Factory soil and groundwater pollution risk management and control method

Abstract

The invention discloses a factory soil and groundwater pollution risk management and control method. The method comprises the following steps: acquiring a factory data set at a current moment and a historical moment; establishing a risk prediction model based on the factory data set; dividing risk grade distribution in the factory based on the risk prediction model and repairing the risk grade distribution; for the high-risk area, in-situ chemical oxidation and a permeable reactive barrier are adopted for repairing; phytoremediation or microbial remediation is adopted for medium risk areas; a low-risk area does not need to be repaired; the method can accurately predict the risk grade distribution in the factory according to the geological parameters, pollutant monitoring data and environmental dynamics of the factory, realizes risk grading output by constructing a risk prediction model fusing a physical mechanism and machine learning advantages, improves the accuracy and repair efficiency of risk assessment, and improves the risk assessment efficiency. And the pertinence, the economical efficiency and the environmental benefits of pollution control are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of water treatment, and specifically relates to a method for controlling the pollution risks of factory area soil and groundwater. Background Art

[0002] At present, the methods for controlling the pollution risks of soil and groundwater mainly include traditional physical, chemical and biological remediation technologies: Physical remediation separates or immobilizes pollutants through means such as soil replacement, electrokinetic remediation, and thermal desorption, which is applicable to shallow pollution but has a high cost; Chemical remediation changes the form of pollutants through reactions such as chemical leaching, oxidation / reduction, and solidification / stabilization, with significant effects but may cause secondary pollution; Biological remediation repairs pollution through natural processes such as phytoextraction and microbial degradation, with low cost but a long cycle; The selection of technologies needs to comprehensively consider the type of pollutants, remediation cycle, cost, and secondary pollution risks. The future trend is to use combined remediation, application of nanomaterials, and in-situ remediation technologies to improve efficiency and reduce environmental impacts;

[0003] In the prior art, the risk control steps include source control, end treatment, pollution monitoring, and emergency response. Source control reduces the risk of pollutant leakage through measures such as optimizing equipment layout, anti-corrosion and anti-seepage treatment, and reducing the laying of underground pipelines; End treatment prevents the spread of pollutants through strict management of zoned anti-seepage, leakage collection systems, and hazardous waste temporary storage sites; The pollution monitoring system relies on groundwater pollution monitoring wells and a regular monitoring system to achieve early detection and disposal of pollution. However, these methods have obvious limitations: Source control depends on the active fulfillment of responsibilities by enterprises, but some enterprises may not implement the measures effectively due to insufficient understanding of regulations or lack of motivation; The anti-seepage layer in end treatment may leak due to aging or construction defects after long-term use, and the layout density and monitoring frequency of pollution monitoring wells are difficult to cover all hidden pollution sources; Although emergency response can control the spread of pollution, it cannot completely eliminate the pollutants that have penetrated into the soil and groundwater.

[0004] In addition, the existing risk control methods have limitations in predicting and evaluating the migration of pollution. Traditional risk assessment often relies on limited sampling data and empirical models, and it is difficult to accurately predict the dynamic changes of pollutants in soil and groundwater, which leads to a high degree of uncertainty and conservatism in the risk assessment results, and may cause waste of resources or insufficient remediation measures. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a method for controlling the pollution risks of factory area soil and groundwater.

[0006] A method for controlling the pollution risks of factory area soil and groundwater includes the following steps:

[0007] Obtain the factory area dataset at the current moment and historical moments; the factory area dataset includes geological parameters, pollutant monitoring data, and environmental dynamic data;

[0008] Based on the factory area dataset, establish a risk prediction model. The inputs of the risk prediction model are geological parameters, pollutant monitoring data, and environmental dynamic data, and the output of the risk prediction model is the risk level distribution within the factory area. The risk level distribution within the factory area includes high-risk areas, medium-risk areas, and low-risk areas; the risk prediction model includes a physical migration module for generating training data and physical constraint conditions from the input dataset, a learning and prediction module for performing spatio-temporal feature operations based on the training data and physical constraint conditions, and a dynamic decision-making module for risk grading output;

[0009] Based on the risk prediction model, divide and repair the risk level distribution within the factory area; for high-risk areas, use in-situ chemical oxidation method and permeable reactive barrier method for short-term repair; for medium-risk areas, use phytoremediation method or bioremediation method for medium- and long-term repair; low-risk areas do not require repair.

[0010] Note: The above method can accurately predict the risk level distribution within the factory area according to the geological parameters, pollutant monitoring data, and environmental dynamic data of the factory area. By constructing a risk prediction model that integrates the advantages of physical mechanisms and machine learning, it accurately quantifies the spatial distribution of the factory area risk level. Based on the physical migration module to strengthen data reliability, the learning and prediction module to capture spatio-temporal evolution laws, and the dynamic decision-making module to achieve risk grading output, this method not only improves the accuracy of risk assessment and repair efficiency, but also can dynamically adjust the repair strategy, thus providing a scientific basis for formulating an efficient repair plan of in-situ chemical oxidation and permeable reactive barrier for high-risk areas, and adopting eco-friendly phytoremediation or bioremediation technologies for medium-risk areas, while avoiding over-interference in low-risk areas, and finally forming a full-chain control system of "data-driven - risk grading - precise repair", significantly enhancing the pertinence, economy, and environmental benefits of pollution control.

[0011] Further, the geological parameters include permeability coefficient and porosity, the pollutant monitoring data includes pollutant concentration; the environmental dynamic data includes meteorological data and hydrological dynamic data.

[0012] Note: The above explanation clarifies the specific content of the factory area dataset, providing comprehensive basic information for establishing a risk prediction model.

[0013] Further, the method for establishing a risk prediction model based on the dataset includes:

[0014] Construct a physical migration module: Select the HYDRUS software that can couple the simulation of saturated-unsaturated water flow, solute transport, and multiple processes. According to the plant dataset, set up a migration equation in the HYDRUS software to obtain a physical migration module with the output being the concentration distribution of pollutants in soil and groundwater, the source strength of pollutants, and the diffusion rate.

[0015] Construct a learning and prediction module: Design a spatio-temporal attention graph neural network. The spatio-temporal attention graph neural network uses a multi-head self-attention mechanism to capture long-range dependencies, uses a dynamic graph convolution module to adaptively adjust the weights between nodes, and adds a mass conservation regularization term to the loss function. Train the learning and prediction module based on the concentration distribution of pollutants in soil and groundwater and environmental data to obtain a learning and prediction module. The output of the learning and prediction module is the concentration distribution of pollutants at future times.

[0016] Based on the concentration distribution of pollutants at future times obtained from the learning and prediction module, divide the high-risk area, medium-risk area, and low-risk area.

[0017] Note: The above method uses the HYDRUS software to construct a physical migration module, accurately coupling the saturated-unsaturated water flow and solute transport processes, ensuring that the simulation results of pollutant concentration distribution, source strength, and diffusion rate conform to physical laws, providing a reliable data basis for the model. Using a spatio-temporal attention graph neural network as the learning and prediction module, capturing the long-range spatio-temporal dependencies of pollutant migration through a multi-head self-attention mechanism, and realizing the dynamic division of risk levels based on the prediction results, which not only ensures the sensitivity of high-risk area identification but also optimizes the resource allocation in medium- and low-risk areas, providing a technical support that combines science and practicality for plant pollution prevention and control.

[0018] Furthermore, the method for dividing the high-risk area, medium-risk area, and low-risk area based on the concentration distribution of pollutants at future times obtained from the learning and prediction module includes: When the predicted concentration of the concentration distribution of pollutants at future times is less than or equal to 50% of the risk threshold, it is classified as a low-risk area; when the predicted concentration of the concentration distribution of pollutants at future times is greater than 50% and less than or equal to 75% of the risk threshold, it is classified as a medium-risk area; when the predicted concentration of the concentration distribution of pollutants at future times is greater than 75%, it is classified as a high-risk area.

[0019] Note: The above method maps the future pollutant concentration prediction results directly to high, medium, and low risk areas by setting clear and quantitative classification criteria, avoiding the arbitrariness of subjective judgment and realizing the gradient differentiation of risk levels through the classification thresholds, ensuring that the high-risk area can accurately lock in the core area of pollution diffusion to give priority to strong intervention measures, the medium-risk area can implement dynamic monitoring and ecological restoration for potential risks, and the low-risk area is exempt from excessive treatment.

[0020] Further, the risk threshold is the environmental risk screening value.

[0021] Note: The above environmental risk screening value is taken from the "Soil Environmental Quality - Risk Control Standards for Soil Pollution of Construction Land (Trial)" (GB 36600 - 2018), which realizes the seamless connection between risk assessment and the current environmental management standards, and ensures the compliance of risk classification by using the legal authority of the screening value.

[0022] Further, the environmental data includes groundwater level data, surrounding vegetation distribution data, and land use change and irrigation data.

[0023] Note: The above method takes into account key environmental data such as groundwater level, surrounding vegetation distribution, and land use change and irrigation, constructs a multi - dimensional risk prediction input system, reflects the impact of hydrogeological conditions on pollutant migration, reveals the natural purification ability of the ecosystem and pollution exposure risk through vegetation distribution data, and captures the dynamic driving effect of human activities on pollution diffusion by combining land use change and irrigation data, thus significantly improving the comprehensiveness and accuracy of risk assessment.

[0024] Further, in the HYDRUS software, different rainfall boundary conditions are set according to rainfall data in different seasons; the relationship between air temperature and soil temperature is established through empirical formulas or experimental data, and then the corresponding soil temperature boundary is set in the software.

[0025] Note: By setting seasonal rainfall boundary conditions and soil temperature boundaries, the spatio - temporal simulation accuracy of the model for pollution migration process is significantly improved: adjusting the boundary conditions based on rainfall data in different seasons can accurately depict the impact of seasonal differences in rainfall infiltration intensity and frequency on pollutant leaching and diffusion.

[0026] Further, the method further includes:

[0027] Determine the repair cycle. When the repair cycle is completed, obtain the plant area dataset after the completion of the repair cycle, and use the plant area dataset after the completion of the repair cycle to update the risk prediction model;

[0028] Based on the risk prediction model, predict new high - risk areas, medium - risk areas, and low - risk areas until the risk level distribution within the plant area is all low - risk, then the control is completed.

[0029] Description: The above method realizes the real-time feedback and iterative optimization of the risk prediction model on the actual repair effect by setting the repair cycle and regularly obtaining the data of the repaired plant area, avoiding the prediction deviation caused by the static model due to environmental changes; continuously predicting and adjusting the risk area based on the updated model to ensure that the control measures always focus on the current high-risk points until the overall risk is reduced to a low-risk level, thus avoiding the waste of resources caused by over-repair.

[0030] Further, the method for determining the repair cycle is to select multiple repair cycles. Among the multiple repair cycles, the numerical range of recent repair is 0.1 - 1 year, and the numerical range of medium- and long-term repair is 0.5 - 5 years. Use the risk prediction model to predict the risk level area of pollutants after each cycle, and select the repair cycle within the range of 0.1 - 5 years that makes the risk level area all close to or all be low-risk areas.

[0031] Description: By presetting multiple candidate cycles within the range of 0.1 - 5 years and using the risk prediction model to quantify the spatial distribution of pollution risks after each cycle, it not only avoids the subjectivity of setting a single cycle, but also uses the degree of approach of the risk level area to the low-risk area as the screening basis to ensure that the selected cycle can balance the repair efficiency and cost and maximize the reduction of environmental risks.

[0032] The beneficial effects of the present invention are:

[0033] The method of the present invention can accurately predict the risk level distribution in the plant area according to the geological parameters, pollutant monitoring data and environmental dynamic data of the plant area. By constructing a risk prediction model that integrates the advantages of physical mechanisms and machine learning, it accurately quantifies the spatial distribution of the risk level in the plant area. Based on the physical migration module to strengthen data reliability, the learning prediction module to capture the spatio-temporal evolution law, and the dynamic decision-making module to achieve risk classification output, this method not only improves the accuracy of risk assessment and repair efficiency, but also can dynamically adjust the repair strategy, thus providing a scientific basis for formulating efficient repair schemes for in-situ chemical oxidation and permeable reactive barriers in high-risk areas, and using eco-friendly plant or microbial repair technologies in medium-risk areas, while avoiding over-interference in low-risk areas, and finally forming a full-chain control system of "data-driven - risk classification - precise repair", significantly improving the pertinence, economy and environmental benefits of pollution control. Description of the Drawings

[0034] Figure 1 It is a schematic structural diagram of the risk prediction model of the embodiment of the present invention;

[0035] Figure 2 is The MIP test PID results of some monitoring points in the plant area in the embodiment of the present invention;

[0036] Figure 3It is the PID result of MIP test at some monitoring points in the factory area in the embodiment of the present invention;

[0037] Figure 4 It is the PID result of MIP test at some monitoring points in the factory area in the embodiment of the present invention;

[0038] Figure 5 It is the PID result of MIP test at some monitoring points in the factory area in the embodiment of the present invention;

[0039] Figure 6 It is the PID result of MIP test at some monitoring points in the factory area in the embodiment of the present invention;

[0040] Figure 7 It is the PID result of MIP test at some monitoring points in the factory area in the embodiment of the present invention;

[0041] Figure 8 It is the schematic diagram of the method flow for recent repair in the embodiment of the present invention. Detailed implementation manners

[0042] To further elaborate on the methods and effects adopted by the present invention, the technical solutions of the present invention will be clearly and completely described below in combination with experiments.

[0043] Since traditional control methods usually adopt a one - size - fits - all approach for control, which cannot meet the requirements of different pollution scenarios, and in the existing pollutant remediation process, some pollutants need to be remediated in the soil to significantly reduce the risk, but some pollutants automatically reduce the risk to soil and groundwater over time without the need for remediation. Therefore, accurately dividing the risk areas can greatly improve the remediation and control effects and reduce unnecessary remediation processes.

[0044] In response to this, the embodiments of the present invention construct a risk prediction model including elements such as pollutant source characteristics, hydrogeological conditions, and receptor exposure paths, which can quantitatively evaluate the pollution diffusion probability and potential hazards in each area, and then divide high, medium, and low risk levels.

[0045] For example, in high - risk areas, in - situ remediation technologies (such as enhanced bioremediation, nano - zero - valent iron injection) need to be preferentially used to block pollution diffusion. In medium - risk areas, exposure risks can be reduced through engineering controls (such as impermeable barriers) combined with institutional management. In low - risk areas, monitoring is mainly carried out. This hierarchical control strategy can not only optimize resource allocation but also achieve precise prevention and control of pollution risks, avoiding over - remediation or insufficient treatment, and providing a scientific basis for the sustainable operation of the factory area.

[0046] Embodiment 1: A method for controlling the pollution risk of factory area soil and groundwater, comprising the following steps:

[0047] S1. Obtain the plant area data sets at the current moment and historical moments; the plant area data sets include geological parameters, pollutant monitoring data, and environmental dynamic data;

[0048] The above geological parameters include permeability coefficient and porosity, and the pollutant monitoring data includes pollutant concentration; the environmental dynamic data includes meteorological data and hydrological dynamic data;

[0049] Exemplarily, at the current moment (December 2024, the current moment in the following embodiments of this example refers to December 2024), the permeability coefficient of the plant area is 1.2×10 -4 , and the porosity is 32, which is obtained through experiments; at the historical moment (March 2024, the historical moment in the following embodiments of this example refers to March 2024), the permeability coefficient of the plant area is 1.0×10 -4 , and the porosity is 30, which is obtained through experiments;

[0050] The meteorological dynamic data includes temperature, humidity, wind speed, and precipitation; at the current moment, the temperature is 8°C, the humidity is 65%, the wind speed is 3.2 m / s, and the precipitation is 2.0 mm; at the historical moment, the temperature is 15°C, the humidity is 50%, the wind speed is 2.8 m / s, and the precipitation is 3.4 mm; at the current moment, the groundwater level is 5.2 m, and the surface water flow velocity is 0.15 m 3 / s; at the historical moment, it is 4.9 m, and the surface water flow velocity is 0.12 m 3 / s;

[0051] As Figure 2, Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 shown, the pollutant concentration data in the pollutant monitoring data is as follows: the target pollutant in the above plant area is ethylbenzene, and the monitoring points of the concentration include the MIP01 - MIP20 monitoring points shown in Figures 2 - 7. The monitoring results include: there are obvious peaks between approximately 1 - 3.2 meters for PID, and the maximum value reaches 5×107 μV, and a certain amount of pollutants has accumulated;

[0052] The values at the MIP11 point are relatively high at 1.6 meters and 2.2 meters, and there may be pollutants;

[0053] The PID value of MIP18 at 18 points is relatively high between 1 and 2 meters, and there may be pollutants. The PID values of other MIP measurement points are relatively low; the PID values of MIP17 at 6 meters and 8 meters are relatively high, and there may be pollutants; the PID value of MIP18 at 2.6 meters is relatively high, and there may be pollutants. The FID values of other MIP measurement points are relatively low; the XSD of MIP11 has an obvious peak between about 2 and 9 meters, and the detected value is greater than 0.5×106 μV, indicating that a certain amount of pollutants should have accumulated. The XSD of MIP18 has an obvious peak between 1 and 3 meters, indicating that a certain amount of pollutants should have accumulated.

[0054] The XSD values of other MIP measurement points are relatively low; the PID, FID, and XSD detection values of MIP20 are relatively small.

[0055] S2. Based on the plant area dataset, establish a risk prediction model (as Figure 1 shown), the input of the risk prediction model is geological parameters, pollutant monitoring data, and environmental dynamic data, and the output of the risk prediction model is the risk level distribution within the plant area. The risk level distribution within the plant area includes high-risk areas, medium-risk areas, and low-risk areas; the risk prediction model includes a physical migration module for generating training data and physical constraint conditions from the input dataset, a learning prediction module for performing spatio-temporal feature operations based on the training data and physical constraint conditions, and a dynamic decision-making module for risk classification output;

[0056] The method for establishing a risk prediction model based on the dataset includes:

[0057] S2-1. Construct a physical migration module: Select the HYDRUS software (3D software) that can couple and simulate saturated-unsaturated water flow, solute transport, and multiple processes. According to the plant area dataset, set up a migration equation in the HYDRUS software to obtain a physical migration module with the output of the concentration distribution of pollutants in soil and groundwater, source strength, and diffusion rate;

[0058] Specifically, the water flow equation in the migration equation is the Richards equation to simulate saturated-unsaturated water flow; the solute transport equation in the migration equation is the advection-dispersion equation (ADRE) to simulate the migration of ethylbenzene;

[0059] In the boundary conditions, the surface is rainfall infiltration (Neumann boundary), and the bottom is free drainage (Dirichlet boundary, pressure head = 0);

[0060] S2-2. Construct a learning and prediction module: Design a spatio-temporal attention graph neural network, which uses a multi-head self-attention mechanism to capture long-range dependencies, an adaptive dynamic graph convolution module to adjust the weights between nodes, and adds a mass conservation regularization term to the loss function. Train the learning and prediction module based on the concentration distribution of pollutants in soil and groundwater and environmental data to obtain the learning and prediction module, and the output of the learning and prediction module is the concentration distribution of pollutants at future times.

[0061] Convert the 3D grid output by HYDRUS into a graph structure, with each grid node being a node in the graph. The node features include: geological parameters (permeability coefficient, porosity); pollutant concentration (at the current time step); environmental dynamic data (rainfall, temperature encoded as a time series); slice by hour to generate time series graph data (e.g., 7 days × 24 hours = 168 time steps).

[0062] Use a learnable similarity matrix to calculate edge weights based on node features (such as concentration gradient, distance). Through multi-head self-attention (in the time dimension): capture dependencies at different time steps, such as the delayed impact of rainfall events on pollution diffusion; identify long-range associations between high-risk areas (such as nodes that are far apart but hydraulically connected). Loss function: mean squared error (MSE) to compare predictions with true concentrations.

[0063] The above step S2-2 is implemented through three tools: PyTorch Geometric (graph neural network), DGL (dynamic graph convolution), and Transformer (attention).

[0064] S2-3. Based on the concentration distribution of pollutants at future times obtained from the learning and prediction module, divide into high-risk areas, medium-risk areas, and low-risk areas. The method includes: when the predicted concentration of the pollutant concentration distribution at future times is less than or equal to 50% of the risk threshold, it is classified as a low-risk area; when the predicted concentration of the pollutant concentration distribution at future times is greater than 50% and less than or equal to 75% of the risk threshold, it is classified as a medium-risk area; when the predicted concentration of the pollutant concentration distribution at future times is greater than 75%, it is classified as a high-risk area. The risk threshold is the environmental risk screening value, which is 28 mg / kg in the "Soil Environmental Quality - Risk Control Standards for Soil Pollution of Construction Land (Trial)" (GB 36600-2018). It is implemented through the ArcGIS API (spatial analysis).

[0065] The environmental data includes groundwater level data, surrounding vegetation distribution data, and land use change and irrigation data.

[0066] In the HYDRUS software, different rainfall boundary conditions are set according to rainfall data in different seasons; the relationship between air temperature and soil temperature is established through empirical formulas or experimental data, and then the corresponding soil temperature boundary is set in the software.

[0067] S3. Divide the risk level distribution within the factory area based on the risk prediction model and conduct remediation; use in-situ chemical oxidation method and permeable reactive barrier method for short-term remediation; for medium-risk areas, use phytoremediation method or microbial remediation method for medium- and long-term remediation; low-risk areas do not require remediation; the method further includes:

[0068] Determine the remediation cycle. After the remediation cycle is completed, obtain the factory area dataset after the completion of the remediation cycle, and use the factory area dataset after the completion of the remediation cycle to update the risk prediction model;

[0069] Based on the risk prediction model, predict new high-risk areas, medium-risk areas, and low-risk areas until the risk level distribution within the factory area is all low-risk, then the control is completed.

[0070] The above method for determining the remediation cycle is to select multiple remediation cycles. Among the multiple remediation cycles, the numerical range of short-term remediation is 0.1 - 1 year, and the numerical range of medium- and long-term remediation is 0.5 - 5 years. Use the risk prediction model to predict the risk level area of pollutants after each cycle, and select a remediation cycle within the range of 0.1 - 5 years that makes the risk level area all close to or all low-risk areas;

[0071] Exemplarily, using in-situ chemical oxidation method and permeable reactive barrier method for short-term remediation includes:

[0072] In-situ chemical oxidation (ISCO) short-term remediation measures:

[0073] By injecting strong oxidants (such as Fenton's reagent, persulfate, permanganate, etc.), directly oxidize and degrade organic pollutants (such as benzene series, petroleum hydrocarbons, chlorinated hydrocarbons, etc.) in soil and groundwater, and convert them into non-toxic or low-toxic products (such as CO2, water or inorganic salts); select Fenton's reagent, activated persulfate or ozone, etc. according to the properties of pollutants; use a high-pressure injection drill to inject the agent into the contaminated layer through the drill pipe. Permeable reactive barrier (PRB) short-term remediation measures: Set an active material wall (such as zero-valent iron, modified zeolite, etc.) on the path of the contaminated water flow, and remove pollutants (such as Cr(VI), petroleum hydrocarbons, heavy metals) through adsorption, reduction or precipitation; the wall design adopts a continuous type or a funnel-gate type, and the wall thickness is calculated as 1.5 times the safety factor of the hydraulic retention time (HRT); Exemplarily, the method adopted for short-term remediation can also be as Figure 8 shown in the method.

[0074] For medium- and long-term remediation using phytoremediation or bioremediation methods, specifically:

[0075] Phytoremediation technology: Plant hyperaccumulator plants (such as Sedum alfredii, Pteris vittata, etc.), which absorb heavy metals in the soil through their roots and transfer them to the above-ground parts, and the pollutants are centrally treated after harvesting. Soil improvement: Apply chelating agents to activate heavy metals and improve the absorption efficiency of plants. Phytostabilization: Use plant root exudates (such as organic acids) or add passivators (phosphate fertilizers, biochar) to fix heavy metals and reduce their bioavailability. Microbial combination: Inoculate mycorrhizal fungi or plant growth-promoting rhizobacteria (PGPR) to promote plant growth and heavy metal absorption. Biological communities or genetically modified microorganisms for the directional degradation of recalcitrant pollutants (such as polycyclic aromatic hydrocarbons). Through the above methods, phytoremediation and bioremediation technologies can achieve the stabilization or removal of pollutants in medium- and long-term remediation, taking into account ecological restoration and sustainability.

Claims

1. A method for controlling the pollution risk of factory area soil and groundwater, characterized in that, It includes the following steps: Obtain the factory area dataset at the current moment and historical moments; the factory area dataset includes geological parameters, pollutant monitoring data, and environmental dynamic data; Based on the factory area dataset, establish a risk prediction model. The input of the risk prediction model is geological parameters, pollutant monitoring data, and environmental dynamic data, and the output of the risk prediction model is the risk level distribution within the factory area. The risk level distribution within the factory area includes high-risk areas, medium-risk areas, and low-risk areas; The risk prediction model includes a physical migration module for generating training data and physical constraint conditions from the input dataset, a learning and prediction module for performing spatio-temporal feature operations based on the training data and physical constraint conditions, and a dynamic decision module for risk classification output; Based on the risk prediction model, divide and repair the risk level distribution within the factory area; for high-risk areas, use in-situ chemical oxidation method and permeable reactive barrier method for short-term repair; for medium-risk areas, use phytoremediation method or bioremediation method for medium- and long-term repair; low-risk areas do not need to be repaired.

2. The method according to claim 1, wherein The geological parameters include permeability coefficient and porosity, and the pollutant monitoring data includes pollutant concentration; the environmental dynamic data includes meteorological data and hydrological dynamic data.

3. The method according to claim 1, wherein The method for establishing a risk prediction model according to the dataset includes: Construct a physical migration module: Select the HYDRUS software that can couple and simulate saturated-unsaturated water flow, solute transport, and multiple processes. According to the factory area dataset, set the migration equation in the HYDRUS software to obtain a physical migration module with the output of the concentration distribution of pollutants in soil and groundwater, source strength, and diffusion rate; Construct a learning and prediction module: Design a spatio-temporal attention graph neural network. The spatio-temporal attention graph neural network uses a multi-head self-attention mechanism to capture long-range dependencies, uses a dynamic graph convolution module to adaptively adjust the weights between nodes, and adds a mass conservation regularization term to the loss function; based on the concentration distribution of pollutants in soil and groundwater and environmental data, train the learning and prediction module to obtain the learning and prediction module, and the output of the learning and prediction module is the concentration distribution of pollutants at future times; Based on the concentration distribution of pollutants at future times obtained by the learning and prediction module, divide high-risk areas, medium-risk areas, and low-risk areas.

4. The method according to claim 1, wherein The method for dividing high-risk areas, medium-risk areas, and low-risk areas based on the concentration distribution of pollutants at future times obtained by the learning and prediction module includes: when the predicted concentration of the concentration distribution of pollutants at future times is less than or equal to 50% of the risk threshold, it is classified as a low-risk area; when the predicted concentration of the concentration distribution of pollutants at future times is greater than 50% and less than or equal to 75% of the risk threshold, it is classified as a medium-risk area; when the predicted concentration of the concentration distribution of pollutants at future times is greater than 75%, it is classified as a high-risk area.

5. The method according to claim 4, wherein The risk threshold is the environmental risk screening value.

6. The method according to claim 1, characterized in that The environmental data includes groundwater level data, surrounding vegetation distribution data, and land use change and irrigation data.

7. The method according to claim 1, wherein In the HYDRUS software, different rainfall boundary conditions are set according to rainfall data in different seasons; the relationship between air temperature and soil temperature is established through empirical formulas or experimental data, and then the corresponding soil temperature boundary is set in the software.

8. The method according to claim 1, wherein The method further includes: Determining a repair cycle. After the repair cycle is completed, obtaining the plant area dataset after the repair cycle is completed, and using the plant area dataset after the repair cycle is completed to update the risk prediction model; Based on the risk prediction model, predicting new high-risk areas, medium-risk areas, and low-risk areas, and completing the control until the risk level distribution within the plant area is all low risk.

9. The method according to claim 6, characterized in that, The method for determining the repair cycle is to select multiple repair cycles. Among the multiple repair cycles, the numerical range of recent repairs is 0.1 - 1 year, and the numerical range of medium- and long-term repairs is 0.5 - 5 years. The risk level areas of pollutants after each cycle are predicted using the risk prediction model, and a repair cycle that makes the risk level areas all close to or all be low-risk areas is selected within the range of 0.1 - 5 years.

Images