pollutant parallel analysis.

[0027] After extracting the analytical elements, the system verifies their completeness. If the verification reveals missing key information, the system does not directly terminate the process or fill in the missing default values. Instead, it generates guiding prompts through a natural language interface, prompting the user to provide supplementary information. For example, if the user enters "Analyze the pollution in a certain city," and the system recognizes that the time range and target pollutant type are missing, it will return the prompt: "Please supplement the analysis time range (e.g., 'the past three days') and target pollutant (e.g., PM2.5, ozone)." This interactive process employs a multi-turn dialogue mechanism until a complete set of analytical elements is obtained.

[0028] Through the above methods, this step realizes the transformation from unstructured natural language to structured analytical elements. At the same time, through integrity verification and interactive supplementation mechanisms, it ensures that the subsequent data processing and analysis process has complete input parameters, avoiding process interruption or analysis result deviation due to missing information, and improving the robustness of the system and user experience.

[0029] Step S3: Based on the analysis elements, execute multiple data acquisition tasks in parallel. These tasks include obtaining multi-source data related to the pollution source tracing request from multiple data sources. The multi-source data includes pollutant time-series data, meteorological time-series data, and enterprise emission inventory data.

[0030] After receiving the analysis elements (target area, time range, target pollutant) output in step S2, a multi-threaded parallel data acquisition task is initiated. Through thread pool management and asynchronous non-blocking I / O technology, multiple acquisition tasks are executed concurrently, which significantly reduces the end-to-end time consumption of the entire process and provides technical support for meeting the timeliness requirements of pollution emergency response.

[0031] Specifically, the following four dimensions of data acquisition tasks are executed in parallel: Step S31: Pollutant Time Series Data Acquisition Pollutant time-series data refers to data records arranged in chronological order, including the location of monitoring stations and the corresponding concentration values ​​of various pollutants at specific times. The encapsulated NL2SQL module automatically converts natural language analysis elements into structured query language (SQL) statements conforming to the structure of environmental monitoring databases, thus completing the collection of pollutant time-series data.

[0032] The NL2SQL module embeds schema mapping rules for the environmental monitoring database, including table structure mapping, field semantic mapping, and conditional filtering rules. Table structure mapping predefines the relationships between core data tables in the database, such as the monitoring station table, pollutant hourly concentration table, and equipment information table. Field semantic mapping establishes a precise mapping between pollutant names in natural language, such as "PM2.5 concentration," "SO2," and "NO2," and their corresponding field names in the database. Conditional filtering rules can convert "target area" into station code filtering conditions and "time range" into timestamp interval conditions.

[0033] Based on the above mapping rules, the NL2SQL module can accurately convert user semantic descriptions such as "PM2.5 concentration at a certain site" into complete SQL statements containing table joins, conditional filtering, and time sorting, and execute the query to return hourly concentration time series data of the target area and surrounding monitoring stations within a specified time range.

[0034] Step S32: Meteorological Time Series Data Acquisition Meteorological time-series data refers to data records arranged chronologically, containing monitoring station locations or latitude and longitude coordinates and corresponding meteorological element values ​​such as wind direction, wind speed, temperature, and relative humidity. This embodiment utilizes the encapsulated NL2API module to automatically parse natural language analysis elements into API call parameters conforming to external meteorological data source interface specifications, thus completing the collection of meteorological data. The NL2API module pre-configures standard API call templates for multiple meteorological data sources (such as public APIs from meteorological departments and commercial meteorological service interfaces), including endpoint URL configuration, parameter mapping rules, and parameter validation logic. The endpoint URL configuration pre-configures the request addresses of each data source; the parameter mapping rules convert the target area into latitude and longitude ranges or station codes, convert time ranges into start and end time parameters, and convert meteorological elements into fields such as wind direction, wind speed, temperature, and relative humidity; the parameter validation logic verifies the legality of the generated time format, coordinate range, and parameter combinations, ensuring that the API call conforms to the interface specifications and avoiding call failures due to incorrect parameters. Through this mechanism, the NL2API module can automatically construct compliant HTTP requests, achieving accurate meteorological data collection.

[0035] Step S33: Enterprise Emission Inventory Data Collection Enterprise emission inventory data refers to data that includes the enterprise's geographical coordinates, industry type, VOCs, and NO. x This embodiment records data on the emission intensity and VOCs composition information of pollutants. It extracts enterprise emission data related to the target area from a locally or cloud-deployed enterprise information database. The acquisition logic includes two stages: spatial range filtering and emission inventory extraction. Spatial range filtering sets a spatial buffer zone (e.g., a 5-kilometer radius) based on the center coordinates or administrative boundaries of the target area, extracting the geographical coordinates, industry type, and other information of all enterprises within this zone. Emission inventory extraction obtains the pollutant emission inventory data of these enterprises, including VOCs and NOx. x Information on the emission intensity and emission components of key precursors.

[0036] Step S34: Domain Knowledge Support Set Retrieval By employing retrieval-enhanced generation techniques, background knowledge relevant to the current pollution event is retrieved from a vectorized knowledge base, forming a domain knowledge support set. Specifically, unstructured text data such as historical pollution event cases, environmental policies and regulations, technical standards, and academic literature are pre-divided into text blocks. Each text block is then converted into a high-dimensional vector using an embedding model and stored in a vector database. Subsequently, the analytical elements obtained in step S2 are vectorized, and a similarity search is performed in the vector database to recall the text blocks most relevant to the current analytical scenario. Finally, the recalled relevant text blocks are organized into structured contextual information for reference by the expert module and the large language model during subsequent analysis and reasoning.

[0037] Subsequently, all parallel data acquisition tasks are monitored and coordinated in a unified manner. When an anomaly occurs in a certain acquisition channel (such as data source unavailability, interface timeout, or abnormal data format), the module records the anomaly log and handles it according to the preset strategy: for non-core data (such as historical cases), acquisition failure is allowed and the data is marked as missing; for core data (such as pollutant concentration), a retry mechanism is triggered or feedback is provided to the user to ensure the robustness of the overall analysis process.

[0038] Step S4: Perform multi-source heterogeneous data fusion processing on pollutant time series data, meteorological time series data, and enterprise emission inventory data to generate a structured dataset.

[0039] Before performing multi-source heterogeneous data fusion processing, the data is first standardized, including four steps: format conversion and extraction, data cleaning, correction of missing and outlier values, and word segmentation and semantic processing.

[0040] In the format conversion and extraction stage, various monitoring data exported from the environmental monitoring database (such as Excel and CSV formats) are uniformly converted into JSON format for storage. During the conversion process, the definition, data type, and unit of each field, as well as other metadata information, are clearly defined, and a structured data dictionary is established to ensure that data from different sources are consistent in format, facilitating subsequent programmatic processing and field indexing.

[0041] In the data cleaning stage, outliers and missing values ​​in the monitoring data are systematically processed. A multi-strategy fusion method is employed to identify outliers, avoiding misjudgments caused by skewed data distribution or short-term fluctuations due to a single statistical method. Specifically, this multi-strategy fusion method includes three aspects: First, dynamic threshold correction, which dynamically calibrates the 3σ principle threshold by combining historical data (such as pollutant concentration distribution over the past three years) to eliminate the influence of seasonal extreme values; second, multi-model fusion judgment, which comprehensively judges suspected outliers using the 3σ principle, box plot method (based on interquartile range IQR), and moving average deviation method. If at least two methods mark it as an outlier, it is confirmed as an outlier; third, business rule overlay, which combines the absolute concentration limits set by national standards such as the "Ambient Air Quality Standard" (GB 3095-2012). Data exceeding the limits and marked by statistical methods are directly marked as outliers and enter the manual review process to ensure the business rationality of the outlier determination.

[0042] In the missing and outlier correction process, engineered processing is used based on fixed formulas and strategies, rather than arbitrarily generating values. All correction operations are recorded in the data traceability log, supporting manual backtracking and review. Specific correction strategies include: for cases where valid data exists before and after the missing value, and the time interval is no more than 2 hours, linear interpolation is used for filling in the missing value. The linear interpolation formula is: ; in, The concentration at the moment to be filled. , The effective concentrations before and after are given, and t is the time to be filled.

[0043] For consecutive or isolated missing data points, the historical moving window mean is used for filling. The formula for filling with the historical moving window mean is as follows: ; The average concentration at the same time over the previous n days (usually n=3~7) is used for filling to avoid intraday periodic deviations.

[0044] For outliers (such as sudden jumps to zero or high values), a weighted moving average with front and back windows is used for correction, as shown in the formula: ; Among them, C t C represents the corrected concentration value of the outlier to be corrected at time t; t i The original concentration value (forward window data) at the i-th time before the anomalous time t; C t+i The original concentration value (backward window data) at the i-th time after the anomalous time t; ω iThe weight coefficient for the i-th data point is calculated using a Gaussian kernel or the inverse of distance. Data points closer to the anomaly point have a higher weight, ensuring the correction value is smooth and conforms to the short-term trend of the data; k is the number of data points selected in the forward window; m is the number of data points selected in the backward window; ∑ ω This is the sum of all weight coefficients for the forward and backward windows, used for normalization calculations.

[0045] In the word segmentation and semantic processing stage, the system uses natural language processing technology to segment and lexically analyze textual data (such as policies and regulations, historical cases, technical reports, etc.) to extract keywords and professional terms; through semantic understanding and annotation, it constructs preliminary knowledge associations, laying the foundation for subsequent semantic segmentation and vector generation of the knowledge base.

[0046] Through the collaborative processing of the above four steps, the original multi-source heterogeneous data can be transformed into a standard dataset with a unified format, reliable quality, and clear semantics, providing high-quality data input for subsequent data screening, integration, and multi-expert collaborative analysis.

[0047] This embodiment systematically filters and structurally integrates preprocessed multi-source data to form a comprehensive analytical view that includes pollutant concentration, meteorological conditions, and emission source information, providing data support for the efficient operation and accurate reasoning of a multi-expert collaborative analysis engine.

[0048] Specifically, systematic screening and structured integration include three stages: target area screening, data correlation integration, and data hierarchy division, ultimately generating a structured dataset.

[0049] In the target area screening stage, based on the target area information obtained from step S2, the collected pollutant time-series data and meteorological time-series data are geographically filtered. Specifically, a list of monitoring station codes covered by the target area is extracted, and only data from stations belonging to this list are retained, while irrelevant data outside the area is removed, ensuring that subsequent analysis focuses on the target area and reducing redundant calculations.

[0050] In the data correlation and integration stage, a spatiotemporal correlation mapping is established between different types of data to provide directly usable aligned data for each expert module. This stage includes alignment processing in three dimensions: First, temporal alignment, which unifies pollutant concentration data, meteorological element data, and enterprise emission data to the same time resolution (e.g., hourly level) and establishes a time index to support joint time series analysis across data sources; second, spatial alignment, which calculates the spatial distance between monitoring stations and enterprises based on the latitude and longitude coordinates of the monitoring stations and the geographical coordinates of the enterprises in the emission inventory, constructs a spatial proximity table, and enables rapid association between monitoring stations and surrounding enterprises, providing a spatial mapping basis for meteorological transmission path analysis and source selection; third, knowledge association, which associates policy and regulation documents, historical cases, and other document data with domain knowledge classification tags, facilitating the accurate retrieval of background knowledge relevant to the current analysis scenario through subsequent retrieval enhancement generation technology.

[0051] In the data hierarchy segmentation stage, based on the importance and frequency of use of the data during the analysis process, the data is divided into three levels: core, auxiliary, and reference, and a hierarchical storage mechanism is established. Core data includes hourly concentrations of target pollutants, air quality index, and key policies and regulations, which are essential data for each analysis scenario and are prioritized for loading into memory. Auxiliary data, such as meteorological derivative indicators (mixing layer height, atmospheric stability), and industry technical guidelines, are called up as needed in specific analysis stages. Reference data includes historical comparison data for the same period and regional background concentrations, used for result verification and in-depth interpretation, and is loaded only when generating reports or when requested by users.

[0052] The aforementioned hierarchical strategy works closely with the dynamic scheduling mechanism based on the ReAct framework. When scheduling the meteorological analysis expert module, it prioritizes loading the core data it needs (such as pollutant concentration data and meteorological time-series data) from memory, while reference data is loaded only on demand during the report generation stage. This intelligent caching and loading strategy significantly reduces memory usage and data access latency, making it one of the key performance optimization methods to ensure that the entire analysis process is completed within ten minutes.

[0053] Through the systematic processing of the above three steps, this step integrates the preprocessed multi-source heterogeneous data into a unified data view that includes spatiotemporal alignment information, knowledge association information, and a hierarchical storage mechanism, providing a data foundation for the efficient operation of the subsequent multi-expert collaborative analysis engine.

[0054] Furthermore, to ensure that the data quality throughout the entire data processing process meets the requirements of subsequent analysis, this step also establishes a quality control mechanism covering the entire process of "collection-preprocessing-screening-database entry," employing a combination of automated verification and manual review to achieve systematic assurance of data quality.

[0055] Specifically, the quality control system includes three core verification steps: In the data acquisition quality verification stage, the multi-source data acquired in parallel in step S3 is verified for acquisition quality. This verification includes integrity verification, which checks whether core fields are missing; timeliness verification, which ensures that real-time data delays do not exceed 1 hour and forecast data is updated on the same day; and authority verification, which verifies that the data source is compliant and traceable. In case of acquisition failure or data anomalies, the system promptly triggers an early warning mechanism and takes appropriate action to prevent low-quality data from flowing into subsequent processing flows.

[0056] In the preprocessing quality verification stage, automated tools are used to perform quality checks on the preprocessed data. Specifically, this includes: checking the accuracy of format conversion and confirming the integrity of table structures; verifying the effectiveness of data cleaning, ensuring that redundant information has been removed and outliers have been marked; and verifying the consistency of standardization, confirming uniform formats and standardized terminology. Meanwhile, for critical documents such as policies, regulations, and technical standards, the system employs manual sampling review to ensure that the preprocessing results meet business requirements.

[0057] In the integration quality verification phase, the integrated data view undergoes quality verification. This verification includes: verifying the accuracy of data associations, confirming no deviation in spatiotemporal matching and no errors in label association; and verifying the accuracy of filtering, ensuring the absence of irrelevant data from outside the designated area. Through these verifications, the integrated data is ensured to be logically coherent and directly usable for subsequent multi-expert collaborative analysis engine calls.

[0058] Furthermore, to ensure efficient data storage and long-term availability, this step designed a hybrid storage architecture and established a routine data update and maintenance mechanism.

[0059] In terms of storage architecture design, a hybrid storage architecture combining vector databases, relational databases, and document databases is adopted. The vector database stores document vector data transformed by the embedding model, supporting efficient retrieval using retrieval enhancement generation technology. The relational database (supporting mainstream databases such as MySQL, SQL Server, Highgo, and DM) stores structured monitoring data and statistical indicators such as pollutant concentrations, meteorological elements, and enterprise emission inventories. The document database stores raw documents, preprocessed text resources, and intermediate processing records, ensuring efficient and targeted data storage.

[0060] Regarding data updates and maintenance, a routine data update mechanism has been established. For real-time monitoring data, incremental updates are automatically performed based on the data source update frequency; for document resources, a regular update cycle of once a month is set. Simultaneously, for dynamic information such as policies, regulations, and standards that may affect the accuracy of analysis, a dedicated tracking mechanism has been established. The knowledge base content is promptly updated after relevant standards are released and updated to ensure the timeliness of the knowledge base and the compliance of the analysis conclusions.

[0061] Through the synergistic effect of the aforementioned quality control system and data storage and maintenance mechanism, this step ensures that the data quality is controllable, the storage architecture is reasonable, and the update mechanism is sound throughout the entire process from data collection to storage, providing a reliable data foundation for the stable operation of the subsequent multi-expert collaborative analysis engine.

[0062] Step S5: Based on the multi-expert collaborative analysis engine, dynamically schedule multiple domain experts to analyze the structured dataset. These experts include meteorological experts who perform dominant transmission path analysis and component analysis experts who perform source resolution.

[0063] Based on the structured dataset generated in step S4, a multi-expert collaborative analysis engine is launched, with a large language model as the core controller. The ReAct framework is used to dynamically schedule each expert module, realizing the automated transformation from multi-source data to professional analysis conclusions.

[0064] Specifically, such as Figure 2 As shown, the multi-expert collaborative analysis engine includes a meteorological analysis expert module, a component analysis expert module, a visualization expert module, and a report generation expert module. These modules work collaboratively under unified scheduling to jointly complete the multi-dimensional analysis of pollution source tracing.

[0065] Among them, the meteorological analysis expert module simulates the possible long-distance transport paths of pollutants by calling the HYSPLIT backward trajectory model, providing meteorological basis for the analysis of pollution sources.

[0066] The meteorological analysis expert module first receives the coordinates of the monitoring station, the type of target pollutant, and the time range information to determine the simulation start time and duration. The HYSPLIT model uses the Lagrange method, tracking air masses, and takes preprocessed three-dimensional meteorological grid data (including wind direction, wind speed, temperature, relative humidity, etc.) as input. The HYSPLIT model calculates the backward trajectory of the air mass within a set time period (e.g., 72 hours) before it reaches the monitoring station by continuously integrating the meteorological elements at discrete points in space and time. During the calculation, the air mass is treated as a closed system, its altitude changing with vertical atmospheric motion, and the time interval between trajectory points is controlled by the model step size.

[0067] After calculating multiple backward trajectories, the meteorological analysis expert module performs cluster analysis on the trajectories to identify the dominant transport paths and prevailing wind directions for pollutant transmission. Based on the identified prevailing wind directions, the meteorological analysis expert module, in conjunction with the spatial proximity table constructed in step S4, calculates the angle between the geographical coordinates of each enterprise and the prevailing wind direction, and filters out a set of potential contributing enterprises located upwind, providing spatial directional basis for subsequent source analysis.

[0068] The component analysis expert module calls on the corresponding professional model to perform source apportionment or sensitivity analysis based on the type of target pollutant, thereby achieving qualitative and quantitative analysis of pollution sources.

[0069] When the target pollutant is ozone, the component analysis expert module calls the EKMA (Empirical Kinetic Simulation) model to perform ozone formation sensitivity analysis. The EKMA model generates ozone isoconcentration curves by simulating photochemical reactions at different initial VOCs to NOx concentration ratios. The component analysis expert module combines this with measured precursor concentration ratios to automatically determine the sensitivity type of the current pollution event.

[0070] It should be explained that the measured precursor concentration ratio is calculated based on the VOCs component concentration data and NOx concentration data in the structured dataset.

[0071] The component analysis expert module calls the EKMA model to map the measured precursor concentration ratios onto the ozone isoconcentration curve, automatically determining whether the current ratio falls within the VOCs control zone, NOx control zone, or transition zone. If it is in the VOCs control zone, reducing VOCs concentration is more effective in reducing ozone; if it is in the NOx control zone, reducing NOx is more effective. Simultaneously, the component analysis expert module combines VOCs component data from enterprise emission inventories and the maximum incremental reactivity coefficient of each component to calculate and rank the ozone generation potential of each enterprise, obtaining an ozone generation potential ranking result that accurately identifies the major emission sources that contribute significantly to ozone generation.

[0072] When the target pollutant is particulate matter, the component analysis expert module calls the PMF (Positive Deterministic Matrix Factorization) model for source apportionment. This model, based on the positive definite matrix factorization method, decomposes particulate matter component data (such as sulfate, nitrate, ammonium salt, organic matter, elemental carbon, etc.) in the pollutant time-series data into multiple factors. It then identifies the physical meaning of each factor based on its chemical composition characteristics, quantifies the contribution rate of various pollution sources such as industrial sources, mobile sources, dust sources, and secondary transformation sources, and achieves qualitative and quantitative analysis of pollution sources.

[0073] The visualization expert module receives the output results from the meteorological analysis expert module and the component analysis expert module, and automatically generates various types of visualization charts to intuitively display the analysis conclusions. Specific generated charts include: time-series overlay plots reflecting pollutant concentration changes over time; transport trajectory plots showing the movement path of air masses; bar charts ranking enterprise contributions; pie charts showing the industry share of pollution sources; and spatial distribution maps showing the distribution of pollution.

[0074] The report generation expert module receives the output results from the above expert modules, including meteorological analysis conclusions, component analysis conclusions, and visualization charts. It integrates this information into structured analysis information, providing a unified input format for the subsequent comprehensive reasoning of the large language model.

[0075] The activation and invocation of the aforementioned multi-expert modules are not pre-set in a fixed order, but are dynamically scheduled by the large language model as the core controller based on the ReAct framework. Specifically, the large language model generates inference actions to invoke specific expert modules based on the current analysis task objective and the acquired structured dataset; in response to this inference action, the system invokes the corresponding expert module, passing a subset of data in the structured dataset associated with that expert module as input; after obtaining the execution result of the expert module, the result is integrated into the context as observation information, driving the large language model to generate the next inference action or enter the final conclusion output stage. Through this "reasoning-action-observation" cyclical mechanism, the analysis path can be flexibly adjusted according to the actual data characteristics and analysis progress, achieving efficient collaboration among multiple models.

[0076] Step S6: Through the large language model, the analysis results of experts from various fields are comprehensively reasoned and information is integrated to generate a comprehensive analysis conclusion that includes the sources and causes of pollution.

[0077] The large language model, combined with the dominant transport path output by meteorological analysis experts and cluster analysis results, determines whether the pollution event is of regional transport type or locally generated type. If the dominant transport path shows that the air mass mainly originates from outside the target area and its trajectory clustering is relatively concentrated, it is determined to be "regional transport type"; if the dominant transport path shows that the air mass trajectory is relatively dispersed or lingers within the target area, it is determined to be "locally generated type". The determination results are output in structured field format.

[0078] The large language model cross-integrates the potential contributing enterprises in the upwind direction output by meteorological analysis experts with the enterprise contribution ranking output by component analysis experts to generate a list of key suspected enterprises. The specific integration logic includes: For ozone pollution scenarios, the top-ranked companies in the ozone formation potential ranking output by component analysis experts are extracted as potential key suspected companies. Then, the intersection of this list with the upwind companies output by meteorological analysis experts is taken to obtain a list of key suspected companies that simultaneously meet the two conditions of "high ozone formation potential" and "located upwind". For particulate matter pollution scenarios, we identify the industry types with a high contribution rate from the contribution rates of each pollution source output by the component analysis experts, and then combine them with the set of upwind enterprises output by the meteorological analysis experts to screen out enterprises that belong to the industry type and are located upwind, and generate a list of key suspected enterprises. The contribution assessment uses a combination of qualitative description and quantitative ranking. The companies on the list are ranked according to their ozone generation potential or industry contribution rate, and their relative contribution level is marked (such as "high contribution", "medium contribution" and "low contribution").

[0079] The large language model directly references the contribution rate data of each pollution source output by component analysis experts, and outputs the industry proportion of pollution sources in a structured form. For ozone pollution scenarios, this proportion reflects the contribution percentage of ozone generation potential from different industry enterprises; for particulate matter pollution scenarios, this proportion reflects the percentage contribution of categories such as industrial sources, mobile sources, dust sources, and secondary conversion.

[0080] The large language model classifies sources based on the source apportionment results output by component analysis experts, combined with pollutant types. For particulate matter pollution, direct emission sources such as industrial sources, mobile sources, and dust sources are classified as "primary sources," while secondary transformations (such as sulfates, nitrates, and ammonium salts) are classified as "secondary sources." For ozone pollution, direct emissions of VOCs and NOx are classified as primary precursor emissions, while ozone generated by photochemical reactions is classified as secondary formation. The model outputs the proportion or qualitative description of primary and secondary sources.

[0081] After completing the aforementioned cross-dimensional fusion and logical reasoning, the large language model integrates the reasoning results into a structured comprehensive analysis conclusion. This conclusion includes at least the pollution transmission type (regional transmission / local generation), a list of key suspected enterprises (including enterprise name, contribution ranking, and relative contribution assessment), the industry share of pollution sources (presented as percentages or pie chart data), primary / secondary source apportionment results (qualitative descriptions or percentage data), and ozone sensitivity type (applicable only to ozone pollution scenarios, outputting VOCs control zone / NOx control zone / transition zone). This structured conclusion is ultimately used to generate a comprehensive source tracing report, which users can view and download online (Word / PDF).

[0082] The method for analyzing the causes and sources of air pollution based on a large language model in this embodiment has the following beneficial effects: 1. Significantly improved analysis efficiency: Through multi-threaded parallel data acquisition, multi-expert collaborative analysis, and automated report generation, the traditional source tracing cycle of several days is shortened to less than ten minutes, fully meeting the real-time requirements of pollution emergency response; 2. Significantly reduced professional threshold: Users can trigger complex source tracing analysis simply by using natural language, without needing to master model operation and professional data processing skills. The system automatically completes data query, model invocation, and result integration, achieving intelligent processing throughout the entire process. 3. Comprehensive and accurate analysis dimensions: It integrates multi-source data such as meteorological, component, and emission inventory data, and combines professional models such as HYSPLIT, EKMA, and PMF to achieve a comprehensive analysis of transmission path, source type, and contribution ratio; 4. High interpretability of results: Through logical reasoning of large language models and intuitive presentation of visual charts, structured data is transformed into easy-to-understand textual and graphical conclusions and management recommendations, directly supporting management decisions; 5. Excellent system scalability: The system adopts a layered architecture and modular design. Each functional module communicates through standardized interfaces, which can easily integrate new analysis models or connect to new data sources, and has good functional expansion capabilities. 6. Highly practical: Supports exporting reports in multiple formats such as Word and PDF. The report content is editable, fully adapting to the actual work needs of environmental management departments in report preparation and archiving.

[0083] Example 2 This embodiment provides an air pollution cause and source tracing analysis system, which executes the air pollution cause and source tracing analysis method as described in Embodiment 1. The system includes: The Natural Language Interaction Unit is used to provide a natural language interaction interface to obtain pollution source tracing requests input by users in natural language form.

[0084] The core unit of the large language model is trained on domain corpus and is used to perform semantic parsing on pollution source tracing requests to extract analysis elements. It also verifies the integrity of the extracted analysis elements and prompts the user to supplement information through the natural language interaction unit when information is missing. The core unit of the large language model also serves as the core controller for multi-expert collaborative analysis.

[0085] The parallel data acquisition unit is used to execute multiple data acquisition tasks in parallel according to the analysis elements. The multiple data acquisition tasks include: obtaining pollutant time series data from the environmental monitoring database through the NL2SQL module, obtaining meteorological time series data from the meteorological data source through the NL2API module, obtaining enterprise emission inventory data from the enterprise information database, and retrieving knowledge support sets in the fields of historical cases and policy standards from the vectorized knowledge base through retrieval enhancement generation technology.

[0086] The data fusion processing unit is used to perform multi-source heterogeneous data fusion processing on pollutant time series data, meteorological time series data, and enterprise emission inventory data. This includes: anomaly detection and missing data correction for time series data through multi-strategy fusion; correlation and alignment of the corrected data in terms of time and spatial dimensions; obtaining a structured dataset and constructing a comprehensive analysis view containing pollutant concentration, meteorological conditions, and emission source information; and storing the data in a hierarchical manner according to three levels: core, auxiliary, and reference.

[0087] The multi-expert collaborative analysis unit operates based on the ReAct framework. The core unit of the large language model acts as the core controller, dynamically calling multiple domain expert modules to analyze structured datasets. These multiple domain expert modules include at least: a meteorological analysis expert module for calling the HYSPLIT model for transmission path analysis, a component analysis expert module for calling the EKMA model for ozone pollution and the PMF model for particulate matter pollution for source apportionment, a visualization expert module for generating visualization charts, and a report generation expert module for integrating the output results of various experts.

[0088] The results integration and output unit is used to input the structured information integrated by the report generation expert module into the core unit of the large language model. The core unit of the large language model performs cross-result fusion and causal reasoning to generate a comprehensive analysis conclusion that includes pollution transmission type, list of key suspected enterprises, industry proportion of pollution sources, and primary and secondary source analysis results, and outputs a comprehensive source tracing report.

[0089] This embodiment uses an ozone pollution event that occurred at a monitoring station in Guangzhou on September 20, 2025, as the subject of source tracing analysis using this system. For example... Figure 3 As shown, the system has been deployed according to a five-layer architecture. The data layer stores hourly air quality data, hourly meteorological data, basic enterprise information, and emission inventories. The model layer integrates the BAAI / bge-m3 vector coding model, the Tencent HunyuanOCR text extraction model, and the Deep Seek large language model. The tool support layer encapsulates the NL2SQL conversion tool, the NL2API calling tool, the RAG retrieval tool, and the HYSPLIT trajectory calculation tool. The expert layer includes four expert modules: meteorological analysis, component analysis, visualization, and report generation. The application layer provides a web-based interactive interface.

[0090] The user enters "Analyze the cause of O3 pollution at a certain site in Guangzhou on September 20, 2025" in the dialog box of the web-based interactive interface. The natural language interaction unit forwards this request to the core unit of the large language model. The core unit of the large language model extracts the analysis elements through semantic parsing: the target area is a certain site in Guangzhou, the time range is the entire day of September 20, 2025, the target pollutant is O3, and no further information is required after completeness verification.

[0091] The parallel data acquisition unit initiates multi-threaded acquisition tasks based on the aforementioned analytical elements. The NL2SQL subunit generates SQL queries to retrieve the hourly O3 concentration sequence for the site on September 20th from the environmental monitoring database. The NL2API subunit calls the meteorological department's API interface to obtain hourly data on wind direction (°), wind speed (m / s), temperature (°C), and relative humidity (%) for the same period. The enterprise information database queries a list of all enterprises within a 5-kilometer radius of the site, including enterprise name, latitude and longitude, and industry, and further retrieves a list of VOCs emission components for these enterprises. The RAG retrieval subunit retrieves policy, regulatory, and technical guideline fragments related to "ozone source tracing" and "VOCs control" from a vectorized environmental knowledge base as background knowledge for subsequent analysis.

[0092] The data fusion processing unit cleans and aligns the collected multi-source data in terms of time and space. Specifically, this includes: removing outliers from monitoring stations whose O3 concentration exceeds three times the standard deviation; after removing outliers, establishing a unified time index for pollutant data, meteorological data, and enterprise emission data by hour, and constructing a comprehensive analysis view.

[0093] The multi-expert collaborative analysis engine initiates the analysis based on the ReAct framework. The large language model first determines the analysis scenario as ozone source tracing and plans to call the meteorological analysis expert module. This module receives station coordinates and time range parameters, calls the HYSPLIT model to simulate the forward trajectory after 72 hours, and cluster analysis identifies the prevailing wind direction as southeast, indicating regional transmission characteristics. The module calculates the azimuth angle relative to the station based on the enterprise's geographical coordinates, filters out enterprises located within ±30° of the prevailing wind direction, and performs a comprehensive analysis and ranking based on enterprise distance and emission intensity, selecting the 10 enterprises with the highest suspicion of contribution.

[0094] Based on meteorological analysis results, the large language model plans to call the component analysis expert module to perform ozone precursor analysis. The component analysis expert module receives a list of enterprises upwind, extracts VOCs emission data for the above 10 enterprises from the enterprise data, and calculates the ozone generation potential of each enterprise according to the MIR coefficient library. The results show that the OFP of enterprises A and C are significantly higher than those of other enterprises. At the same time, the EKMA model is called, and the measured VOCs and NOx concentration data are input to calculate the precursor ratio, determining that the area belongs to the VOCs-controlled zone.

[0095] Based on the comprehensive analysis of meteorological findings (southeast wind, regional transport), component analysis findings (VOCs-controlled type, with companies A and C having the highest OFP), and the spatial location of the companies (companies A and C are located southeast of the monitoring point), the large-scale language model deduced that the ozone pollution was a "complex type of local generation and regional transport," with companies A and C being the primary suspects. The final structured conclusion was: "This ozone pollution is a complex type of local generation and regional transport. Local photochemical processes contribute approximately 60%, mainly influenced by precursors brought by southeast winds. The key contributing companies are company A (chemical) and company C (coating), whose combined OFP contribution accounts for 45% of local precursors. It is recommended to strengthen the control of VOCs emissions from the above-mentioned companies, which can effectively reduce the local O3 formation potential."

[0096] The system then invokes the visualization expert module to generate ozone concentration-wind speed time series plots, backward trajectory plots, and bar charts ranking enterprise OFP contributions. The report generation expert module integrates structured conclusions, visualization charts, and policy evidence retrieved from RAG, generating a Word format source analysis report for users to view or download online. All data throughout the process is automatically standardized and externalized for storage. After execution, the tool returns a data ID that can be directly referenced in subsequent analyses, ensuring data traceability and reusability.

[0097] Example 3 This embodiment provides an electronic device. Figure 4 A structural block diagram of an electronic device according to an embodiment of the present invention is shown. Figure 4 As shown, the electronic device includes a memory 100 and a processor 200. The memory 100 stores a computer program that can run on the processor 200. When the processor 200 executes the computer program, it implements the air pollution cause and source tracing analysis method described in the above embodiments. The number of memories 100 and processors 200 can be one or more.

[0098] The electronic device also includes: The communication interface 300 is used to communicate with external devices and perform data exchange and transmission.

[0099] If the memory 100, processor 200, and communication interface 300 are implemented independently, they can be interconnected via a bus to communicate with each other. This bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc.

[0100] Optionally, in a specific implementation, if the memory 100, processor 200, and communication interface 300 are integrated on a single chip, then the memory 100, processor 200, and communication interface 300 can communicate with each other through an internal interface.

[0101] This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method provided in this invention.

[0102] This invention also provides a chip, which includes a processor for calling and executing instructions stored in a memory, causing a communication device on which the chip is installed to perform the method provided in this invention.

[0103] This invention also provides a chip, including: an input interface, an output interface, a processor, and a memory. The input interface, output interface, processor, and memory are connected through an internal connection path. The processor is used to execute code in the memory. When the code is executed, the processor is used to execute the method provided in this invention.

[0104] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting the Advanced Reduced Instruction Set Computing (RISC) machine (ARM) architecture.

[0105] Further, optionally, the aforementioned memory may include read-only memory and random access memory, and may also include non-volatile random access memory. The memory may be volatile or non-volatile, or may include both. Non-volatile memory may include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which serves as an external cache. Many forms of RAM are available by way of example, but not limitation. Examples include static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0106] In the above embodiments, implementation can be achieved, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another.

[0107] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0108] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for analyzing the causes and sources of air pollution, characterized in that, include: The system acquires pollution source tracing requests input in natural language form, performs semantic parsing on the pollution source tracing requests based on a large language model, and extracts analytical elements including target area, time range, and target pollutant. Based on the analysis elements, multiple data acquisition tasks are executed in parallel. The multiple data acquisition tasks include acquiring multi-source data associated with the pollution source tracing request from multiple data sources. The multi-source data includes pollutant time-series data, meteorological time-series data, and enterprise emission inventory data. The pollutant time-series data, the meteorological time-series data, and the enterprise emission inventory data are subjected to multi-source heterogeneous data fusion processing to generate a structured dataset; Based on a multi-expert collaborative analysis engine, multiple domain experts are dynamically scheduled to analyze the structured dataset. These multiple domain experts include meteorological experts who perform transmission path analysis and component analysis experts who perform source resolution. By integrating and fusing the analytical results of experts from various fields using the large language model, a comprehensive analytical conclusion including the sources and causes of pollution is generated.

2. The method for analyzing the causes and sources of air pollution according to claim 1, characterized in that, The multi-source heterogeneous data fusion processing of the pollutant time-series data, the meteorological time-series data, and the enterprise emission inventory data includes: Anomaly detection using a multi-strategy fusion method is performed on the pollutant time-series data and the meteorological time-series data to identify outliers or missing values; wherein, the multi-strategy fusion includes: statistical anomaly detection based on dynamic thresholds, anomaly determination based on multi-model voting, and rule verification based on national standard limits; The outliers or missing values ​​are corrected using a variety of interpolation strategies that match the data missing pattern to obtain corrected time series data; wherein the various interpolation strategies include linear interpolation, historical mean filling and weighted moving average; The corrected time-series data is correlated and aligned with the enterprise emission inventory data in both time and spatial dimensions to generate a structured dataset; Based on the structured dataset, a comprehensive analytical view containing pollutant concentration, meteorological conditions, and emission source information is constructed.

3. The method for analyzing the causes and sources of air pollution according to claim 1, characterized in that, The parallel execution of multiple data acquisition tasks includes: Based on NL2SQL technology, the analysis elements are converted into structured query statements to obtain the time series data of the pollutants from the environmental monitoring database; Based on NL2API technology, the analysis elements are parsed into API call parameters, and the meteorological time series data is obtained from the meteorological data source; Extract the latitude and longitude coordinates, industry, and emission inventory data of enterprises within a preset range surrounding the target area from the enterprise information database; the emission inventory data includes the emission intensity and composition information of pollutants.

4. The method for analyzing the causes and sources of air pollution according to claim 1, characterized in that, The meteorological experts' transmission path analysis method includes: The backward movement trajectory of the air mass was simulated based on the meteorological time-series data by invoking the Lagrange particle diffusion model. Cluster analysis was performed on multiple backward trajectories to identify the dominant transport pathways of pollutants; By combining the geographical coordinates of the enterprises in the enterprise emission inventory data, potential contributing source enterprises located upwind of the dominant transmission path are screened out.

5. The method for analyzing the causes and sources of air pollution according to claim 1, characterized in that, The source resolution method of the component analysis experts includes: When the target pollutant is ozone, the empirical kinetic simulation model is invoked to determine the sensitivity type of ozone formation based on the precursor concentration ratio in the time series data of the pollutant. Combined with the volatile organic compound component data in the enterprise emission inventory and the maximum incremental reactivity coefficient of each component, the ozone formation potential of each enterprise is calculated to obtain the ozone formation potential ranking result. When the target pollutant is particulate matter, the positive definite matrix factorization model is invoked to quantify and output the contribution rate of each pollution source based on the particulate matter component data in the pollutant time series data.

6. The method for analyzing the causes and sources of air pollution according to claim 1, characterized in that, By comprehensively reasoning and fusing the analytical results of experts from various fields using the aforementioned large language model, a comprehensive analytical conclusion including the sources and causes of pollution is generated, including: Based on the dominant transmission path output by meteorological experts, the transmission type of the pollution event is determined to be either regional transmission type or locally generated type. Based on the ozone formation potential ranking results or the contribution rate of each pollution source output by the component analysis experts, and combined with the spatial relationship between the geographical location of the enterprise and the dominant transmission path, a list of key suspected enterprises and a relative contribution assessment are generated. The transmission type of the pollution incident, the list of key suspected enterprises, and the contribution rate of each pollution source are integrated into a structured comprehensive analysis conclusion.

7. The method for analyzing the causes and sources of air pollution according to claim 1, characterized in that, The experts in these multiple domains also include visualization experts and report generation experts; The visualization expert is used to generate visualization charts including time series plots, trajectory plots, ranking plots, and distribution plots. The report generation expert is used to integrate the comprehensive analysis conclusions with the visualization charts into a structured analysis report.

8. A system for analyzing the causes and sources of air pollution, characterized in that, Performing the method for analyzing the causes and sources of air pollution as described in any one of claims 1 to 7 includes: The natural language interaction unit is used to obtain pollution source tracing requests input by users in natural language form; The core unit of the large language model, which is trained on a domain corpus, is used to perform semantic parsing on the pollution source tracing request to extract analytical elements. A parallel data acquisition unit is used to execute multiple data acquisition tasks in parallel according to the analysis elements. The multiple data acquisition tasks include at least obtaining pollutant time-series data from an environmental monitoring database, obtaining meteorological time-series data from a meteorological data source, and obtaining enterprise emission inventory data from an enterprise information database. The data fusion processing unit is used to perform multi-source heterogeneous data fusion processing on the pollutant time series data, the meteorological time series data, and the enterprise emission inventory data to generate a structured dataset; A multi-expert collaborative analysis unit is used to dynamically call multiple domain expert modules to analyze the structured dataset; the multiple domain expert modules include at least a meteorological analysis expert module for transmission path analysis and a component analysis expert module for source resolution; The result integration and output unit is used to input the analysis results of expert modules in various fields into the core unit of the large language model. The core unit of the large language model performs cross-result fusion and causal reasoning to generate a comprehensive analysis conclusion that includes the source and cause of pollution.

9. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores instructions that are loaded and executed by the processor to implement the method for analyzing the causes and sources of air pollution as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for analyzing the causes and sources of air pollution as described in any one of claims 1 to 7.

Images