Soil remediation control method and system

Abstract

The invention provides a soil remediation control method and system, and relates to the technical field of soil remediation control, and the method comprises the steps: configuring a pollutant concentration distribution model architecture which comprises an input layer, a convolution full-connection layer and a plurality of separated long and short-term memory neural networks which are connected in sequence; obtaining soil environment information and pollutant distribution information of the target area, and chemical agent distribution information and a plurality of chemical reaction equilibrium formulas of the soil remediation scheme; constructing a plurality of balance loss items, and constructing a numerical loss item according to the output deviation; generating a plurality of sub-loss functions, adding to generate a total loss function, generating a pollutant concentration distribution model, and adjusting a soil remediation scheme. The problems that an existing soil remediation control model is low in prediction precision, insufficient in model generalization ability and incapable of being widely applied to soil remediation practice can be solved, and the effect of improving the prediction precision of pollutant concentration distribution and the model generalization ability in the soil remediation process is achieved.

Description

Technical Field

[0001] This application relates to the technical field of soil remediation control, and particularly to a soil remediation control method and system. Background Art

[0002] Soil pollution is one of the important environmental problems at present, posing a serious threat to the ecosystem, human health and sustainable development. With the rapid development of industrialization and urbanization, a large amount of heavy metals, organic pollutants and other harmful substances have accumulated in the soil. In order to effectively control soil pollution, soil remediation technologies have emerged. However, the complexity and uncertainty of the soil remediation process make it an urgent problem to accurately control the remediation process. Traditional soil remediation control models mainly rely on data-driven methods. By collecting a large amount of historical data, predictive models are constructed using supervised learning algorithms. Although these models can reflect the changing trend of soil pollutant concentration to a certain extent, due to the lack of physical and chemical constraints on the actual treatment process, their prediction accuracy is significantly insufficient. Exemplarily, traditional data-driven models usually cannot consider the equilibrium relationship of chemical reactions in the soil remediation process, resulting in a large deviation between the prediction results and the actual situation when the model faces a complex soil environment and a multi-pollutant system. In addition, these models can often only handle the remediation problems of single pollutants and are difficult to cope with the complex scenarios of multi-pollutant coupling.

[0003] In summary, the soil remediation control models in the prior art are mostly data-driven, lacking constraints on the chemical reaction equilibrium in the actual treatment process, and at the same time being difficult to handle the complex scenarios of multi-pollutant coupling. There are problems such as low prediction accuracy, insufficient model generalization ability, and inability to be widely applied to soil remediation practice. Summary of the Invention

[0004] The purpose of this application is to provide a soil remediation control method and system to solve the problems in the prior art that the soil remediation control models are mostly data-driven, lacking constraints on the chemical reaction equilibrium in the actual treatment process, and at the same time being difficult to handle the complex scenarios of multi-pollutant coupling, with low prediction accuracy, insufficient model generalization ability, and inability to be widely applied to soil remediation practice.

[0005] In view of the above problems, this application provides a soil remediation control method and system.

[0006] In a first aspect, the present application provides a soil remediation control method, which is implemented through a soil remediation control system. Among them, the soil remediation control method includes: configuring a pollutant concentration distribution model architecture, where the pollutant concentration distribution model architecture includes an input layer, a convolutional fully connected layer, and multiple separate long short-term memory neural networks connected in sequence; obtaining the soil environmental information, pollutant distribution information of the target area, the chemical agent distribution information of the soil remediation plan, and several chemical reaction equilibrium equations; placing the soil environmental information, pollutant distribution information, and chemical agent distribution information in the input layer, placing the pollutant concentration distribution prediction information in the output layer of the long short-term memory neural network, and any one long short-term memory neural network corresponds uniquely to an attribute pollutant concentration distribution prediction information; constructing several equilibrium loss terms according to the several chemical reaction equilibrium equations, and constructing a numerical loss term according to the output deviation; respectively combining the several equilibrium loss terms and the numerical loss term according to the multiple pollutant attribute identifications of the multiple long short-term memory neural networks to generate multiple sub-loss functions, adding them up to generate a total loss function, training the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, and adjusting the soil remediation plan to control soil remediation.

[0007] In a second aspect, the present application further provides a soil remediation control system for executing the soil remediation control method described in the first aspect. Among them, the soil remediation control system includes: an architecture configuration module for configuring a pollutant concentration distribution model architecture, where the pollutant concentration distribution model architecture includes an input layer, a convolutional fully connected layer, and multiple separate long short-term memory neural networks connected in sequence; an information acquisition module for obtaining the soil environmental information, pollutant distribution information of the target area, the chemical agent distribution information of the soil remediation plan, and several chemical reaction equilibrium equations; a model training module for placing the soil environmental information, pollutant distribution information, and chemical agent distribution information in the input layer, placing the pollutant concentration distribution prediction information in the output layer of the long short-term memory neural network, and any one long short-term memory neural network corresponds uniquely to an attribute pollutant concentration distribution prediction information; a loss analysis module for constructing several equilibrium loss terms according to the several chemical reaction equilibrium equations and constructing a numerical loss term according to the output deviation; a soil remediation module for respectively combining the several equilibrium loss terms and the numerical loss term according to the multiple pollutant attribute identifications of the multiple long short-term memory neural networks to generate multiple sub-loss functions, adding them up to generate a total loss function, training the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, and adjusting the soil remediation plan to control soil remediation.

[0008] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0009] By configuring a pollutant concentration distribution model architecture, wherein the pollutant concentration distribution model architecture includes an input layer, a convolutional fully-connected layer, and multiple separate long short-term memory neural networks connected in sequence; obtaining soil environmental information, pollutant distribution information, chemical agent distribution information of a soil remediation plan, and several chemical reaction equilibrium equations of the target area; placing the soil environmental information, pollutant distribution information, and chemical agent distribution information in the input layer, placing the pollutant concentration distribution prediction information in the output layer of the long short-term memory neural network, and any one long short-term memory neural network corresponds uniquely to one attribute pollutant concentration distribution prediction information; constructing several equilibrium loss terms according to the several chemical reaction equilibrium equations, and constructing a numerical loss term according to the output deviation; respectively combining the several equilibrium loss terms and the numerical loss term according to the multiple pollutant attribute identifications of the multiple long short-term memory neural networks to generate multiple sub-loss functions, summing them to generate a total loss function, training the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, and adjusting the soil remediation plan for soil remediation control. That is to say, by introducing chemical reaction equilibrium equations to construct equilibrium loss terms, the soil remediation control model can better reflect the physical and chemical change laws in the soil remediation process, thereby improving the prediction accuracy of the pollutant concentration distribution. In addition, by comprehensively considering chemical reaction equilibrium and numerical deviation, generating multiple sub-loss functions and summing them to generate a total loss function, the soil remediation control model has better generalization ability in different soil environments and remediation scenarios and can adapt to a wider range of soil remediation requirements. Finally, the technical goal of dynamic soil remediation control is achieved. The technical effect of improving the prediction accuracy and generalization ability of the soil remediation control model, and further improving the stability and reliability of soil remediation is achieved.

[0010] The above description is only an overview of the technical solution of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present application more obvious and understandable, the following specifically gives the specific embodiments of the present application. It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present application, nor is it used to limit the scope of the present application. Other features of the present application will become easily understandable through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0012] Figure 1Schematic flow chart of a soil remediation control method of the present application;

[0013] Figure 2 Schematic structural diagram of a soil remediation control system of the present application.

[0014] Explanation of reference numerals in the drawings:

[0015] 11. Architecture configuration module; 12. Information acquisition module; 13. Model training module; 14. Loss analysis module; 15. Soil remediation module. Detailed implementation manners

[0016] By providing a soil remediation control method and system, the present application solves the problems in the prior art that most of the soil remediation control models are data-driven, lack of constraints on the chemical reaction equilibrium in the actual treatment process, and are difficult to handle complex scenarios of multi-pollutant coupling, resulting in low prediction accuracy and insufficient model generalization ability, and being unable to be widely applied to soil remediation practice. By introducing a chemical reaction equilibrium formula to construct an equilibrium loss term, the soil remediation control model can better reflect the physical and chemical change laws in the soil remediation process, thereby improving the prediction accuracy of pollutant concentration distribution. In addition, by comprehensively considering the chemical reaction equilibrium and numerical deviation, generating multiple sub-loss functions and summing them to generate a total loss function, the soil remediation control model has better generalization ability under different soil environments and remediation scenarios and can adapt to a wider range of soil remediation requirements. Finally, the technical goal of dynamic soil remediation control is achieved. The technical effects of improving the prediction accuracy and generalization ability of the soil remediation control model, and further improving the stability and reliability of soil remediation are achieved.

[0017] Next, the technical solutions in the present application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application. In addition, it should be noted that for the sake of description, only the parts related to the present application are shown in the drawings rather than all of them.

[0018] Example 1, please refer to the attached Figure 1 , the present application provides a soil remediation control method, wherein the method is applied to a soil remediation control system, and the soil remediation control method specifically includes the following steps:

[0019] Step P10: Configure the pollutant concentration distribution model architecture, wherein the pollutant concentration distribution model architecture includes an input layer, a convolutional fully connected layer, and multiple separate long short-term memory neural networks connected in sequence;

[0020] Specifically, a deep learning model architecture for predicting pollutant concentration distribution is constructed, which mainly consists of an input layer, a convolutional fully-connected layer, and multiple separate long short-term memory neural networks (LSTMs). This architecture can effectively process complex soil environment data and pollutant dynamic changes, and through hierarchical processing and the design of memory networks, achieve accurate prediction of pollutant concentration distribution.

[0021] Specifically, first, the input layer receives various data such as soil environment information, pollutant distribution information, and chemical agent distribution information of the target area. These data provide rich background information and initial conditions for the model. Then, the convolutional fully-connected layer performs feature extraction and preliminary processing on the input data. Through convolutional operations, it can effectively capture the spatial correlation in the data, laying a foundation for subsequent in-depth analysis. Next, multiple separate long short-term memory neural networks respectively model and predict the pollutant concentration distribution of different attributes. Each LSTM network focuses on a specific pollutant and can effectively handle the dynamic characteristics of pollutant concentration changes over time. In addition, the separate design of the LSTM networks enables the model to simultaneously handle the problem of multi-pollutant coupling, further enhancing the applicability and flexibility of the model. Finally, by constructing a balance loss term and a numerical loss term, a total loss function is generated to train the model, ensuring that the model output not only conforms to the chemical reaction equilibrium law but also accurately reflects the actual data, thereby achieving high-precision prediction of pollutant concentration distribution. This architecture design makes full use of the feature extraction ability of the deep learning model and the time series processing ability of the memory network, providing strong technical support for soil remediation control.

[0022] In summary, through the construction of a hierarchical and functionally clear deep learning model architecture, efficient prediction of soil pollutant concentration distribution is achieved. Its technical goal is to improve prediction accuracy, enhance the adaptability of the model to multi-pollutant systems, and provide a reliable prediction tool for the optimization of soil remediation plans. Through this architecture design, the model can effectively process complex soil environment data, capture the dynamic change rules of pollutant concentration, thereby providing accurate guidance for the soil remediation process and significantly improving the efficiency and effect of soil remediation.

[0023] Step P20: Obtain the soil environment information, pollutant distribution information, chemical agent distribution information of the soil remediation plan, and several chemical reaction equilibrium equations of the target area;

[0024] Specifically, by obtaining the soil environmental information, pollutant distribution information, chemical agent distribution information, and chemical reaction equilibrium equations of the target area, comprehensive and accurate data support is provided for the soil remediation process. These data are the basis for constructing the pollutant concentration distribution model, which can help the model better understand the physical and chemical change laws in the soil remediation process, so as to achieve high-precision prediction of the pollutant concentration distribution and optimization of the remediation plan. Specifically, first, obtain the soil environmental information of the target area, which includes the physical properties of the soil (such as texture, porosity, etc.) and chemical properties (such as pH value, redox potential, etc.). These information are the basis for understanding the impact of soil on the migration and transformation of pollutants. Then collect the pollutant distribution information to clarify the types, concentrations, and spatial distribution of pollutants in the target area, which is crucial for the model to predict the change of pollutant concentration. Next, obtain the chemical agent distribution information in the soil remediation plan, including the types, dosages, and application methods of the agents. These information can help the model understand the direct impact of chemical agents on the change of pollutant concentration during the remediation process. In addition, several chemical reaction equilibrium equations need to be collected. These equilibrium equations reflect the chemical reaction laws between pollutants and chemical agents in the soil remediation process and are important bases for the model to predict the change of pollutant concentration. Through the integration of these data, the model can more comprehensively simulate the physical and chemical changes in the soil remediation process. It should be noted that the acquisition of these data requires a combination of field sampling, laboratory analysis, and theoretical derivation and other methods to ensure the accuracy and integrity of the data. Obviously, the integration of these information provides rich inputs for the model, enabling it to more accurately predict the pollutant concentration distribution.

[0025] In summary, by obtaining the soil environmental information, pollutant distribution information, chemical agent distribution information, and chemical reaction equilibrium equations, comprehensive and accurate data support is provided for the prediction of the pollutant concentration distribution in the soil remediation process. The integration of these data can help the model more accurately reflect the physical and chemical change laws in the soil remediation process, so as to achieve high-precision prediction of the pollutant concentration distribution, provide a scientific basis for the optimization and control of the soil remediation plan, and significantly improve the efficiency and effect of soil remediation.

[0026] Step P30: Place the soil environmental information, pollutant distribution information, and chemical agent distribution information in the input layer, and place the pollutant concentration distribution prediction information in the output layer of the long short-term memory neural network. Any one long short-term memory neural network corresponds uniquely to an attribute pollutant concentration distribution prediction information;

[0027] Specifically, by reasonably configuring data input and output, the effective integration of soil environmental information, pollutant distribution information, and chemical agent distribution information is achieved, and the long short-term memory neural network (LSTM) is used to accurately predict the pollutant concentration distribution of different attributes. This design not only ensures that the model can fully capture the dynamic changes during the soil remediation process but also improves the adaptability and prediction accuracy of the model for multi-pollutant systems through a one-to-one network configuration, thereby providing reliable technical support for soil remediation control.

[0028] Specifically, first, the soil environmental information, pollutant distribution information, and chemical agent distribution information are placed as input data in the input layer of the model. Among them, the soil environmental information includes the physical and chemical properties of the soil, the pollutant distribution information clarifies the types and concentrations of pollutants, and the chemical agent distribution information reflects the agents used in the remediation process and their action modes. Then, these input data are passed to the subsequent network layers for analysis and processing. Subsequently, the pollutant concentration distribution prediction information is placed in the output layer of the long short-term memory neural network. The long short-term memory neural network is a neural network structure specifically used to process time series data and can effectively capture the changing patterns of pollutant concentrations over time. Among them, each long short-term memory neural network corresponds uniquely to the pollutant concentration distribution prediction information of a specific attribute, which means that each network focuses on predicting the concentration changes of one pollutant. This one-to-one network configuration enables the model to perform specialized modeling and optimization for the characteristics of different pollutants, thereby improving the prediction accuracy. In addition, through separate long short-term memory neural networks, the model can handle multiple pollutants simultaneously, further enhancing its adaptability to complex soil pollution environments. Finally, through this input-output configuration, the input soil environment, pollutant distribution, and chemical agent distribution information can be converted into accurate pollutant concentration distribution prediction results. By leveraging the powerful feature extraction ability of the deep learning model and the dynamic modeling ability of the long short-term memory neural network, high-precision prediction of pollutant concentration changes during the soil remediation process is achieved.

[0029] In summary, by using the soil environmental information, pollutant distribution information, and chemical agent distribution information as input, and using the output layer of the long short-term memory neural network to predict the pollutant concentration distribution, and through a one-to-one network configuration, accurate modeling and prediction of multi-pollutant systems are achieved. This design not only improves the prediction accuracy of the model but also enhances its adaptability to complex soil pollution environments.

[0030] Step P40: Construct a number of equilibrium loss terms according to the several chemical reaction equilibrium equations, and construct a numerical loss term according to the output deviation.

[0031] Specifically, by introducing chemical reaction equilibrium equations to construct equilibrium loss terms and combining output deviations to construct numerical loss terms, a dual constraint mechanism is provided for model training. This design not only takes into account the physical and chemical laws of pollutant concentration changes during the soil remediation process but also considers the deviation between the model prediction value and the actual value, enabling the model to improve prediction accuracy and reliability on the premise of satisfying chemical reaction equilibrium and ultimately achieving precise control of the soil remediation process. Specifically, first, several equilibrium loss terms are constructed according to the several chemical reaction equilibrium equations. The chemical reaction equilibrium equations are the basic laws of the reaction between pollutants and chemical agents during the soil remediation process, reflecting the transformation and migration relationships of substances during the remediation process. By converting these equilibrium equations into equilibrium loss terms, the training process can be constrained by physical and chemical laws to ensure that the prediction results conform to the actual chemical reaction process. For example, when a chemical agent reacts with a pollutant, its concentration change must follow the stoichiometric relationship, and the equilibrium loss term can effectively monitor whether the model output meets this requirement. In addition, the introduction of the equilibrium loss term enables the model to better handle complex chemical reactions in a multi-pollutant system, further enhancing the applicability and accuracy of the model. Then, a numerical loss term is constructed based on the output deviation. The numerical loss term is an important indicator for measuring the difference between the model prediction value and the actual observed value, usually calculated using the mean square error or other statistical indicators. By constructing the numerical loss term, the accuracy of the prediction results can be quantitatively evaluated, and the model parameters can be continuously optimized during the training process to minimize the prediction error. Obviously, the introduction of the numerical loss term provides a direct feedback mechanism for model training, enabling the model to adjust its prediction ability according to actual data. Finally, the equilibrium loss term and the numerical loss term are combined to form the total loss function for model training. This dual constraint mechanism takes into account both the equilibrium law of chemical reactions and the deviation between the model prediction value and the actual value, enabling the model to improve prediction accuracy while satisfying physical and chemical constraints. In this way, the model can more accurately reflect the dynamic changes during the soil remediation process, providing a reliable basis for the optimization and control of the soil remediation plan. In addition, the weight allocation of the equilibrium loss term and the numerical loss term needs to be adjusted according to the specific application scenario to ensure that the model can achieve the best performance under different conditions.

[0032] In summary, by constructing equilibrium loss terms and numerical loss terms, a dual constraint of physical and chemical laws and prediction accuracy is provided for model training, improving the adaptability of the model to complex soil remediation processes and significantly enhancing the prediction accuracy and reliability of the model.

[0033] Step P50: According to the pollutant property identifiers of multiple long short-term memory neural networks respectively, combine the several balance loss terms and the numerical loss terms to generate multiple sub-loss functions, sum them up to generate a total loss function, train the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, and adjust the soil remediation plan for soil remediation control.

[0034] Specifically, by combining the pollutant property identifiers of the long short-term memory neural networks, the balance loss terms and the numerical loss terms are combined to generate multiple sub-loss functions, and further summed up to generate a total loss function for training the pollutant concentration distribution model. By optimizing the total loss function, the precise regulation of the soil remediation plan is achieved, thereby achieving the purpose of efficiently remediating soil pollution.

[0035] First, for each of the multiple long short-term memory neural networks, according to its corresponding pollutant property identifier, combine the balance loss term and the numerical loss term respectively. Each long short-term memory neural network focuses on predicting a specific pollutant concentration distribution. Therefore, the corresponding balance loss term and numerical loss term can be optimized according to the characteristics of the pollutant. For example, for heavy metal pollutants, their chemical reaction equilibrium equations may be different from those of organic pollutants. Therefore, the corresponding balance loss term needs to be constructed according to the specific chemical reaction equilibrium equation. At the same time, the numerical loss term is used to measure the deviation between the predicted value and the actual observed value of the network. In this way, each sub-loss function can accurately reflect the prediction accuracy and chemical reaction equilibrium constraints of a specific pollutant. Then, sum up these sub-loss functions for different pollutants to generate a total loss function. The construction of the total loss function is to comprehensively consider the prediction accuracy and chemical reaction equilibrium constraints of all pollutants, so that the model can balance the differences between different pollutants during the training process. In addition, the optimization process of the total loss function can ensure that the model reaches the best performance as a whole. Among them, when constructing the total loss function, different weights can be assigned to each sub-loss function according to the importance or concentration level of different pollutants to optimize the training effect of the model. Finally, train the pollutant concentration distribution model architecture by optimizing the total loss function to generate the final pollutant concentration distribution model. This model can output accurate pollutant concentration distribution prediction results based on the input soil environment information, pollutant distribution information, and chemical agent distribution information. Based on these prediction results, the soil remediation plan can be dynamically adjusted, such as optimizing the dosage of chemical agents or adjusting the remediation process parameters, so as to achieve precise control of the soil remediation process.

[0036] In summary, by constructing sub-loss functions for different pollutants and combining them into a total loss function, efficient training of the pollutant concentration distribution model is achieved. Not only the physicochemical properties of the pollutants are considered, but also the accuracy of model prediction is taken into account. Finally, by optimizing the total loss function, a precise regulation basis is provided for the soil remediation plan, significantly improving the efficiency and effect of soil remediation.

[0037] Furthermore, according to the several chemical reaction equilibrium equations, several equilibrium loss terms are constructed, and according to the output deviation, a numerical loss term is constructed, including:

[0038] Step P41: Obtain the first chemical reaction equilibrium equation and construct the first equilibrium loss term:

[0039] a1r1+...+a i r i +...+a Q r Q →b1p1+...+b j p j +...+b Y p Y ,

[0040]

[0041] where a i and b j represent stoichiometric coefficients, r i represents the reactant, p j represents the product, represents the reactant concentration, k represents the training sample serial number, represents the predicted product concentration, N represents the number of training samples, a1r1+...+a i r i +...+a Q r Q →b1p1+...+b j p j +...+b Y p Y represents the chemical reaction equilibrium equation;

[0042] Step P42: Until the L-th chemical reaction equilibrium equation is used to construct the L-th equilibrium loss term e L , add the first equilibrium loss term e1 to the L-th equilibrium loss term e L to the several equilibrium loss terms;

[0043] Step P43: Construct the numerical loss term according to the deviation between the supervised data and the output data.

[0044] Specifically, by gradually constructing the equilibrium loss term and the numerical loss term, a comprehensive constraint mechanism is provided for the training of the soil remediation model. First, obtain the first chemical reaction equilibrium equation a1r1 +... + a i r i +... + a Q r Q →b1p1 +... + b j p j +... + b Y p Y , and construct the first equilibrium loss term accordingly. The chemical reaction equilibrium equation is the basic law of the reaction between pollutants and chemical agents in the soil remediation process. The introduction of the equilibrium loss term can ensure that the model prediction results conform to these physicochemical constraints. For example, when a chemical agent reacts with a pollutant, its concentration change must follow the stoichiometric relationship, and the equilibrium loss term can effectively monitor whether the model output meets this requirement. In this way, the model can gradually learn the chemical reaction laws in the soil remediation process during training, providing a basis for subsequent predictions. Then continue to construct the subsequent equilibrium loss terms until the Lth equilibrium loss term e L is constructed according to the Lth chemical reaction equilibrium equation, ensuring that all relevant chemical reaction equilibrium equations are incorporated into the model's constraint conditions. Add the first equilibrium loss term to the Lth equilibrium loss term e L all into several equilibrium loss terms, enabling the model to comprehensively consider the equilibrium relationships of various chemical reactions in the soil remediation process, further improving the accuracy and reliability of the model. Finally, construct the numerical loss term according to the deviation between the supervised data and the output data. By introducing the numerical loss term, the model can continuously adjust its own parameters during training to minimize the prediction error, thereby improving the prediction accuracy of the model.

[0045] Furthermore, according to the multiple pollutant attribute identifiers of multiple long short-term memory neural networks respectively, combine the several equilibrium loss terms and the numerical loss term to generate multiple sub-loss functions, including:

[0046] Step P51: Extract the first pollutant attribute identifier of the first long short-term memory neural network among the multiple long short-term memory neural networks;

[0047] Step P52: Extract the first selected equilibrium loss term set whose reactants have the first pollutant attribute identifier from the several equilibrium loss terms;

[0048] Step P53: Perform weighted summation on the first selected equilibrium loss term set to generate the chemical equilibrium loss function;

[0049] Step P54: Add the chemical equilibrium loss function and the numerical loss term to generate a first sub-loss function, and add it to the multiple sub-loss functions.

[0050] Specifically, the corresponding equilibrium loss terms are extracted through long short-term memory neural networks for different pollutant attributes, and combined with the numerical loss terms to generate sub-loss functions, thereby providing a refined constraint mechanism for model training.

[0051] First, extract the first pollutant attribute identifier corresponding to the first network among multiple long short-term memory neural networks. The long short-term memory neural network is a model for predicting the concentration distribution of specific pollutants, and each network focuses on the characteristics of one pollutant. The purpose of extracting the pollutant attribute identifier is to be able to subsequently select the chemical reaction equilibrium constraints related to this pollutant specifically. Exemplarily, if the first LSTM network is used to predict the concentration distribution of heavy metal mercury, then the extracted pollutant attribute identifier is the related characteristics of mercury. This process provides a clear target for the subsequent construction of the loss function. Then, extract the first selected set of equilibrium loss terms whose reactants have the first pollutant attribute identifier from several equilibrium loss terms. This means screening out the content related to the pollutant corresponding to the first LSTM network from all possible chemical reaction equilibrium constraints. Exemplarily, if the first LSTM network focuses on the concentration distribution of mercury, then the first selected set of equilibrium loss terms will include the loss terms corresponding to all chemical reaction equilibrium equations related to mercury. Then, perform weighted summation on the first selected set of equilibrium loss terms to generate a chemical equilibrium loss function. The process of weighted summation considers the importance of different chemical reaction equilibrium equations. By assigning weights to each equilibrium loss term, the model can better balance the constraint strength between different reactions. Exemplarily, some chemical reactions may have a greater impact on the change in pollutant concentration, so higher weights can be assigned. The generation of the chemical equilibrium loss function enables the model to comprehensively consider all chemical reaction equilibrium constraints related to a specific pollutant, further optimizing the prediction results. Finally, add the chemical equilibrium loss function and the numerical loss term to generate a first sub-loss function, and add it to the multiple sub-loss functions. The numerical loss term is used to measure the deviation between the model prediction value and the actual observed value, while the chemical equilibrium loss function ensures that the prediction result conforms to the chemical reaction equilibrium law. By combining the two, the first sub-loss function takes into account both the prediction accuracy and the physicochemical constraints.

[0052] In summary, by extracting specific pollutant attribute identifiers, screening relevant equilibrium loss terms, and combining numerical loss terms to generate sub-loss functions, a refined constraint on model training is achieved, further improving the prediction performance of the model and the reliability of soil remediation control.

[0053] Further, perform weighted summation on the first selected set of equilibrium loss terms to generate a chemical equilibrium loss function, including:

[0054] Step P531: When the first pollutant attribute identifier consumes a unit weight, count the set of weights of co-reactants consumed by the first selected set of equilibrium loss terms, where the co-reactants are the set of substances that react with the first pollutant in any one selected equilibrium loss term;

[0055] Step P532: Sum up the set of weights of co-reactants consumed to obtain the total weight of co-reactants consumed;

[0056] Step P533: Calculate the ratio of the set of weights of co-reactants consumed to the total weight of co-reactants consumed respectively, and then subtract the ratio from 1 to obtain the weight distribution of the selected equilibrium loss terms;

[0057] Step P534: Perform weighted summation on the first selected set of equilibrium loss terms according to the weight distribution of the selected equilibrium loss terms to generate a chemical equilibrium loss function.

[0058] Specifically, through a refined weight assignment mechanism, weighted processing is performed on the equilibrium loss terms related to the first pollutant to generate a chemical equilibrium loss function. First, when the first pollutant attribute identifier consumes a unit weight, count the set of weights of co-reactants in the first selected set of equilibrium loss terms. Co-reactants refer to other substances that react with the first pollutant in the selected chemical reaction equilibrium formula. Thus, when a unit weight of pollutant is consumed, the consumption of other substances participating in the reaction can be understood at the same time. For example, if the first pollutant is a certain heavy metal, the co-reactants may be chemical agents used for remediation or other soil components participating in the reaction. Then sum up the set of weights of co-reactants consumed to obtain the total weight of co-reactants consumed. Then calculate the ratio of the set of weights of co-reactants consumed to the total weight of co-reactants consumed respectively, and subtract the ratio from 1 to obtain the weight distribution of the selected equilibrium loss terms. Demonstratively, if the proportion of the weight of a certain co-reactant consumed in the total weight is large, the corresponding equilibrium loss term will be given a smaller weight (because the result of subtracting the ratio from 1 is smaller), and vice versa. Finally, perform weighted summation on the first selected set of equilibrium loss terms according to the weight distribution of the selected equilibrium loss terms to generate a chemical equilibrium loss function. The chemical equilibrium loss function can comprehensively consider the importance of different co-reactants in chemical reactions, enabling the model to more reasonably reflect the chemical reaction equilibrium law in the soil remediation process during training.

[0059] In summary, by statistically analyzing the weight loss of co-reactants, calculating the weight distribution, and weighted summing the equilibrium loss terms, a chemical equilibrium loss function is generated. Among them, the generated chemical equilibrium loss function can accurately reflect the dynamic change law of pollutants during soil remediation, providing an important basis for model optimization.

[0060] Furthermore, according to the deviation between the supervised data and the output data, the numerical loss term is constructed as follows:

[0061]

[0062] where e num represents the numerical loss term, C z,s,1 represents the predicted concentration of pollutants at the s-th position in the z-th training, C z,s,0 represents the actual concentration of pollutants at the s-th position in the z-th training, U represents the number of samples, and I represents the number of positions.

[0063] Specifically, according to the deviation between the supervised data and the output data, the numerical loss term is constructed The numerical loss term is used to measure the difference between the model prediction value and the actual observation value, and is the key to optimizing the model prediction accuracy. Among them, e num represents the numerical loss term, C z,s,1 represents the predicted concentration of pollutants at the s-th position in the z-th training, C z,s,0 represents the actual concentration of pollutants at the s-th position in the z-th training, U represents the number of samples, and I represents the number of positions.

[0064] Furthermore, the pollutant concentration distribution model architecture is trained to generate a pollutant concentration distribution model, including:

[0065] Step P55: According to the several chemical reaction equilibrium equations, divide the multiple pollutant attributes of the multiple separated long short-term memory neural networks into reaction stages to obtain a pollutant reaction directed network, where the first pollutant at the end point of the arrow of the pollutant reaction directed network is the origin pollutant of the second pollutant at the starting point of the arrow, and the origin pollutant means that the co-reactant of the second pollutant is the first pollutant, or a substance generated by the reaction of the first pollutant and the co-reactant of the first pollutant;

[0066] Step P56: Divide the multiple separated long short-term memory neural networks according to the pollutant reaction directed network to obtain a set of long short-term memory neural networks in the first stage until a set of long short-term memory neural networks in the B-th stage;

[0067] Step P57: Respectively according to the first-stage pollutant attribute identification set of the first-stage long short-term memory neural network ensemble, combine the several balance loss terms and the numerical loss terms to generate multiple first-stage sub-loss functions, sum them up to generate the first-stage total loss function, and train the first-stage pollutant concentration distribution model;

[0068] Step P58: Freeze the first-stage long short-term memory neural network ensemble, respectively according to the second-stage pollutant attribute identification set of the second-stage long short-term memory neural network ensemble, combine the several balance loss terms and the numerical loss terms to generate multiple second-stage sub-loss functions, sum them up to generate the second-stage total loss function, and train the second-stage pollutant concentration distribution model;

[0069] Step P59: Until training the pollutant concentration distribution model of the Bth stage, which is set as the pollutant concentration distribution model.

[0070] Specifically, by constructing a directed network of pollutant reactions and based on this, dividing multiple separate long short-term memory neural networks (LSTMs) into stages, and then training the pollutant concentration distribution model stage by stage. First, according to several chemical reaction equilibrium equations, the reaction stages of multiple pollutant attributes of multiple separate long short-term memory neural networks are divided to obtain a directed network of pollutant reactions. This network represents the chemical reaction relationship between pollutants through the starting and ending points of the arrows, where the first pollutant at the end point of the arrow is the originating pollutant of the second pollutant at the starting point of the arrow. The originating pollutant means that the co-reactant of the second pollutant is the first pollutant, or the substance formed by the reaction of the first pollutant with its co-reactant. In this way, the model can clearly describe the generation, transformation, and consumption relationships between pollutants, providing a basis for subsequent stage division. Then, according to the directed network of pollutant reactions, multiple separate long short-term memory neural networks are divided into stages to obtain a set of long short-term memory neural networks from the first stage to the Bth stage. This process decomposes the complex pollutant reaction process into multiple stages, and each stage corresponds to a specific set of LSTM networks and pollutant attribute identification sets. This stage-by-stage processing method can effectively reduce the complexity of model training while improving the adaptability and accuracy of the model to different reaction stages. Then, for the set of long short-term memory neural networks in the first stage, according to its first-stage pollutant attribute identification set, several balance loss terms and numerical loss terms are combined to generate multiple first-stage sub-loss functions, and they are summed to generate the first-stage total loss function. By optimizing the first-stage total loss function, the first-stage pollutant concentration distribution model is trained. This process ensures that the model can accurately reflect the pollutant concentration distribution and chemical reaction laws in the initial stage, laying a foundation for the training of subsequent stages. Next, freeze the set of long short-term memory neural networks in the first stage. For the set of long short-term memory neural networks in the second stage, according to its second-stage pollutant attribute identification set, several balance loss terms and numerical loss terms are combined to generate multiple second-stage sub-loss functions, and they are summed to generate the second-stage total loss function. By optimizing the second-stage total loss function, the second-stage pollutant concentration distribution model is trained. Freezing the network parameters of the previous stage can ensure the stability of the model in the training of subsequent stages while gradually optimizing the model performance of each stage. Finally, gradually train until the pollutant concentration distribution model of the Bth stage and set it as the final pollutant concentration distribution model. Through the method of stage-by-stage optimization, the model can gradually adapt to the dynamic changes of pollutants in different stages of the soil remediation process, and finally generate a model architecture that can accurately reflect the entire remediation process.

[0071] In summary, by constructing a directed network of pollutant reactions, optimizing the long short-term memory neural network stage by stage, and gradually training the pollutant concentration distribution model, the accurate modeling of the dynamic changes of pollutants in the soil remediation process is achieved.

[0072] Further, a pollutant concentration distribution model is generated, and the soil remediation plan is adjusted to control soil remediation, including:

[0073] Step P60: Statistically calculate the proportion of the number of abnormal positions where the concentration deviation at the same position between the pollutant concentration distribution prediction information output by the pollutant concentration distribution model and the pollutant concentration distribution expected information exceeds the concentration deviation threshold;

[0074] Step P61: When the proportion of the number of abnormal positions is greater than or equal to the preset proportion, adjust the soil remediation plan a preset number of times to obtain an updated soil remediation plan, where the adjustment strategy is to continue adjusting based on the updated soil remediation plan if it is advantageous, and to return to the soil remediation plan to continue adjusting if it is disadvantageous;

[0075] Step P62: When the proportion of the number of abnormal positions in the updated soil remediation plan is less than the preset proportion, output the updated soil remediation plan;

[0076] Step P63: Otherwise, continue to iterate.

[0077] Specifically, by evaluating the deviation between the prediction result and the expected result of the pollutant concentration distribution model, the soil remediation plan is dynamically adjusted to optimize the soil remediation effect. First, statistically calculate the concentration deviation at the same position between the pollutant concentration distribution prediction information output by the pollutant concentration distribution model and the pollutant concentration distribution expected information. Calculate the proportion of the number of abnormal positions where these deviations exceed the concentration deviation threshold. For example, if the predicted concentration at a certain position deviates too much from the expected concentration, it indicates that the remediation effect at this position has not reached the expectation and needs to be further optimized. Then, when the proportion of the number of abnormal positions is greater than or equal to the preset proportion, adjust the soil remediation plan a preset number of times to obtain an updated soil remediation plan. The adjustment strategy is to continue to optimize based on the updated soil remediation plan: if the adjustment result shows an advantage (i.e., the proportion of the number of abnormal positions decreases), then continue to use the updated plan; if the adjustment result shows a disadvantage (i.e., the proportion of the number of abnormal positions has not improved or has deteriorated), then return to the original soil remediation plan to continue adjusting. The core of this strategy is to dynamically adjust the soil remediation plan to ensure that each adjustment moves towards the optimization goal. Then, when the proportion of the number of abnormal positions in the updated soil remediation plan is less than the preset proportion, it indicates that the plan has achieved the expected remediation effect, and at this time, output the updated soil remediation plan. This process marks the end of the current iteration stage and provides an optimized plan for actual soil remediation operations. Finally, if the proportion of the number of abnormal positions in the updated soil remediation plan still does not reach the expected value, continue to perform iterative optimization. Through continuous iteration, gradually approach the ideal soil remediation effect. In summary, by quantifying the deviation between the model prediction and the expectation, dynamically adjust the soil remediation plan, and gradually approach the ideal remediation effect through iterative optimization.

[0078] In summary, a soil remediation control method provided by the present application has the following technical effects:

[0079] By configuring the pollutant concentration distribution model architecture, where the pollutant concentration distribution model architecture includes an input layer, a convolutional fully connected layer, and multiple separate long short-term memory neural networks connected in sequence; obtaining the soil environment information, pollutant distribution information, chemical agent distribution information of the soil remediation plan, and several chemical reaction equilibrium equations of the target area; placing the soil environment information, pollutant distribution information, and chemical agent distribution information in the input layer, placing the pollutant concentration distribution prediction information in the output layer of the long short-term memory neural network, and any one long short-term memory neural network corresponds uniquely to a pollutant concentration distribution prediction information of an attribute; constructing several equilibrium loss terms according to the several chemical reaction equilibrium equations, and constructing a numerical loss term according to the output deviation; respectively combining the several equilibrium loss terms and the numerical loss term according to the multiple pollutant attribute identifications of the multiple long short-term memory neural networks to generate multiple sub-loss functions, summing them to generate a total loss function, training the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, and adjusting the soil remediation plan to control soil remediation. That is to say, by introducing chemical reaction equilibrium equations to construct equilibrium loss terms, the soil remediation control model can better reflect the physical and chemical change laws in the soil remediation process, thereby improving the prediction accuracy of pollutant concentration distribution. In addition, by comprehensively considering chemical reaction equilibrium and numerical deviation, generating multiple sub-loss functions and summing them to generate a total loss function, the soil remediation control model has better generalization ability under different soil environments and remediation scenarios and can adapt to a wider range of soil remediation requirements. Finally, the technical goal of dynamic soil remediation control is achieved. The technical effects of improving the prediction accuracy and generalization ability of the soil remediation control model, and further improving the stability and reliability of soil remediation are achieved.

[0080] Embodiment 2, based on the same inventive concept as a soil remediation control method in the foregoing embodiment, the present application also provides a soil remediation control system. Please refer to the attached Figure 2 , the soil remediation control system includes:

[0081] An architecture configuration module 11, configured to configure a pollutant concentration distribution model architecture, where the pollutant concentration distribution model architecture includes an input layer, a convolutional fully connected layer, and multiple separate long short-term memory neural networks connected in sequence;

[0082] An information acquisition module 12, configured to obtain the soil environment information, pollutant distribution information, chemical agent distribution information of the soil remediation plan, and several chemical reaction equilibrium equations of the target area;

[0083] The model training module 13 is configured to place the soil environmental information, pollutant distribution information, and chemical agent distribution information in the input layer, and place the pollutant concentration distribution prediction information in the output layer of the long short-term memory neural network. Any long short-term memory neural network corresponds uniquely to a piece of pollutant concentration distribution prediction information for an attribute.

[0084] The loss analysis module 14 is configured to construct a number of equilibrium loss terms according to the several chemical reaction equilibrium equations, and construct a numerical loss term according to the output deviation.

[0085] The soil remediation module 15 is configured to combine the several equilibrium loss terms and the numerical loss term respectively according to the multiple pollutant attribute identifiers of multiple long short-term memory neural networks to generate multiple sub-loss functions, sum them up to generate a total loss function, train the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, and adjust the soil remediation plan for soil remediation control.

[0086] Further, the loss analysis module 14 in the soil remediation control system is further configured to:

[0087] Obtain a first chemical reaction equilibrium equation and construct a first equilibrium loss term:

[0088] a1r1+...+a i r i +...+a Q r Q →b1p1+...+b j p j +...+b Y p Y ,

[0089]

[0090] where a i and b j represent stoichiometric coefficients, r i represents the reactant, p j represents the product, represents the reactant concentration, k represents the training sample serial number, represents the predicted product concentration, N represents the number of training samples, a1r1+...+a i r i +...+a Q r Q →b1p1+...+b j p j +...+b Y p Y represents the chemical reaction equilibrium equation;

[0091] Until the $L$-th chemical reaction equilibrium equation is used to construct the $L$-th equilibrium loss term $e$ L Add the first equilibrium loss term $e_1$ to the $L$-th equilibrium loss term $e$ L to the several equilibrium loss terms;

[0092] Construct the numerical loss term according to the deviation between the supervised data and the output data.

[0093] Furthermore, the soil remediation module 15 in the soil remediation control system is further configured to:

[0094] Extract the first pollutant attribute identifier of the first long short-term memory neural network among the multiple long short-term memory neural networks;

[0095] Extract the first selected equilibrium loss term set in which the reactants have the first pollutant attribute identifier from the several equilibrium loss terms;

[0096] Perform weighted summation on the first selected equilibrium loss term set to generate a chemical equilibrium loss function;

[0097] Sum the chemical equilibrium loss function and the numerical loss term to generate a first sub-loss function, and add it to the multiple sub-loss functions.

[0098] Furthermore, the soil remediation module 15 in the soil remediation control system is further configured to:

[0099] Statistically calculate the set of consumption weights of co-reactants in the first selected equilibrium loss term set when the first pollutant attribute identifier consumes a unit weight, where the co-reactant is the set of substances that react with the first pollutant in any one selected equilibrium loss term;

[0100] Sum the set of consumption weights of co-reactants to obtain the total consumption weight of co-reactants;

[0101] Calculate the ratio of the set of consumption weights of co-reactants to the total consumption weight of co-reactants respectively, and then subtract the ratio from 1 to obtain the weight distribution of the selected equilibrium loss terms;

[0102] Perform weighted summation on the first selected equilibrium loss term set according to the weight distribution of the selected equilibrium loss terms to generate a chemical equilibrium loss function.

[0103] Furthermore, the loss analysis module 14 in the soil remediation control system is further configured to:

[0104] Construct the numerical loss term:

[0105]

[0106] where $e$num Characterize the numerical loss term, C z,s,1 Characterize the predicted pollutant concentration at the s-th position in the z-th training, C z,s,0 Characterize the actual pollutant concentration at the s-th position in the z-th training. U represents the number of samples, and I represents the number of positions.

[0107] Furthermore, the soil remediation module 15 in the soil remediation control system is further configured to:

[0108] According to the several chemical reaction equilibrium equations, divide the pollutant attributes of the multiple separated long short-term memory neural networks into reaction stages to obtain a pollutant reaction directed network, where the first pollutant at the end point of the arrow in the pollutant reaction directed network is the origin pollutant of the second pollutant at the starting point of the arrow, and the origin pollutant means that the co-reactant of the second pollutant is the first pollutant, or the substance formed by the reaction of the first pollutant and the co-reactant of the first pollutant;

[0109] According to the pollutant reaction directed network, divide the multiple separated long short-term memory neural networks into stages to obtain the long short-term memory neural network set of the first stage until the long short-term memory neural network set of the B-th stage;

[0110] Respectively, according to the first-stage pollutant attribute identification set of the long short-term memory neural network set of the first stage, combine the several equilibrium loss terms and the numerical loss term to generate multiple first-stage sub-loss functions, sum them to generate the first-stage total loss function, and train the first-stage pollutant concentration distribution model;

[0111] Freeze the long short-term memory neural network set of the first stage, respectively, according to the second-stage pollutant attribute identification set of the long short-term memory neural network set of the second stage, combine the several equilibrium loss terms and the numerical loss term to generate multiple second-stage sub-loss functions, sum them to generate the second-stage total loss function, and train the second-stage pollutant concentration distribution model;

[0112] Until the pollutant concentration distribution model of the B-th stage is trained, which is set as the pollutant concentration distribution model.

[0113] Furthermore, the soil remediation module 15 in the soil remediation control system is further configured to:

[0114] Statistically calculate the proportion of the number of abnormal positions where the concentration deviation at the same position between the pollutant concentration distribution prediction information and the pollutant concentration distribution expected information output by the pollutant concentration distribution model exceeds the concentration deviation threshold;

[0115] When the proportion of the number of abnormal positions is greater than or equal to a preset proportion, adjust the soil remediation plan a preset number of times to obtain an updated soil remediation plan. Among them, the adjustment strategy is to continue adjusting based on the updated soil remediation plan for advantages and return to the soil remediation plan to continue adjusting for disadvantages;

[0116] When the proportion of the number of abnormal positions in the updated soil remediation plan is less than the preset proportion, output the updated soil remediation plan;

[0117] Otherwise, continue to iterate.

[0118] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The foregoing Figure 1 A soil remediation control method and specific example in Embodiment 1 are equally applicable to a soil remediation control system in this embodiment. Through the foregoing detailed description of a soil remediation control method, those skilled in the art can clearly know a soil remediation control system in this embodiment. Therefore, for the sake of simplicity of the specification, it will not be elaborated here. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For the relevant parts, refer to the description in the method part.

[0119] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0120] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application also intends to include these changes and modifications.

Claims

1. A soil remediation control method, characterized in that, Including: Configuring a pollutant concentration distribution model architecture, wherein the pollutant concentration distribution model architecture includes an input layer, a convolutional fully-connected layer, and multiple separate long short-term memory neural networks connected in sequence; Obtaining soil environmental information, pollutant distribution information of a target area, chemical agent distribution information of a soil remediation plan, and several chemical reaction equilibrium equations; Placing the soil environmental information, pollutant distribution information, and chemical agent distribution information in the input layer, placing the pollutant concentration distribution prediction information in the output layer of the long short-term memory neural network, and any one long short-term memory neural network corresponds uniquely to one piece of attribute pollutant concentration distribution prediction information; Constructing several equilibrium loss terms according to the several chemical reaction equilibrium equations, and constructing a numerical loss term according to the output deviation; Combining the several equilibrium loss terms and the numerical loss term respectively according to the multiple pollutant attribute identifiers of the multiple long short-term memory neural networks to generate multiple sub-loss functions, summing them to generate a total loss function, training the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, and adjusting the soil remediation plan to control soil remediation.

2. The method according to claim 1, wherein Constructing several equilibrium loss terms according to the several chemical reaction equilibrium equations, and constructing a numerical loss term according to the output deviation, including: Obtaining a first chemical reaction equilibrium equation and constructing a first equilibrium loss term: a1r1+...+a i r i +...+a Q r Q →b1p1+...+b j p j +...+b Y p Y , Among them, a i and b j represent stoichiometric coefficients, r i represents the reactant, p j represents the product, represents the reactant concentration, k represents the training sample number, represents the predicted product concentration, N represents the number of training samples, a1r1+...+a i r i +...+a Q r Q →b1p1+...+b j p j +...+b Y p Y represents the chemical reaction equilibrium equation; Until the L-th chemical reaction equilibrium equation is used to construct the L-th equilibrium loss term e L , the first equilibrium loss term e1 to the L-th equilibrium loss term e L are added to the several equilibrium loss terms; Constructing the numerical loss term according to the deviation between the supervised data and the output data.

3. The method according to claim 1, characterized in that Combining the several equilibrium loss terms and the numerical loss term respectively according to the multiple pollutant attribute identifiers of the multiple long short-term memory neural networks to generate multiple sub-loss functions, including: Extracting the first pollutant attribute identifier of the first long short-term memory neural network among the multiple long short-term memory neural networks; Extracting a first selected set of equilibrium loss terms whose reactants have the first pollutant attribute identifier from the several equilibrium loss terms; Performing weighted summation on the first selected set of equilibrium loss terms to generate a chemical equilibrium loss function; Summing the chemical equilibrium loss function and the numerical loss term to generate a first sub-loss function and adding it to the multiple sub-loss functions.

4. The method according to claim 3, characterized in that Performing weighted summation on the first selected set of equilibrium loss terms to generate a chemical equilibrium loss function, including: Statistical consumption weight set of co-reactants of the first selected set of equilibrium loss terms when the first pollutant attribute identifier consumes a unit weight, where the co-reactants are the set of substances that react with the first pollutant in any one selected equilibrium loss term; Summing the consumption weight set of co-reactants to obtain the total consumption weight of co-reactants; Calculating the ratio of the consumption weight set of co-reactants to the total consumption weight of co-reactants respectively, and then using 1 minus the ratio to obtain the weight distribution of the selected equilibrium loss terms; Performing weighted summation on the first selected set of equilibrium loss terms according to the weight distribution of the selected equilibrium loss terms to generate a chemical equilibrium loss function.

5. The method according to claim 2, wherein Constructing the numerical loss term according to the deviation between the supervised data and the output data: where, e num represents the numerical loss term, C z,s,1 represents the predicted concentration of pollutants at the s-th position in the z-th training, C z,s,0 represents the actual concentration of pollutants at the s-th position in the z-th training, U represents the number of samples, and I represents the number of positions.

6. The method according to claim 1, wherein Training the pollutant concentration distribution model architecture to generate a pollutant concentration distribution model, including: According to the several chemical reaction equilibrium equations, divide the pollutant attributes of the multiple separated long short-term memory neural networks into reaction stages to obtain a pollutant reaction directed network, where the first pollutant at the arrow end of the pollutant reaction directed network is the origin pollutant of the second pollutant at the arrow start, and the origin pollutant means that the co-reactant of the second pollutant is the first pollutant, or the substance generated by the reaction of the first pollutant and the co-reactant of the first pollutant; According to the pollutant reaction directed network, divide the multiple separated long short-term memory neural networks into stages to obtain the long short-term memory neural network set of the first stage until the long short-term memory neural network set of the Bth stage; Respectively, according to the pollutant attribute identification set of the long short-term memory neural network set of the first stage, combine the several balance loss terms and the numerical loss terms to generate multiple first-stage sub-loss functions, sum them to generate the first-stage total loss function, and train the first-stage pollutant concentration distribution model; Freeze the long short-term memory neural network set of the first stage, respectively, according to the pollutant attribute identification set of the long short-term memory neural network set of the second stage, combine the several balance loss terms and the numerical loss terms to generate multiple second-stage sub-loss functions, sum them to generate the second-stage total loss function, and train the second-stage pollutant concentration distribution model; Until the pollutant concentration distribution model of the Bth stage is trained, set it as the pollutant concentration distribution model.

7. The method according to claim 1, characterized in that, Generate a pollutant concentration distribution model, and adjust the soil remediation plan for soil remediation control, including: Count the proportion of the number of abnormal positions where the concentration deviation at the same position between the pollutant concentration distribution prediction information output by the pollutant concentration distribution model and the expected pollutant concentration distribution information exceeds the concentration deviation threshold; When the proportion of the number of abnormal positions is greater than or equal to the preset proportion, adjust the soil remediation plan a preset number of times to obtain an updated soil remediation plan, where the adjustment strategy is to adjust based on the updated soil remediation plan for advantages and return to the soil remediation plan for adjustment for disadvantages; When the proportion of the number of abnormal positions of the updated soil remediation plan is less than the preset proportion, output the updated soil remediation plan; Otherwise, continue to iterate.

8. A soil remediation control system, characterized in that, The system is used to execute the method described in any one of claims 1 to 7, and the system includes: An architecture configuration module for configuring the architecture of the pollutant concentration distribution model, where the architecture of the pollutant concentration distribution model includes an input layer, a convolutional fully connected layer, and multiple separated long short-term memory neural networks connected in sequence; An information acquisition module for obtaining the soil environment information, pollutant distribution information of the target area, as well as the chemical agent distribution information of the soil remediation plan and several chemical reaction equilibrium equations; A model training module for placing the soil environment information, pollutant distribution information, and chemical agent distribution information on the input layer, placing the pollutant concentration distribution prediction information on the output layer of the long short-term memory neural network, and any one long short-term memory neural network corresponds uniquely to a predicted pollutant concentration distribution information of an attribute pollutant; A loss analysis module, configured to construct a number of equilibrium loss terms according to the several chemical reaction equilibrium equations, and construct a numerical loss term according to the output deviation; A soil remediation module, configured to respectively combine the several equilibrium loss terms and the numerical loss term according to the multiple pollutant attribute identifications of multiple long short-term memory neural networks, generate multiple sub-loss functions, sum them up to generate a total loss function, train the pollutant concentration distribution model architecture, generate a pollutant concentration distribution model, and adjust the soil remediation plan to control soil remediation.

Images