CN-122022830-A - Machine learning-based pollutant and greenhouse gas dynamic tracing method and system

CN122022830ACN 122022830 ACN122022830 ACN 122022830ACN-122022830-A

Abstract

The invention discloses a pollutant and greenhouse gas dynamic tracing method and system based on machine learning, which relate to the technical field of environmental monitoring and data processing. According to the method, two different types of homologous substance data are utilized for cross verification, contradiction in independent analysis is intelligently solved through a deep fusion model, analysis uncertainty caused by factors such as atmospheric chemical reaction, background concentration noise or multi-source mixing of a single tracing method is effectively overcome, a tracing result which is closer to a real physical process and has high credibility is obtained, and quantitative evaluation of tracing conclusion reliability is realized through introducing a link of mutual authentication confidence evaluation and comprehensive credibility evaluation, so that follow-up environment supervision, law enforcement and treatment actions are more targeted and efficient.

Inventors

Xie Diesong
XIE FANGJIAN
TIAN FENG
ZHENG XINMEI
LIU CHUNLEI
Dou Daodao
WANG YAN
SUN YUEYIN

Assignees

南京市生态环境保护科学研究院

Dates

Publication Date: 20260512
Application Date: 20251217

Claims (10)

1. The pollutant and greenhouse gas dynamic tracing method based on machine learning is characterized by comprising the following steps: s1, synchronously acquiring pollutant concentration data, greenhouse gas concentration data and meteorological field data in a target area to form a multi-mode synchronous data set; s2, inputting pollutant concentration data in the multi-mode synchronous data set and corresponding meteorological field data into a pollutant traceability analysis model to generate a first traceability result, and simultaneously, inputting greenhouse gas concentration data in the multi-mode synchronous data set and corresponding meteorological field data into a greenhouse gas traceability analysis model to generate a second traceability result; S3, performing mutual evidence consistency analysis on the first tracing result and the second tracing result to generate a mutual evidence confidence score; s4, determining and triggering a subsequent processing flow based on a comparison result of the mutual authentication confidence coefficient score and a confidence coefficient threshold value interval; S5, when a deep fusion tracing flow is triggered, carrying out optimization calculation through a coupling tracing model based on the first tracing result, the second tracing result, the multi-mode synchronous data set and a symbiotic emission knowledge base so as to simultaneously and optimally fit concentration distribution of pollutants and greenhouse gases as targets on the premise of meeting optimization constraint conditions and emission activity rules and generate a fusion tracing result; and S6, carrying out comprehensive credibility assessment on the finally determined emission source according to the mutual authentication confidence score or the analysis quality index of the fusion traceability result, and outputting a traceability conclusion with credibility assessment.
2. The machine learning based pollutant and greenhouse gas dynamic tracing method of claim 1, wherein said performing an inter-evidence consistency analysis on said first tracing result and said second tracing result, generating an inter-evidence confidence score, comprises: Extracting first main candidate source space-time information and a first inferred emission activity type from the first tracing result, wherein the first main candidate source space-time information comprises geographic coordinates and an active time window; Extracting second main candidate source space-time information and a second inferred emission activity type from the second tracing result; calculating the space-time coincidence degree of the first main candidate source space-time information and the second main candidate source space-time information, wherein the space-time coincidence degree is specifically determined through a distance on a geographic space and an overlapping interval on a time sequence; querying the symbiotic emission knowledge base by taking the first inferred emission activity type and the second inferred emission activity type as query conditions, judging that the association of the first inferred emission activity type and the second inferred emission activity type is consistent if the first inferred emission activity type and the second inferred emission activity type point to the same source or are defined as a strong symbiotic relationship, otherwise judging that the association is inconsistent, and thus obtaining a judging result of type consistency; and based on the space-time coincidence degree and the association coincidence degree, carrying out weighted summation calculation after normalization processing to generate the mutual evidence confidence score.
3. The machine learning based pollutant and greenhouse gas dynamic tracing method of claim 1, wherein said determining and triggering a subsequent process flow comprises: If the mutual authentication confidence score is higher than an upper threshold value in the confidence threshold value interval, triggering a direct evaluation flow, and performing comprehensive credibility evaluation based on the first tracing result and the second tracing result; if the cross-correlation confidence score is lower than a lower threshold value in the confidence threshold value interval, triggering a deep fusion tracing flow; And if the mutual authentication confidence score is between the upper limit threshold and the lower limit threshold, packaging the current tracing case data, the first tracing result, the second tracing result and the mutual authentication confidence score to form a sample to be learned, and storing the sample to be learned into a case learning pool.
4. The method for dynamically tracing the source of the pollutant and the greenhouse gas based on the machine learning according to claim 1, wherein the optimizing calculation is performed through the coupling tracing model to generate the fusion tracing result, and the method comprises the following steps: formalizing inconsistent information between the first tracing result and the second tracing result to construct a set of optimization constraint conditions; Providing the multimodal synchronization data set, emission activity rules stored in the symbiotic emission knowledge base, and the optimization constraint conditions as inputs to the coupling traceability model; And operating the coupling traceability model to solve the concentration distribution of the best fitting pollutant and greenhouse gas as a target on the premise of meeting the optimization constraint condition and the emission activity rule, and outputting the fusion traceability result.
5. The method for dynamically tracing a contaminant and a greenhouse gas according to claim 4, further comprising the step of evolving knowledge after outputting the result of the fusion tracing: Evaluating an interpretation degree index of the fusion traceability result on the multi-mode synchronous data set, and if the interpretation degree index exceeds a preset interpretation degree index threshold and the interaction confidence score corresponding to the fusion traceability result is in a preset new experience generation interaction confidence score interval, generating new experience knowledge about specific emission activity and characteristic substance combination; evaluating the validity of the new experience knowledge to obtain the evaluation confidence of the new experience knowledge; And if the evaluation confidence exceeds a preset updating threshold, integrating and updating the new experience knowledge to the symbiotic emission knowledge base.
6. The machine learning based pollutant and greenhouse gas dynamic tracing method of claim 5, further comprising the following knowledge batch evolution step based on the case learning pool: Triggering a batch learning flow when the number of the samples to be learned stored in the case learning pool reaches a preset batch learning threshold; Extracting all or part of the sample to be learned from the case learning pool, and mining implicit emission pattern relations in the sample to be learned through a machine learning algorithm to generate batch updated knowledge; and updating the batch updated knowledge to the symbiotic emission knowledge base after auditing.
7. The machine learning based pollutant and greenhouse gas dynamic tracing method of claim 6, wherein when updating new experience knowledge or batch updated knowledge to the symbiotic emission knowledge base, further comprising: adding the new empirical knowledge or the batch updated knowledge as candidate knowledge items to the symbiotic emission knowledge base; Associating initial confidence weights for the candidate knowledge items according to the evaluation confidence degrees of the new experience knowledge or the evaluation results when the knowledge is updated in batches; and in subsequent traceability analysis, weighting and applying different knowledge items in the symbiotic emission knowledge base according to the initial confidence weight.
8. The machine learning-based pollutant and greenhouse gas dynamic tracing method according to claim 1, wherein the pollutant tracing analysis model and the greenhouse gas tracing analysis model are obtained by training by the following methods: acquiring pollution source list data, greenhouse gas emission list data and corresponding historical meteorological data in a historical period; generating a historical pollutant concentration distribution data set based on the pollution source list data and the historical meteorological data by utilizing an atmospheric physical model in a forward simulation mode; the historical pollutant concentration distribution data set and the corresponding historical meteorological data are used as training input, the pollution source list data are used as training targets, the pollutant traceability analysis model is obtained through training, and the greenhouse gas traceability analysis model is obtained through training in a similar mode.
9. The machine learning based pollutant and greenhouse gas dynamic tracing method of claim 1, wherein said integrated credibility assessment comprises: When the subsequent processing flow is a direct evaluation flow, the credibility evaluation is to call the mutual authentication confidence score and directly convert the mutual authentication confidence score into the credibility grade according to a preset mapping rule; And when the subsequent processing flow is the deep fusion tracing flow, the credibility assessment is obtained by integrating an optimized objective function convergence value of the coupling tracing model, a data fitting residual error of the fusion tracing result on the multi-mode synchronous data set, and a weighted summation of the matching degree of the fusion tracing result and the updated symbiotic emission knowledge base.
10. A system for performing the machine learning based contaminant and greenhouse gas dynamic tracing method of claims 1-9, comprising: The data synchronization and preprocessing module is used for executing the steps of synchronously acquiring and forming a multi-mode synchronous data set; the parallel tracing analysis module is configured with a pollutant tracing analysis model and a greenhouse gas tracing analysis model and is used for receiving the multi-mode synchronous data set and generating a first tracing result and a second tracing result; The cross-certification evaluation and dynamic decision module is used for receiving the first tracing result and the second tracing result, generating a cross-certification confidence score, and determining and triggering a subsequent processing flow according to the score; the depth fusion tracing module is configured with a coupling tracing model and is used for generating a fusion tracing result when triggered by the interaction evaluation and dynamic decision module; The symbiotic knowledge base and knowledge evolution engine is used for storing emission activity rules and executing knowledge evolution or batch evolution steps so as to dynamically update the emission activity rules; And the credibility synthesis and output module is used for carrying out credibility assessment on the traceability conclusion and outputting the credibility assessment.

Description

Machine learning-based pollutant and greenhouse gas dynamic tracing method and system Technical Field The invention relates to the technical field of environmental monitoring and data processing, in particular to a pollutant and greenhouse gas dynamic tracing method and system based on machine learning. Background The atmospheric traceability analysis refers to a technology for reversely tracing and identifying pollutant or greenhouse gas emission sources and contributions thereof by analyzing atmospheric environment monitoring data and combining weather information and a physicochemical model. With the development of environmental monitoring technology, the utilization of advanced computing models such as machine learning and the like to process massive environmental data so as to improve the efficiency and the accuracy of traceability analysis has become an important development direction in the field. The technology is widely applied to the fields of regional air quality management, climate change treatment, environmental law enforcement and the like, and has important significance for realizing accurate pollution control and scientific emission reduction. The prior art discloses a greenhouse gas point source emission automatic identification method, a system, equipment and a medium disclosed by the patent application publication No. CN120298897B, and an atmospheric pollutant tracing system and a method based on big data technology disclosed by the publication No. CN114757687A, and the conventional tracing technology is usually implemented by using the tracing of conventional pollutants and greenhouse gases as two independent technical tasks. For example, the tracing of pollutants such as sulfur dioxide and nitrogen oxides mainly depends on the reverse calculation based on an atmospheric-chemical diffusion model, while the tracing of greenhouse gases such as carbon dioxide and methane focuses more on the estimation and spatial localization of the emission flux. In these processes, the analytical flow, data input and model system of each type of material are mutually split, running independently within respective frameworks, although it is also possible to apply some machine learning algorithm to optimize model parameters or to perform pattern recognition. However, the above prior art has obvious technical drawbacks in practical applications. Firstly, the isolated analysis mode ignores the physical fact that pollutants and greenhouse gases are emitted in symbiosis in a large amount of emission activities, so that information is not fully utilized, two independent analysis clues cannot form mutual evidence or constraint, and uncertainty of tracing results is large when facing complex mixed sources or weak signals. Second, prior methods use prior knowledge of emissions factors, which are typically static and cannot accommodate dynamic changes in emissions characteristics in the real world. Finally, when different analysis methods draw contradictory conclusions for the same area, the prior art lacks an effective and systematic contradiction solving mechanism, often relies on manual judgment, and the objectivity and reliability of the result are difficult to ensure. Disclosure of Invention The invention aims to provide a pollutant and greenhouse gas dynamic tracing method and system based on machine learning, which solve the problems in the background technology. In order to solve the problems, the first aspect of the invention provides a pollutant and greenhouse gas dynamic tracing method based on machine learning, which adopts a dynamic decision and knowledge evolution feedback mechanism based on mutual evidence confidence assessment, can realize intelligent verification and dynamic correction of emission source tracing results, and improves the accuracy, reliability and self-adaptation capability of tracing analysis, and comprises the steps of S1, synchronously acquiring pollutant concentration data, greenhouse gas concentration data and meteorological field data in a target area to form a multi-mode synchronous data set. And S2, inputting the pollutant concentration data in the multi-mode synchronous data set and the corresponding meteorological field data into a pollutant traceability analysis model to generate a first traceability result, and simultaneously, inputting the greenhouse gas concentration data in the multi-mode synchronous data set and the corresponding meteorological field data into a greenhouse gas traceability analysis model to generate a second traceability result. And S3, performing mutual evidence consistency analysis on the first tracing result and the second tracing result to generate a mutual evidence confidence score. And S4, determining and triggering a subsequent processing flow based on a comparison result of the mutual authentication confidence coefficient score and a confidence coefficient threshold interval. And S5, when the deep fusion tracing flow is triggered, carrying out