Ozone pollution dynamic prediction and control area discrimination method and system based on mixed symbol-numerical reasoning

Abstract

The invention relates to the technical field of atmospheric environment intelligent treatment, and discloses an ozone pollution dynamic prediction and control area discrimination method and system based on mixed symbol-numerical reasoning. The method comprises the following steps: constructing a knowledge graph oriented to the atmospheric photochemical reaction field, wherein the knowledge graph comprises chemical species nodes and edges with temperature-irradiance dependent reaction rates; taking the knowledge graph as a chemical kinetics constraint, taking real-time monitoring data as input, guiding the large model to carry out end-to-end prediction on the ozone peak concentration, and introducing a chemical kinetics consistency penalty term; reversely updating the reaction rate weight in the knowledge graph based on the prediction error and the residual error of the observation value to form a closed-loop optimization mechanism; and judging whether the current atmospheric environment belongs to a VOC control region, a NOx control region or a transition region by combining the uncertainty of large model prediction and the VOC dominant reaction path confidence of the knowledge graph. According to the invention, credible discrimination of ozone generation dominant factors on the principle level is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent atmospheric environmental governance technology, specifically to a method and system for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning. Background Technology

[0002] Ozone (O3), as a major secondary pollutant in the troposphere, has a highly nonlinear formation mechanism, depending on volatile organic compounds (VOCs) and nitrogen oxides (NOx). x The coupling effects of solar radiation, temperature, and free radical concentration are considered. Traditional EKMA (Empirical Kinetic Modeling Approach) curves simulate different VOC / NO concentrations by fixing the chemical mechanism. x The combined ozone peak has three major drawbacks: (1) the chemical mechanism is simplified, ignoring free radical chain reactions and VOC species specificity; (2) the meteorological conditions are static and cannot adapt to hourly changes; and (3) it cannot support dynamic emission reduction decisions.

[0003] In recent years, large-scale models (such as Galactica and ChatWeather) have been attempted for environmental prediction, but they are "black box" models, and their outputs lack chemical rationality constraints. Although knowledge graphs (KGs) can encode chemical mechanisms, traditional knowledge graphs (such as Wikidata) cannot express the relationship between dynamic rates and differential equations. Existing retrieval-enhanced generation (RAG) schemes combined with large language models only support qualitative question answering (such as "What is ozone generated?") and cannot output quantitative predictions (such as "The ozone peak in the next 6 hours will be 152 ppb").

[0004] More importantly, existing methods have not established a "prediction-feedback-correction" closed loop: when there is a deviation between model prediction and actual measurement, it is impossible to automatically correct the underlying chemical mechanism parameters (such as reaction rate), which leads to the system's inability to adapt to new pollution sources (such as new industrial solvents) or mechanism deviations. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning, including: A knowledge graph for atmospheric photochemical reactions is constructed, which includes chemical species nodes and edges with temperature-irradiance dependent reaction rates. The knowledge graph is used as a chemical kinetic constraint, and real-time monitoring data is used as input to guide the large model to make end-to-end predictions of ozone peak concentration. A chemical kinetic consistency penalty term is introduced into the loss function of the large model training. Based on the residual between the prediction error and the observed value, the reaction rate weights in the knowledge graph are updated in reverse to form a closed-loop optimization mechanism. By combining the uncertainties in large-scale model predictions with the confidence levels of VOC-dominated response pathways from knowledge graphs, it is determined whether the current atmospheric environment belongs to a VOC-controlled area and a NO-controlled area. x Control zone or transition zone.

[0007] In one embodiment, the knowledge graph's node set , , This represents the collection of chemical species involved in the atmospheric photochemical reaction system. This represents the first to nth volatile organic compounds. It is a general term for nitrogen oxides. It represents nitrogen dioxide. Represents hydroxyl radicals. It represents a peroxyalkyl radical. Indicates ozone; This represents the i-th chemical species node in the knowledge graph, corresponding to a chemical species in S; set of edges ; Absolute temperature Indicates ultraviolet irradiance; reaction rate Using Arrhenius-optical decoupling: ; To start from chemical species nodes Transformation into chemical species nodes The pre-exponential factor of the corresponding chemical reaction, To start from chemical species nodes Transformation into chemical species nodes The activation energy of the corresponding chemical reaction To start from chemical species nodes Transformation into chemical species nodes The photosensitivity coefficient corresponding to the chemical reaction, where R is the gas constant. Let be the optical resolution efficiency function. This is the critical wavelength.

[0008] In one embodiment, the knowledge graph is used as a chemical kinetic constraint, real-time monitoring data is used as input, and a large model is guided to perform end-to-end prediction of ozone peak concentration. A chemical kinetic consistency penalty term is introduced into the loss function, specifically including: Vector of real-time monitoring data ; Relative humidity, Atmospheric pressure This is the index of the t-th time step; The loss function for: ; This represents the actual peak ozone concentration. For the predicted peak ozone concentration, For learnable balance coefficients, The ozone formation differential equation generated from the knowledge graph: ; These are advection transport and dry deposition terms, obtained by meteorological field interpolation. It is a set of ozone formation reactions. for One of the ozone formation reactions, Ozone generation reaction The stoichiometric coefficient of ozone, Ozone generation reaction rate constant, Ozone generation reaction The product of the concentrations of ozone precursors involved, This is a set of ozone-depleting reactions. for One of the ozone-depleting reactions in the process, Ozone depletion reaction The stoichiometric coefficient of ozone, Ozone depletion reaction The rate constant.

[0009] In one embodiment, the step of updating the reaction rate weights in the knowledge graph in reverse based on the residual between the prediction error and the observations specifically includes: Define residual for: ; Calculate reaction rate sensitivity using backpropagation via a graph neural network: ; in, It is the stoichiometric derivative; The reaction rate is corrected using the exponential decay update rule: ; in, For adaptive learning rate, satisfy γ is the attenuation factor. The initial learning rate, This represents the updated reaction rate.

[0010] In one embodiment, the node update process of the graph neural network includes: ; ; For chemical species nodes The updated embedding vector, It is a non-linear activation function. For chemical species nodes The set of first-order neighbor chemical species nodes in a knowledge graph To obtain from neighboring chemical species nodes To the central chemical species node Attention weights For neighboring nodes The input embedding vector of the l-th layer of the graph neural network The learnable weight matrix of a graph neural network, For chemical species nodes The attention scores of all chemical species nodes are normalized. For the LeakyReLU function, Learnable attention vectors in graph attention mechanisms; Chemical species nodes in a knowledge graph The embedding vector contains information on concentration, reactivity, and uncertainty.

[0011] In one embodiment, the method combines the uncertainty of large model predictions with the confidence of VOC-dominant response pathways from a knowledge graph to determine whether the current atmospheric environment belongs to a VOC control area and NO... x Control zone or transition zone, specifically including: The large model uses Monte Carlo Dropout to output the prediction interval, corresponding to uncertainty. for: ; This represents the upper bound of the ozone peak concentration prediction output by the large model at time step t. This is the lower bound of the ozone peak concentration prediction output by the large model at time step t; Calculating the confidence of VOC dominant path based on knowledge graph : ; ; in, It is a hydroxyl radical activity corrector. It is a VOC-limited reaction set. As an intermediate variable, This is the set of all reaction pathways related to ozone formation. This refers to the total concentration of volatile organic compounds. Control area discrimination function Defined as: ; for Dominant path confidence This is the VOC sensitivity threshold coefficient. for Sensitivity threshold coefficient The threshold for VOC path confidence discrimination. for Path confidence threshold The total concentration of nitrogen oxides, The predicted peak ozone concentration; where, calculations are based on a knowledge graph. Dominant path confidence : ; .

[0012] Secondly, this invention provides a dynamic ozone pollution prediction and control zone discrimination system based on hybrid symbolic-numerical reasoning, comprising: Domain knowledge graph construction module: Constructs a knowledge graph for the field of atmospheric photochemical reactions, which includes chemical species nodes and edges with temperature-irradiance dependent reaction rates; Inference engine module: The knowledge graph is used as a chemical kinetic constraint, real-time monitoring data is used as input, and the large model is guided to make end-to-end predictions of ozone peak concentration. A chemical kinetic consistency penalty term is introduced into the loss function. Closed-loop feedback optimization module: Based on the residual between the prediction error and the observed value, the reaction rate weights in the knowledge graph are updated in reverse to form a closed-loop optimization mechanism; Intelligent control area discrimination module: Combining the uncertainty of large model predictions with the confidence of VOC-dominant response pathways from a knowledge graph, it determines whether the current atmospheric environment belongs to a VOC control area or a NO control area.x Control zone or transition zone.

[0013] Compared with the prior art, the beneficial technical effects of the present invention are: This invention fundamentally ensures the accuracy of prediction results in accordance with physicochemical laws by using symbolic chemical mechanism knowledge as a structural constraint of a large numerical model. By leveraging a residual-driven knowledge graph parameter back-optimization mechanism, the system gains the ability to dynamically adapt to changes in the atmospheric environment and pollution sources, thus overcoming the limitations of static models. Finally, by fusing data-driven prediction uncertainties with path confidence derived from mechanisms, a reliable determination of the dominant factors in ozone formation is achieved at the principle level. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0015] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] The purpose of this invention is to provide a hybrid symbolic-numerical reasoning method that integrates domain knowledge graphs and large scientific models. By explicitly encoding atmospheric chemical mechanisms, constraining the output of large models, and introducing closed-loop feedback optimization, it achieves high-precision, interpretable, and self-evolving ozone pollution simulation and control zone identification.

[0017] The method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning in this invention specifically includes: S1. Construct a domain-specific knowledge graph (KG) for atmospheric photochemical reactions. Nodes represent chemical species (e.g., isoprene, ·OH, O3), edges represent chemical reaction pathways, and edge weights are temperature-irradiance dependent reaction rates. ; S2. Design a hybrid symbolic-numerical reasoning engine to encode the knowledge graph into the ozone generation differential equation C( This serves as a penalty term in the loss function of the large model, ensuring that the prediction results satisfy chemical kinetics. S3. Define the prediction residual and calculate it through backpropagation using a differentiable graphical neural network. And dynamically correct This forms a closed-loop optimization. S4. Prediction uncertainty based on the output of the large model Confidence of VOC-dominated path in knowledge graph Construct a control area discrimination function to determine whether the current atmospheric environment belongs to a VOC control area or a NO control area. x The control zone or transition zone is selected, and the emission reduction priority is output.

[0018] The knowledge graph KG=(N,E) was automatically extracted from Master Chemical Mechanism (MCM) v3.3.1 and contains 3,562 species nodes and 15,284 reaction edges.

[0019] For the photochemical reaction NO2 + hν (λ < 420 nm) → NO + O, the reaction rate is defined as: ; Where σ is the absorption cross section. For quantum yield, This is for the actual measurement of solar spectral irradiance.

[0020] The graph supports dynamic injection: when a new VOC species (such as butyl acrylate) is detected by online monitoring, the system automatically obtains its ·OH reaction rate constant from the PubChem database and adds a new node and edge.

[0021] This invention employs a finely tuned large-scale scientific model (based on the Galactica-1.3B architecture), whose input is a 12-dimensional vector. It includes: 8 sets of VOC component concentrations (ppbC); 2 sets of NOx data (NO, NO2); meteorological parameters (T, RH, ... P).

[0022] This invention introduces a chemical kinetic consistency penalty: ; in, Approximated by numerical differentiation: ( ) / Δt, Δt=1h.

[0023] To achieve self-evolution of the knowledge graph, this invention treats the knowledge graph as a differentiable graph. Reaction rate. It is parameterized as a learnable scalar and backpropagation is achieved through PyTorch Geometric.

[0024] Experiments show that the mechanism can automatically correct for deviations caused by changes in industrial emissions within 48 hours.

[0025] This invention is the first to integrate uncertainty with mechanism confidence: if and If so, it is determined to be a high-confidence VOC control region; if and If so, it is determined to be a transition zone, and it is recommended that "VOCs and NOs" be separated. x "Coordinated emission reduction".

[0026] The judgment result was directly output as an emission reduction instruction: "It is recommended to shut down the painting enterprise in Area A (VOC contribution rate of 42%), which is expected to reduce O3 by 18 ppb."

[0027] Example: This embodiment identifies the source of the initial chemical mechanism for knowledge graph construction, preferably using the internationally recognized MasterChemical Mechanism (MCM) v3.3.1. This mechanism contains more than 6,700 species and 17,000 reaction pathways, covering the reaction network of major VOCs (such as alkanes, alkenes, aromatic hydrocarbons, and oxygen-containing VOCs) and free radicals (·OH, NO3, RO2) in urban atmosphere.

[0028] refer to Figure 1 This embodiment presents a method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning, comprising the following steps: S1, constructing a domain knowledge graph oriented towards atmospheric photochemical reactions; S2, designing a hybrid symbolic-numerical reasoning engine to perform ozone concentration prediction constrained by chemical mechanisms; S3, dynamically correcting the reaction rate parameters in the knowledge graph based on the prediction residuals through a differentiable graph propagation mechanism; S4, combining prediction uncertainty and path confidence to determine the control zone type of the current atmospheric environment and generate emission reduction recommendations.

[0029] In this embodiment, S1 includes the following steps: S11, identifying and selecting authoritative atmospheric chemical mechanism sources, preferably using the internationally recognized Master Chemical Mechanism (MCM) v3.3.1, which covers detailed reaction pathways of major volatile organic compounds, nitrogen oxides, and free radicals in urban atmosphere; S12, automatically extracting triplet data from the MCM mechanism to construct initial knowledge graph nodes and edges, wherein node types include volatile organic compound species nodes such as isoprene and toluene, oxidant nodes such as hydroxyl radicals and ozone, intermediate product nodes such as peroxyalkyl radicals and formaldehyde, and final product nodes such as carbon dioxide and nitric acid, and edges represent chemical reaction processes, each edge carrying attributes such as reaction equation, Arrhenius parameters, photolysis threshold wavelength, and photosensitivity coefficient, for example, the reaction equation isoprene reacting with hydroxyl radicals to generate oxidation products, and Arrhenius parameters... The parameters include pre-exponential factors and activation energies, used to calculate temperature-dependent reaction rates; S13, access real-time data from the urban environmental monitoring network, obtain a list of volatile organic compound components actually present in the air through an online volatile organic compound monitor, if a component is not in the initial graph, the system automatically calls the PubChem application programming interface to obtain its molecular structure, and estimates its reaction rate constant with hydroxyl radicals based on a quantitative structure-activity relationship model, adding corresponding nodes and reaction edges to the knowledge graph; S14, store the constructed knowledge graph in a graph database, each reaction rate parameter is encapsulated as a trainable floating-point variable, and a fast index is established according to reaction type and control region sensitivity to support subsequent differentiable optimization and path lookup.

[0030] In this embodiment, S2 includes the following steps: S21, deploying a large-scale scientific model, with the encoder portion of Galactica-1.3B used as the basic architecture. The input layer is modified into a twelve-dimensional floating-point vector, containing the concentrations of eight key volatile organic compounds, nitrogen oxides, temperature, relative humidity, ultraviolet irradiance, and atmospheric pressure; S22, implementing a chemical consistency constraint module, which extracts the ozone formation differential equation from the knowledge graph in real time. The equation is in the form that the ozone formation rate equals the sum of the contributions of each precursor oxidation pathway minus the sum of the consumption pathways, plus advection and dry deposition terms. The concentration of hydroxyl radicals is dynamically solved from the knowledge graph using a steady-state approximation method; S23, during the model training phase, the loss function consists of two parts: one part is the mean square error between the predicted and measured ozone values, and the other part is the mean square error between the predicted ozone change rate and the output value of the chemical differential equation. The two are weighted and summed using learnable weight coefficients; S24, the model is trained using hourly daily ozone pollution data from a certain city from 2020 to 2023. After training, the data is exported in ONNX format and deployed on an edge server. The inference latency is controlled within two seconds, meeting the hourly prediction requirements.

[0031] In this embodiment, S3 includes the following steps: S31. Obtain the measured ozone value from the national monitoring station every hour, calculate the residual between it and the model prediction value, and trigger the feedback optimization process if the absolute value of the residual exceeds 10 ppb; S32. Construct the knowledge graph into a differentiable graph structure, and use a graph neural network framework to set the rate parameter of each reaction edge as a learnable tensor; S33. Perform backpropagation with the residual as the loss function, and automatically calculate the gradient of each reaction rate with respect to the residual; S34. Correct the reaction rate using an exponential decay update rule, that is, the new rate is equal to the old rate multiplied by the natural exponential function with the negative learning rate and the gradient as the exponent, and the initial learning rate is set to 0.01 and decays over time to prevent over-adjustment; S35. Set the single update amplitude to not exceed ±10%, and the cumulative update over 72 hours to not exceed ±30%, to ensure that the chemical mechanism is not distorted due to over-correction.

[0032] In this embodiment, S4 includes the following steps: S41. During inference, enable Monte Carlo Dropout, run twenty forward propagations to obtain the ozone prediction set, calculate its 95% confidence interval, and define the half-width of the interval as the prediction uncertainty; S42. Traverse all ozone generation paths in the knowledge graph, screen out volatile organic compound (VOC)-limited reactions, calculate the weighted path confidence, and the weight is jointly determined by the reaction rate, precursor concentration, and free radical activity; S43. Set a preset control region discrimination threshold. If the confidence of a VOC path is greater than 0.7 and the prediction uncertainty is less than 8 ppb, it is determined to be a VOC control region. If the confidence of a nitrogen oxide (NOx) path is greater than 0.65 and the uncertainty is less than 8 ppb, it is determined to be a VOC control region. If the emissions exceed ppb, the area is designated as a nitrogen oxide control zone; otherwise, it is designated as a transition zone. S44. Within the volatile organic compound (VOC) control zone, each enterprise is ranked according to its contribution rate to ozone formation, and a list of emission reduction recommendations is generated. For example, it is recommended that painting enterprises limit production by 50% and printing enterprises stagger their production schedules. S45. The ozone prediction curve, control zone type identifiers, and enterprise emission reduction priority heatmaps are visualized through a web interface to provide decision support for environmental protection departments.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0034] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0035] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0036] Based on the description of the above method embodiments, the present invention also provides a system. The system may be a system that uses the methods, software (applications), modules, components, servers, clients, etc., described in the embodiments of this specification, combined with necessary implementation hardware. Since the implementation schemes and methods for solving the problem are similar, the specific system implementations in the embodiments of this specification can be referred to the implementations of the foregoing methods, and repeated details will not be repeated. As used below, the term "module" or "module group" refers to a combination of software and / or hardware capable of performing a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0037] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0038] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0039] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning, characterized in that, include: A knowledge graph for atmospheric photochemical reactions is constructed, which includes chemical species nodes and edges with temperature-irradiance dependent reaction rates. The knowledge graph is used as a chemical kinetic constraint, and real-time monitoring data is used as input to guide the large model to make end-to-end predictions of ozone peak concentration. A chemical kinetic consistency penalty term is introduced into the loss function of the large model training. Based on the residual between the prediction error and the observed value, the reaction rate weights in the knowledge graph are updated in reverse to form a closed-loop optimization mechanism. By combining the uncertainties in large-scale model predictions with the confidence levels of VOC-dominated response pathways from knowledge graphs, it is determined whether the current atmospheric environment belongs to a VOC-controlled area and a NO-controlled area. x Control zone or transition zone.

2. The method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning according to claim 1, characterized in that, The set of nodes in the knowledge graph , , This represents the collection of chemical species involved in the atmospheric photochemical reaction system. This represents the first to nth volatile organic compounds. It is a general term for nitrogen oxides. It represents nitrogen dioxide. Represents hydroxyl radicals. It represents a peroxyalkyl radical. Indicates ozone; This represents the i-th chemical species node in the knowledge graph, corresponding to a chemical species in S; set of edges ; Absolute temperature Indicates ultraviolet irradiance; reaction rate Using Arrhenius-optical decoupling: ； To start from chemical species nodes Transformation into chemical species nodes The pre-exponential factor of the corresponding chemical reaction, To start from chemical species nodes Transformation into chemical species nodes The activation energy of the corresponding chemical reaction To start from chemical species nodes Transformation into chemical species nodes The photosensitivity coefficient corresponding to the chemical reaction, where R is the gas constant. Let be the optical resolution efficiency function. This is the critical wavelength.

3. The method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning according to claim 2, characterized in that, Using the knowledge graph as a chemical kinetic constraint and real-time monitoring data as input, the large model is guided to perform end-to-end prediction of ozone peak concentration. A chemical kinetic consistency penalty term is introduced into the loss function, specifically including: Vector of real-time monitoring data ; Relative humidity, Atmospheric pressure This is the index of the t-th time step; The loss function for: ； This represents the actual peak ozone concentration. For the predicted peak ozone concentration, For learnable balance coefficients, The ozone formation differential equation generated from the knowledge graph: ； These are advection transport and dry deposition terms, obtained by meteorological field interpolation. It is a set of ozone formation reactions. for One of the ozone formation reactions, Ozone generation reaction The stoichiometric coefficient of ozone, Ozone generation reaction rate constant, Ozone generation reaction The product of the concentrations of ozone precursors involved, This is a set of ozone-depleting reactions. for One of the ozone-depleting reactions in the process, Ozone depletion reaction The stoichiometric coefficient of ozone, Ozone depletion reaction The rate constant.

4. The method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning according to claim 3, characterized in that, The method of updating the reaction rate weights in the knowledge graph in reverse based on the residual between the prediction error and the observed values ​​specifically includes: Define residual for: ； Calculate reaction rate sensitivity using backpropagation via a graph neural network: ； in, It is the stoichiometric derivative; The reaction rate is corrected using the exponential decay update rule: ； in, For adaptive learning rate, satisfy γ is the attenuation factor. The initial learning rate, This represents the updated reaction rate.

5. The method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning according to claim 4, characterized in that, The node update process of the graph neural network includes: ； ； For chemical species nodes The updated embedding vector, It is a non-linear activation function. For chemical species nodes The set of first-order neighbor chemical species nodes in a knowledge graph To obtain from neighboring chemical species nodes To the central chemical species node Attention weights Neighboring nodes The input embedding vector of the l-th layer of the graph neural network The learnable weight matrix of a graph neural network, For chemical species nodes The attention scores of all chemical species nodes are normalized. For the LeakyReLU function, Learnable attention vectors in graph attention mechanisms; Chemical species nodes in a knowledge graph The embedding vector contains information on concentration, reactivity, and uncertainty.

6. The method for dynamic prediction and control zone identification of ozone pollution based on hybrid symbolic-numerical reasoning according to claim 1, characterized in that, The method combines the uncertainty of large model predictions with the confidence of VOC-dominated response pathways from the knowledge graph to determine whether the current atmospheric environment belongs to a VOC-controlled area and NO-controlled area. x Control zone or transition zone, specifically including: The large model uses Monte Carlo Dropout to output the prediction interval, corresponding to uncertainty. for: ； This represents the upper bound of the ozone peak concentration prediction output by the large model at time step t. This is the lower bound of the ozone peak concentration prediction output by the large model at time step t; Calculating the confidence of VOC dominant path based on knowledge graph : ； ； in, It is a hydroxyl radical activity corrector. It is a VOC-limited reaction set. As an intermediate variable, This is the set of all reaction pathways related to ozone formation. This refers to the total concentration of volatile organic compounds. Control area discrimination function Defined as: ； for Dominant path confidence This is the VOC sensitivity threshold coefficient. for Sensitivity threshold coefficient The threshold for VOC path confidence discrimination. for Path confidence threshold The total concentration of nitrogen oxides, This represents the predicted peak ozone concentration.

7. A dynamic prediction and control zone discrimination system for ozone pollution based on hybrid symbolic-numerical reasoning, characterized in that, include: Domain knowledge graph construction module: Constructs a knowledge graph for the field of atmospheric photochemical reactions, which includes chemical species nodes and edges with temperature-irradiance dependent reaction rates; Inference engine module: The knowledge graph is used as a chemical kinetic constraint, real-time monitoring data is used as input, and the large model is guided to make end-to-end predictions of ozone peak concentration. A chemical kinetic consistency penalty term is introduced into the loss function. Closed-loop feedback optimization module: Based on the residual between the prediction error and the observed value, the reaction rate weights in the knowledge graph are updated in reverse to form a closed-loop optimization mechanism; Intelligent control area discrimination module: Combining the uncertainty of large model predictions with the confidence of VOC-dominant response pathways from a knowledge graph, it determines whether the current atmospheric environment belongs to a VOC control area or a NO control area. x Control zone or transition zone.

Images