CN-122024919-A - Ozone pollution dynamic prediction and control area discrimination method and system based on mixed sign-numerical reasoning

CN122024919ACN 122024919 ACN122024919 ACN 122024919ACN-122024919-A

Abstract

The invention relates to the technical field of atmospheric environment intelligent treatment and discloses a method and a system for dynamically predicting ozone pollution and distinguishing a control area based on mixed sign-numerical reasoning; the method comprises the steps of constructing a knowledge graph facing the field of atmospheric photochemical reaction, wherein the knowledge graph comprises chemical species nodes and edges with temperature-irradiance dependent reaction rates, taking the knowledge graph as chemical dynamics constraint, taking real-time monitoring data as input, guiding a large model to conduct end-to-end prediction on ozone peak concentration, introducing a chemical dynamics consistency penalty term, reversely updating reaction rate weights in the knowledge graph based on residual errors of prediction errors and observed values to form a closed-loop optimization mechanism, and judging that the current atmospheric environment belongs to a VOC control area, an NO x control area or a transition area by combining uncertainty of large model prediction and VOC dominant reaction path confidence of the knowledge graph. The invention realizes the credible discrimination of the ozone generation dominant factor on the principle level.

Inventors

YE JIAHAO
LIU GANG
ZHU TONGTONG
LIU YANG

Assignees

合肥中科环境监测技术国家工程实验室有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (7)

1. The ozone pollution dynamic prediction and control area discrimination method based on mixed sign-numerical reasoning is characterized by comprising the following steps: Constructing a knowledge graph oriented to the field of atmospheric photochemical reaction, wherein the knowledge graph comprises chemical species nodes and edges with temperature-irradiance dependent reaction rates; Taking the knowledge graph as chemical kinetics constraint, taking real-time monitoring data as input, guiding the large model to predict the ozone peak concentration end to end, and introducing a chemical kinetics consistency penalty term into a loss function of the training large model; based on the residual error of the prediction error and the observed value, reversely updating the reaction rate weight in the knowledge graph to form a closed-loop optimization mechanism; And judging that the current atmospheric environment belongs to the VOC control area, the NO x control area or the transition area by combining uncertainty of large model prediction and the VOC dominant reaction path confidence of the knowledge graph.
2. The method for ozone pollution dynamic prediction and control area discrimination based on mixed sign-numerical reasoning of claim 1, wherein said set of nodes of knowledge graph , , Represents the collection of chemical species involved in the atmospheric photochemical reaction system, Represents the 1 st to nth volatile organic compounds, Is the general term for the nitrogen oxides, Represents the nitrogen dioxide and is used to treat the waste gas, Represents a hydroxyl radical and is used to form a hydroxyl radical, Represents a radical of a peroxy alkyl group, Represents ozone; Representing an ith chemical species node in the knowledge graph, corresponding to one chemical species in S; Edge collection ; The temperature is set to be the absolute temperature, Represents ultraviolet irradiance; Reaction Rate The arrhenius-photolcoupled form is adopted: ; To be from chemical species node Conversion to chemical species node The pre-finger factor of the corresponding chemical reaction, To be from chemical species node Conversion to chemical species node The activation energy of the corresponding chemical reaction, To be from chemical species node Conversion to chemical species node The photosensitivity coefficient of the corresponding chemical reaction, R is the gas constant, As a function of the efficiency of the photolysis, Is the critical wavelength.
3. The method for distinguishing the ozone pollution dynamic prediction and control area based on mixed sign-numerical reasoning according to claim 2, wherein the knowledge graph is taken as a chemical kinetics constraint, real-time monitoring data is taken as input, a large model is guided to conduct end-to-end prediction on the ozone peak concentration, and a chemical kinetics consistency penalty term is introduced into a loss function, and the method specifically comprises the following steps: Vector of real-time monitoring data ; In order to be of a relative humidity level, Is the air pressure of the air, and is the air pressure of the air, Index for the t-th time step; the loss function The method comprises the following steps: ; for a true peak concentration of ozone, In order to predict the peak concentration of ozone, In order to be able to learn the balance coefficients, Generating differential equations for ozone generated from the knowledge graph: ; respectively carrying out advection conveying items and dry sedimentation items, and interpolating by a meteorological field; For the purpose of the ozone generation reaction set, Is that One of the ozone generation reactions in the process, For ozone generation reactions Is used for the ozone to be used for the ozone, For ozone generation reactions Is used for the rate constant of (a), For ozone generation reactions The concentration product of the ozone precursor involved, For the set of ozone depletion reactions, Is that One of the ozone depletion reactions in the (a) is, For ozone depletion reactions Is used for the ozone to be used for the ozone, For ozone depletion reactions Is a rate constant of (c).
4. The method for distinguishing the ozone pollution dynamic prediction and control area based on mixed sign-numerical reasoning according to claim 3, wherein the reverse updating of the reaction rate weight in the knowledge graph based on the residual error of the prediction error and the observed value comprises the following steps: Defining residuals The method comprises the following steps: ; Reaction rate sensitivity was calculated by back propagation through the graph neural network: ; Wherein, the Is the stoichiometric derivative; The reaction rate is modified by adopting an exponential decay update rule: ; Wherein, the For self-adaptive learning rate, satisfy Gamma is the attenuation factor, and the gamma is the attenuation factor, For the initial rate of learning to be the same, Is the updated reaction rate.
5. The method for dynamically predicting ozone pollution and distinguishing control areas based on mixed sign-value reasoning as recited in claim 4, wherein the node updating process of the graph neural network comprises the following steps: ; ; Is a chemical species node Is used to update the embedded vector of the block, As a function of the non-linear activation, Is a chemical species node A first-order neighbor chemical species node set in the knowledge-graph, To be from a neighbor chemical species node To a central chemical species node Is used for the concentration weight of the person, Is a neighbor node The input at the first layer of the neural network embeds the vector, As a matrix of learnable weights for the graph neural network, To chemical species node The attention scores of all chemical species nodes of (a) are normalized, As a function of LeakyReLU, A learnable attention vector in an attention mechanism for drawing; is the chemical species node in the knowledge graph Contains information on concentration, reactivity and uncertainty.
6. The method for distinguishing the ozone pollution dynamic prediction and control area based on mixed sign-numerical reasoning according to claim 1, wherein the method is characterized in that the method combines uncertainty of large model prediction and confidence of VOC dominant reaction path of knowledge graph to distinguish that the current atmospheric environment belongs to a VOC control area, a NO x control area or a transition area, and specifically comprises the following steps: The large model adopts Monte Carlo Dropout to output a prediction interval and corresponding uncertainty The method comprises the following steps: ; an upper limit value for the ozone peak concentration prediction output by the large model at the t-th time step, A lower limit value of the ozone peak concentration prediction output by the big model in the t-th time step; Calculating VOC dominant path confidence based on knowledge-graph : ; ; Wherein, the Is a hydroxyl radical activity correction factor, Is a set of VOC-limited reactions, As an intermediate variable, the number of the variables, For all reaction paths associated with ozone generation, Is the total concentration of volatile organic compounds; Control zone discriminant function The definition is as follows: ; Is that The degree of confidence of the dominant path, Is the VOC sensitivity threshold coefficient, Is that The sensitivity threshold coefficient is used to determine, For the VOC path confidence discrimination threshold, Is that The path confidence level discrimination threshold value, Is the total concentration of the nitrogen oxides, Is the predicted peak concentration of ozone.
7. An ozone pollution dynamic prediction and control area discrimination system based on mixed sign-numerical reasoning, which is characterized by comprising: the domain knowledge graph construction module is used for constructing a knowledge graph oriented to the atmospheric photochemical reaction domain, wherein the knowledge graph comprises chemical species nodes and edges with temperature-irradiance dependent reaction rates; The reasoning engine module takes the knowledge graph as chemical dynamics constraint, takes real-time monitoring data as input, guides the large model to predict the peak concentration of ozone end to end, and introduces a chemical dynamics consistency penalty term into a loss function; the closed loop feedback optimization module is used for reversely updating the reaction rate weight in the knowledge graph based on the prediction error and the residual error of the observed value to form a closed loop optimization mechanism; And the control area intelligent judging module is used for judging that the current atmospheric environment belongs to the VOC control area, the NO x control area or the transition area by combining uncertainty of large model prediction and the VOC leading reaction path confidence of the knowledge graph.

Description

Ozone pollution dynamic prediction and control area discrimination method and system based on mixed sign-numerical reasoning Technical Field The invention relates to the technical field of atmospheric environment intelligent treatment, in particular to a method and a system for dynamically predicting ozone pollution and distinguishing a control area based on mixed sign-numerical reasoning. Background Ozone (O 3) is a major secondary pollutant in the troposphere, which produces a highly nonlinear mechanism that relies on the coupling of Volatile Organic Compounds (VOCs), nitrogen oxides (NO x), solar radiation, temperature and radical concentration. The traditional EKMA (EMPIRICAL KINETIC Modeling Approach) curve simulates ozone peaks under different VOC/NO x combinations through a fixed chemical mechanism, but has three defects of (1) simplified chemical mechanism, neglected free radical chain reaction and VOC species specificity, (2) static meteorological conditions, incapability of adapting to hour level change, and (3) incapability of supporting dynamic emission reduction decision. In recent years, large models (such as GALACTICA, CHATWEATHER) have been tried for environmental prediction, but are "black box" models, the output lacks chemical rationality constraints, and the Knowledge Graph (KG) can encode chemical mechanisms, but the traditional knowledge graph (such as Wikidata) cannot express the relation between dynamic rate and differential equation. Existing search enhancement generation (RAG) combined large language model schemes only support qualitative questions (e.g. "what is generated by ozone. More importantly, the existing method does not establish a prediction-feedback-correction closed loop, namely when the model prediction and actual measurement deviate, the parameters (such as reaction rate) of the chemical mechanism of the bottom layer cannot be automatically corrected, so that the system cannot adapt to new pollution sources (such as novel industrial solvents) or mechanism deviation. Disclosure of Invention In order to solve the technical problems, the invention provides a method and a system for judging an ozone pollution dynamic prediction and control area based on mixed sign-numerical reasoning. In order to solve the technical problems, the invention adopts the following technical scheme: in a first aspect, the invention provides a method for discriminating an ozone pollution dynamic prediction and control area based on mixed sign-numerical reasoning, which comprises the following steps: Constructing a knowledge graph oriented to the field of atmospheric photochemical reaction, wherein the knowledge graph comprises chemical species nodes and edges with temperature-irradiance dependent reaction rates; Taking the knowledge graph as chemical kinetics constraint, taking real-time monitoring data as input, guiding the large model to predict the ozone peak concentration end to end, and introducing a chemical kinetics consistency penalty term into a loss function of the training large model; based on the residual error of the prediction error and the observed value, reversely updating the reaction rate weight in the knowledge graph to form a closed-loop optimization mechanism; And judging that the current atmospheric environment belongs to the VOC control area, the NO x control area or the transition area by combining uncertainty of large model prediction and the VOC dominant reaction path confidence of the knowledge graph. In one embodiment, the node set of the knowledge-graph,,Represents the collection of chemical species involved in the atmospheric photochemical reaction system,Represents the 1 st to nth volatile organic compounds,Is the general term for the nitrogen oxides,Represents the nitrogen dioxide and is used to treat the waste gas,Represents a hydroxyl radical and is used to form a hydroxyl radical,Represents a radical of a peroxy alkyl group,Represents ozone; Representing an ith chemical species node in the knowledge graph, corresponding to one chemical species in S; Edge collection ;The temperature is set to be the absolute temperature,Represents ultraviolet irradiance; Reaction Rate The arrhenius-photolcoupled form is adopted: ; To be from chemical species node Conversion to chemical species nodeThe pre-finger factor of the corresponding chemical reaction,To be from chemical species nodeConversion to chemical species nodeThe activation energy of the corresponding chemical reaction,To be from chemical species nodeConversion to chemical species nodeThe photosensitivity coefficient of the corresponding chemical reaction, R is the gas constant,As a function of the efficiency of the photolysis,Is the critical wavelength. In one embodiment, the knowledge graph is used as a chemical kinetics constraint, real-time monitoring data is used as input, a large model is guided to conduct end-to-end prediction on the ozone peak concentration, and a chemical kinetics consistency penalty term