Closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic pollutant removal

Abstract

This invention provides a closed-loop optimization method for soil organic matter carbon cycle regulation and pollutant synergistic removal, belonging to the field of environmental remediation technology based on machine learning. First, time-series data on soil physicochemical properties, soil pollutant concentrations, environmental meteorological conditions, and anthropogenic control parameters of the target area are acquired to construct a dataset. Second, a designed soil carbon cycle-pollutant migration coupling prediction model is used to obtain predicted values ​​of soil organic matter content and pollutant concentrations for future time steps. Finally, a set of decision variables is established, and a dual objective function and hard constraints are set. The trained migration coupling prediction model is used as a surrogate model and input into a non-dominated sorting genetic algorithm for global optimization to extract the Pareto front solution set. Finally, the optimal combination of control parameters is obtained through an approximation-ideal-solution sorting method. This invention achieves high-precision joint prediction of soil organic matter dynamics and pollutant migration evolution through a designed multi-objective closed-loop optimization strategy and prediction model.

Description

Technical Field

[0001] This invention belongs to the field of environmental remediation technology based on machine learning, and particularly relates to a closed-loop optimization method for regulating soil organic carbon cycle and synergistic removal of pollutants. Background Technology

[0002] With the continuous advancement of agricultural intensification and industrialization, soil pollution and carbon loss have become increasingly prominent issues. Soil organic matter, as a core component of the soil carbon pool, directly impacts soil fertility, structural stability, and carbon sequestration function. Simultaneously, the accumulation of heavy metals and organic pollutants in the soil not only threatens agricultural product safety but also indirectly accelerates organic matter decomposition by altering the soil microbial community structure, creating a vicious cycle of carbon loss and exacerbated pollution. Traditional soil remediation and carbon management methods are mostly implemented independently, lacking the dynamic perception and synergistic control capabilities of the coupling relationship between carbon cycling and pollutant migration, making it difficult to achieve efficient pollutant removal while ensuring the stability of the soil carbon pool.

[0003] Currently, the main methods for regulating soil organic carbon cycle and removing pollutants include the following: (1) Pollutant immobilization method based on single chemical modification: This method mainly involves adding amendments such as lime, phosphate, or biochar to the soil to cause heavy metals to precipitate or be adsorbed and immobilized, thereby reducing their bioavailability. The operation process is relatively mature and easy to implement in the field. However, this method usually relies on the amount of additives applied based on experience, making it difficult to adjust the ratio of amendments in real time according to the dynamic changes in soil organic matter. Furthermore, excessive addition of alkaline substances may damage the activity of soil microorganisms, accelerate the mineralization of organic matter, and lead to the loss of carbon pool, thus limiting its application in balancing carbon cycle stability and pollution remediation. (2) Pollutant degradation methods based on bioremediation: These methods utilize biological metabolic processes to degrade organic pollutants or enrich and remove heavy metals by inoculating functional microorganisms or planting hyperaccumulating plants. These methods are environmentally friendly. However, due to the complex and variable soil environment, the bioremediation process is easily affected by factors such as temperature, humidity, and nutrient supply, making it difficult to guarantee stable and efficient operation. Furthermore, existing bioremediation methods often focus on pollutant removal efficiency, lacking simultaneous monitoring and control of soil carbon cycle processes, which may lead to accelerated decomposition of soil organic matter during the remediation process. (3) Agricultural management methods based on empirical rules: Some studies and engineering practices have attempted to increase soil organic matter content through crop rotation, straw return to the field, or application of organic fertilizer, while hoping to reduce pollutant concentration through dilution effects. However, these methods rely on long-term experience accumulation, lack scientific analysis of the coupling mechanism between pollutant migration pathways and carbon cycle, and are difficult to formulate precise joint control schemes for different pollution types and soil conditions, thus failing to achieve multi-objective optimization of carbon stabilization and pollution removal.

[0004] In summary, existing methods for regulating soil organic carbon cycle and removing pollutants have certain limitations in terms of multi-source information perception, analysis of carbon-pollution coupling mechanisms, and collaborative optimization decision-making: single chemical improvement methods are difficult to take into account carbon pool stability, bioremediation methods are greatly affected by environmental fluctuations and lack synchronous regulation of carbon cycle, and management methods based on empirical rules are difficult to achieve multi-objective precise optimization and dynamic feedback. Summary of the Invention

[0005] To address the above problems, this invention provides a closed-loop optimization method for regulating soil organic carbon cycle and synergistically removing pollutants, comprising the following processes: S1. Obtain time-series data of soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and human-controlled parameters in the target area to construct a carbon-pollutant coupled dataset. S2. Construct a soil carbon cycle-pollutant migration coupled prediction model based on physical information. Input four types of time series data into the time series feature encoding module. Use a long short-term memory network and a feedforward attention network to perform feature selection and global time series context weight calculation to obtain time series attention features. Input the time series attention features into the coupled prediction output module to obtain the predicted values ​​of soil organic matter content and pollutant concentration for future time steps. S3. Establish the set of decision variables and set the dual objective function and hard constraints. Use the trained soil carbon cycle-pollutant migration coupled prediction model as a surrogate model and input it into the non-dominated sorting genetic algorithm for global optimization to obtain the Pareto front solution set. Construct the Pareto front solution set into the initial evaluation matrix and obtain the optimal combination of control parameters by approximating the ideal solution sorting method.

[0006] Preferably, the soil physicochemical index time series data includes soil organic matter content, total carbon content, microbial biomass carbon, pH value, moisture content and electrical conductivity; The soil pollutant concentration time series data includes soil heavy metal concentration and polycyclic aromatic hydrocarbon concentration. The heavy metals specifically include three elements: cadmium, lead, and arsenic. The polycyclic aromatic hydrocarbons specifically include three components: phenanthrene, pyrene, and benzo[a]pyrene. The time-series data of the environmental meteorological parameters include air temperature, air humidity, precipitation, and light intensity data; The time-series data of the artificially controlled parameters include the amount of soil conditioner added, irrigation water volume, tillage depth, and straw return to the field.

[0007] Preferably, the soil carbon cycle-pollutant migration coupled prediction model first inputs four types of time-series data into the time-series feature encoding module, and performs feature filtering and global time-series context weight calculation through a long short-term memory network and a feedforward attention network to obtain time-series attention features; the time-series attention features are then input into a coupled prediction output module containing two parallel branches and interactive attention units to obtain predicted values ​​of soil organic matter content and pollutant concentration for future time steps. A soil carbon cycle-pollutant migration coupled prediction model was trained using a carbon-pollutant coupled dataset. During the training process, soil carbon dynamics residuals and pollutant migration-transformation residuals were embedded as constraints, and a joint loss function was constructed with the data loss.

[0008] Preferably, the temporal feature encoding module comprises two layers, and each layer has Long Short-Term Memory (LSTM) networks with hidden units and containing The system consists of a feedforward attention network with 10 neurons. First, four types of soil time-series data—soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and human-controlled parameters—are concatenated according to their feature dimensions to form a multi-dimensional input vector, which is then fed into the Long Short-Term Memory (LSTM) network. Information is filtered through a triple-gating mechanism consisting of a forget gate, an input gate, and an output gate. The forget gate automatically discards redundant environmental noise information by comparing the current input with the previous hidden state, while the input gate captures sudden changes in water content caused by rainfall and concentration pulses resulting from human control, updating these features in the cell state. Next, the low-level spatiotemporal feature sequence extracted by the first-layer LSTM is directly fed into the second-layer LSTM to capture long-range dependencies across time periods at a higher semantic dimension. Finally, a feedforward attention network is introduced above the LSM output, using an additive attention mechanism to calculate the weight contributions of the hidden states at different times in the LSM network. The final output of the LSM network contains global temporal context information. Dimensional temporal attention feature vector.

[0009] Preferably, the coupled prediction output module consists of two parallel multilayer perceptrons and an interactive attention unit embedded therebetween; the first branch is the soil organic matter prediction branch, which is composed of... The network consists of cascaded fully connected layers, each using the Tanh activation function; the second branch is for pollutant concentration prediction, and its network topology is completely symmetrical with the first branch; the interactive attention unit is deployed between the penultimate layer of the two branches, and includes a query matrix mapping operator, a key-value matrix mapping operator, and cross-weight calculation; the output of the temporal feature encoding module is... The time-series attention feature vector is simultaneously fed into two parallel branches. In the first branch, a nonlinear mapping is used to extract a feature vector representing the carbon cycle evolution potential, while in the second branch, a feature vector representing the pollutant migration trajectory is extracted. When the data stream reaches the penultimate layer, the interactive attention unit uses a mapping operator to calculate the similarity between the two feature vectors, generating a cross-weight score. This dynamically represents the nonlinear correlation between soil organic matter state and the effectiveness of pollutant adsorption sites within the latent space. The weighted feature stream enters the final output layer, where the first branch directly outputs the future... The predicted soil organic matter content C for each time step is directly output by the second branch for the future time step. Predicted pollutant concentration values ​​at a time step .

[0010] Preferably, the embedding of soil carbon kinetic residuals and pollutant migration-transformation residual constraints during the training process specifically includes: First, the partial derivatives of the predicted values ​​output by the coupled prediction output module with respect to time and spatial coordinates are obtained using an automatic differential engine, and the soil carbon decomposition kinetic equation and the pollutant convection and diffusion equation are transformed into physical residuals; the soil carbon kinetic residuals are defined. as follows: ; in The predicted value of soil organic matter content output by the coupled prediction output module; The carbon input flux is calculated by linearly weighting the collected straw return to the field and the soil conditioner application amount using a preset carbon conversion coefficient. Based on the decomposition rate, The temperature and moisture correction function is based on the collected air temperature. With moisture content As independent variables, the product of the Lloyd-Taylor temperature correction term and the normalized moisture correction term is used to obtain the nonlinear mapping. Define pollutant migration-transformation residuals as follows: ; in The predicted pollutant concentration output by the coupled prediction output module; The dispersion coefficient; The longitudinal velocity of soil pore water is obtained by dynamically solving a one-dimensional water infiltration model by combining the collected precipitation and irrigation water volume with the real-time soil moisture content. and For the predicted value of soil organic matter content Nonlinear functions that are controlled; Among them, the blocking factor The degradation rate was calculated using a nonlinear physical mapping method based on the linear isothermal adsorption theory. The calculations were performed using Monod Kinetics nonlinear equations.

[0011] Preferably, in step S3, establishing the set of decision variables and setting a bi-objective function and hard constraints specifically involves: First, define the set of decision variables. ,in This refers to the dosage of soil conditioner. For irrigation water volume, For tillage depth, This refers to the amount of straw returned to the field; Secondly, a surrogate model mapping operator is established, and the trained soil carbon cycle-pollutant migration coupling prediction model is used as the surrogate model. By inputting decision variables In addition, the collected time-series data of soil physicochemical indicators, soil pollutant concentrations, and environmental meteorological parameters are directly mapped and output for the future... Predicted values ​​of soil organic matter content at each time step Compared with the predicted pollutant concentration ; Redefine the dual objective function: Objective 1 is set as the predicted value of soil organic matter content and the target steady-state value. Minimize the absolute deviation between them, i.e. Objective two is set as minimizing the predicted pollutant concentration, i.e. ; Finally, hard constraints on the decision variables are set based on actual agronomic standards, namely... , , as well as ,in It is 500 kg per mu. It is 100 cubic meters per mu. It is 30 centimeters. The target is 1000 kg per mu, ensuring that the control measures are within the physical boundaries of safe production.

[0012] Preferably, the sorting method for approximating the ideal solution in S3 is as follows: First, the Pareto front solution set is... The bi-objective prediction results are constructed as a dimension of The initial evaluation matrix is ​​then transformed into a standardized matrix by performing vector normalization. ; Secondly, based on the standardized matrix Extract the numerical extrema of each target column and construct positive ideal solution vectors for each. With negative ideal solution vector The positive ideal solution vector From the standardized matrix The minimum value elements of each objective column in the negative ideal solution vector represent the theoretical optimal boundary in the current solution space. From the standardized matrix The maximum value element in each objective column represents the theoretical worst boundary in the current solution space; We then calculate the Euclidean distances of each decision variable to the positive and negative ideal solutions within the standardized feature space. and Calculate the relative similarity of each scheme. , The Pareto front solution set Relative closeness The set of decision variables with the largest values ​​is taken as the optimal combination of control parameters. .

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. Design of a Physically Constrained Coupling Prediction Architecture for Soil Carbon Cycle and Pollutant Migration; This invention constructs a neural network-based joint prediction architecture for soil spatiotemporal evolution that integrates physical information. This architecture dynamically maps the regulatory mechanism of soil organic matter state changes on pollutant adsorption effectiveness through a long short-term memory network and interactive attention units. Simultaneously, embedding physical residual constraints of soil carbon decomposition kinetics and pollutant diffusion during network training improves the reliability of multi-indicator collaborative prediction under complex environmental fluctuations. 2. Intelligent optimization and decision-making control strategy for the dual objectives of maintaining carbon stability and decontamination; This invention designs a multi-objective joint optimization and multi-criteria decision-making system based on a non-dominated sorting genetic algorithm. This system uses a trained coupled prediction network as a surrogate model, combined with agronomic production boundary constraints, to generate a Pareto front solution set for maintaining carbon stability and reducing pollutant concentrations. Subsequently, an approximation-ideal-solution sorting method is introduced to extract the target extrema, screening out suitable combinations of control parameters that balance soil fertility protection and synergistic pollutant removal. 3. A closed-loop control and online fine-tuning system driven by data self-iteration and error awareness; This invention establishes a closed-loop environmental control workflow encompassing multi-source sensing, parameter optimization, terminal execution, and model evolution. The system transforms the control parameters generated by decision-making into physical execution commands and simultaneously collects real environmental response data as iterative feedback after intervention. When the prediction mean square error exceeds a preset threshold, the system automatically extracts the latest time-series data to trigger network weight fine-tuning, thereby adapting to the long-term dynamic drift characteristics of the soil environment. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0016] Figure 2 This is a flowchart illustrating the process of collecting multi-source heterogeneous soil data and constructing a carbon-pollutant coupled dataset for this invention.

[0017] Figure 3 This is a diagram of the soil carbon cycle-pollutant migration coupled prediction model architecture of the present invention.

[0018] Figure 4 This is a flowchart illustrating the multi-objective collaborative optimization strategy generation process of the present invention.

[0019] Figure 5 This is a graph showing the prediction accuracy of soil organic matter (SOM) in the examples.

[0020] Figure 6 This is a graph showing the accuracy analysis of soil pollutant concentration prediction in the examples.

[0021] Figure 7 This is a diagram illustrating the synergistic removal of pollutants and verification of carbon stability in the examples. Detailed Implementation

[0022] This invention proposes a closed-loop optimization method for regulating soil organic carbon cycle and synergistically removing pollutants. The overall process is as follows: Figure 1 As shown: First, time-series data of soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and anthropogenic control parameters for the target area are acquired and preprocessed to obtain four types of time-series data. Feature vector concatenation and sliding window segmentation mapping are then performed on these four types of time-series data to construct and partition a carbon-pollutant coupled dataset. Second, the four types of time-series data are input into a time-series feature encoding module. Feature selection and global time-series context weight calculation are performed using a long short-term memory network and a feedforward attention network to obtain time-series attention features. These time-series attention features are then input into a coupled prediction output module containing two parallel branches and an interactive attention unit to obtain predicted values ​​for soil organic matter content and pollutant concentrations at future time steps. A soil carbon cycle-pollutant migration coupled prediction model was trained using a carbon-pollutant coupled dataset. During the training process, soil carbon dynamics residuals and pollutant migration-transformation residuals were embedded as constraints, and a joint loss function was constructed with data loss to obtain the trained model. Finally, a set of decision variables was established, and a dual objective function and hard constraints were set. The trained soil carbon cycle-pollutant migration coupled prediction model was used as a surrogate model and input into a non-dominated sorting genetic algorithm for global optimization to extract the Pareto front solution set. The Pareto front solution set was used to construct an initial evaluation matrix and input into an approximation ideal solution sorting method. The optimal combination of control parameters was obtained by calculating the Euclidean distance and relative proximity to the positive and negative ideal solutions in the feature space.

[0023] The specific implementation process of the present invention will be further described below with reference to specific embodiments.

[0024] S1. Collaborative acquisition of multi-source heterogeneous soil data and construction of carbon-pollutant coupled datasets Time-series data of soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and anthropogenic control parameters in the target area were acquired and preprocessed to obtain four types of time-series data. Feature vector concatenation and sliding window segmentation were performed on these four types of time-series data to construct and partition a carbon-pollutant coupled dataset. The overall process is as follows: Figure 2 As shown.

[0025] S1-1 Soil physicochemical index collection and spatial topology calibration; a soil sensor array consisting of a nondispersive infrared carbon sensor, an ion-selective electrode, and a frequency domain reflectometer is deployed in a grid pattern in the target farmland or remediation site, with the grid spacing set to [value missing]. The sensor array collects multidimensional physicochemical indicators including soil organic matter content, total carbon content, microbial biomass carbon, pH value, moisture content, and electrical conductivity; the sampling depth is set to... The sampling frequency is set to a centimeter-deep cultivated layer. The total collection period is set to This ensures coverage of the entire crop growth cycle or remediation phase; finally, the collected data is organized according to time series and assigned spatial location indexes to obtain time series data of soil physicochemical indicators. S1-2 Soil pollutant concentration collection; soil heavy metal concentration was simultaneously collected at the same spatial sampling points as S1-1. With polycyclic aromatic hydrocarbon concentration The collection of data; the heavy metals specifically include cadmium, lead, and arsenic; the polycyclic aromatic hydrocarbons specifically include phenanthrene, pyrene, and benzo[a]pyrene; the collection frequency and the collection frequency of soil physicochemical indicators. Maintain synchronization and alignment; the above data were obtained through in-situ monitoring; the soil heavy metal concentration and polycyclic aromatic hydrocarbon concentration were organized into time series data of soil pollutant concentration, thereby quantitatively characterizing the spatiotemporal evolution intensity of pollutants; S1-3 Environmental meteorological parameter acquisition; First, a miniature automatic weather station is deployed at the center of the monitoring area, and the weather station transmits data at a frequency... Real-time data collection of air temperature, air humidity, precipitation, and light intensity is converted into time-series data of environmental meteorological parameters; S1-4 Records of artificial control parameters; records the amount of soil conditioner added, irrigation water, tillage depth and straw return to the field implemented during the monitoring period; finally, the above discrete control measures are mapped to the time axis according to the implementation date to obtain the time series data of artificial control parameters; Construction and sample segmentation of the S1-5 carbon-pollutant coupled dataset; using four types of time-series data collected from S1-1 to S1-4 as the input feature matrix, and then using the future... Soil organic matter content and soil pollutant concentration at each time step were used as prediction labels; according to the time window length... Step size is The sliding window segmentation operation is performed in a certain way; finally, the segmented time-series sample blocks are associated to construct a carbon-pollutant coupled dataset, and divided into training set, validation set and test set according to a preset ratio, which serve as the model construction dataset.

[0026] S2. Construct a soil carbon cycle-pollutant migration coupled prediction model based on physical information. The architecture diagram of the soil carbon cycle-pollutant migration coupled prediction model is as follows: Figure 3As shown, four types of time-series data are input into the time-series feature encoding module. Feature selection and global time-series context weight calculation are performed through a long short-term memory network and a feedforward attention network to obtain time-series attention features. The time-series attention features are then input into a coupled prediction output module containing two parallel branches and interactive attention units to obtain predicted values ​​of soil organic matter content and pollutant concentration for future time steps. The soil carbon cycle-pollutant migration coupled prediction model is trained using a carbon-pollutant coupled dataset. During the training process, soil carbon dynamics residuals and pollutant migration-transformation residual constraints are embedded, and a joint loss function is constructed with the data loss to obtain the trained model.

[0027] S2-1 Temporal Feature Encoding Module Design: Four types of time-series data are concatenated according to feature dimensions to form a multi-dimensional input vector; this vector is then input into... Layers, each layer has In a Long Short-Term Memory (LSTM) network with hidden units, redundant environmental noise information is discarded through a forgetting gate, and key environmental change features are filtered through an input gate, particularly capturing sudden changes in water content caused by rainfall and concentration pulses caused by human intervention, and updating cell states. An additive attention mechanism is introduced on top of the LSTM network output, utilizing... A feedforward network of individual neurons calculates the weight contributions of the hidden states at different time points, ultimately converging to generate a dataset containing global temporal contextual information. 3D temporal attention feature vector; The module consists of two layers, each with... Long Short-Term Memory (LSTM) networks with hidden units and containing The system consists of a feedforward attention network with 10 neurons. First, four types of soil time-series data—soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and human-controlled parameters—are concatenated according to their feature dimensions to form a multi-dimensional input vector, which is then fed into the Long Short-Term Memory (LSTM) network. Information filtering is achieved through a triple-gating mechanism consisting of a forget gate, an input gate, and an output gate. The forget gate automatically discards redundant environmental noise information by comparing the current input with the previous hidden state, while the input gate accurately captures sudden changes in water content caused by rainfall and concentration pulses resulting from human control, updating these features in the cell state. Next, the low-level spatiotemporal feature sequence extracted by the first-layer LSTM is directly fed into the second-layer LSTM to capture long-range dependencies across periods at a higher semantic dimension, overcoming the vanishing gradient problem of soil parameters during long-term evolution. Finally, a feedforward attention network is introduced above the LSM output, using an additive attention mechanism to calculate the weight contribution of the hidden states at different times in the LSM network. The final output of the LSM network contains global temporal context information. 3D temporal attention feature vector; S2-2 Coupled Prediction Output Module Design; This module mainly consists of two parallel multilayer perceptrons and an interactive attention unit embedded between them; The first branch is the soil organic matter prediction branch, which is composed of... The network consists of cascaded fully connected layers, each using the Tanh activation function; the second branch is for pollutant concentration prediction, and its network topology is completely symmetrical with the first branch; the interactive attention unit is deployed between the penultimate layer of the two branches, and includes a query matrix mapping operator, a key-value matrix mapping operator, and cross-weight calculation; the output of S2-1... The time-series attention feature vector is simultaneously fed into two parallel branches. In the first branch, a nonlinear mapping is used to extract a feature vector representing the carbon cycle evolution potential, while in the second branch, a feature vector representing the pollutant migration trajectory is extracted. When the data stream reaches the penultimate layer, the interactive attention unit uses a mapping operator to calculate the similarity between the two feature vectors, generating a cross-weight score. This dynamically represents the nonlinear correlation between soil organic matter state and the effectiveness of pollutant adsorption sites within the latent space. The weighted feature stream enters the final output layer, where the first branch directly outputs the future... The predicted soil organic matter content C for each time step is directly output by the second branch for the future time step. Predicted pollutant concentration values ​​at a time step ; S2-3 Model Training and Loss Function Design Based on Physical Information Constraints; This step aims to embed the prior dynamic equations as soft constraints into the optimization objective. First, the predicted values ​​output by the coupled prediction output module in S2-2 are processed using an automatic differentiation engine. and Partial derivatives are calculated with respect to time and space coordinates, and the soil carbon decomposition kinetics equation and pollutant convection-diffusion equation are transformed into physical residuals. The soil carbon kinetic residuals are defined. as follows: ; in The predicted value of soil organic matter content output by the coupled prediction output module; The carbon input flux is calculated by linearly weighting the amount of straw returned to the field and the amount of soil conditioner added by S1 using a preset carbon conversion coefficient. Based on the decomposition rate, The temperature and moisture correction function is based on the air temperature collected by S1. With moisture content As independent variables, the product of the Lloyd-Taylor temperature correction term and the normalized moisture correction term is used to obtain the nonlinear mapping. Define pollutant migration-transformation residuals as follows: ; in The predicted pollutant concentration output by the coupled prediction output module; The dispersion coefficient; The longitudinal velocity of soil pore water is obtained by dynamically solving a one-dimensional water infiltration model by combining the sum of precipitation and irrigation water collected by S1 with the real-time soil moisture content. and For the predicted value of soil organic matter content Nonlinear functions that are controlled; Among them, the blocking factor The calculation is performed using a nonlinear physical mapping based on the linear isothermal adsorption theory: ; In the formula This refers to the dry bulk density of the soil. Real-time soil moisture content collected for S1; The organic carbon partition coefficient of the target pollutant; this form rigorously anchors the inhibitory and attenuating effect of the soil solid phase organic carbon pool on pollutant migration at the physical level.

[0028] Degradation rate Calculations were performed using the Monod Kinetics nonlinear equations: ; In the formula This is the maximum limiting rate constant for the degradation of the target pollutant by microorganisms; It is a half-saturation constant; this form accurately describes the predicted values ​​of soil organic matter content. The nonlinear degradation process of pollutants by soil microbial communities under the core carbon source-driven mechanism.

[0029] Secondly, construct a data-driven loss term. Soil carbon kinetic residual loss term and pollutant migration and transformation residual loss items The weighted multi-task total loss function is as follows: ; Using homoscedastic uncertainty strategy for weighting coefficients , and Dynamic adaptive adjustment is performed; end-to-end iterative training is conducted on the carbon and pollutant coupled dataset using the Adam optimizer, with an initial learning rate set at... With batch size When the validation set loss is continuous When the wheel stops descending, an early stop mechanism is triggered, and the trained soil carbon cycle-pollutant migration coupled prediction model is output.

[0030] For the data-driven loss term in the multi-task total loss function Its physical meaning is to measure the absolute deviation between the forward propagation prediction of the model network and the actual sensor observation at the same spatiotemporal coordinate anchor point. Its calculation formula is as follows: ; In the formula This represents the total number of valid spatiotemporal observation samples extracted within the training time window. and The coupling prediction output module is in the th Predicted values ​​of soil organic matter content and pollutant concentration at each spatiotemporal coordinate point; and These represent the actual organic matter content label value and pollutant concentration label value obtained from the sensor array and sampling analysis at the corresponding spatiotemporal coordinate points. By minimizing this data-driven loss term, the prediction model is ensured to approximate the actual soil environment evolution trajectory as closely as possible within the theoretical boundaries that conform to physical constraints.

[0031] S3. Generation of a multi-objective optimization strategy for synergistic pollutant removal under soil carbon stability constraints. like Figure 4 As shown, a set of decision variables is established and a dual objective function and hard constraints are set. The trained soil carbon cycle-pollutant migration coupled prediction model is used as a surrogate model and input into a non-dominated sorting genetic algorithm for global optimization to extract the Pareto front solution set. The Pareto front solution set is used to construct an initial evaluation matrix and input into the approximation ideal solution sorting method. The optimal combination of control parameters is obtained by calculating the Euclidean distance and relative proximity to the positive and negative ideal solutions in the feature space.

[0032] S3-1 Definition and Constraint Setting of Multi-Objective Optimization Problem; This step aims to establish the search space and evaluation index of the optimization algorithm; First, define the set of decision variables. ,in This refers to the dosage of soil conditioner. For irrigation water volume, For tillage depth, This refers to the amount of straw returned to the field; Secondly, a surrogate model mapping operator is established, using the soil carbon cycle-pollutant migration coupled prediction model trained by S2 as the surrogate model. By inputting decision variables In addition, the time-series data of soil physicochemical indicators, soil pollutant concentrations, and environmental meteorological parameters collected by S1 are directly mapped and output for the future... Predicted values ​​of soil organic matter content at each time step Compared with the predicted pollutant concentration ; Redefine the dual objective function: Objective 1 is set as the predicted value of soil organic matter content and the target steady-state value. Minimize the absolute deviation between them, i.e. Objective two is set as minimizing the predicted pollutant concentration, i.e. ; Finally, hard constraints on the decision variables are set based on actual agronomic standards, namely... , , as well as ,in It is 500 kg per mu. It is 100 cubic meters per mu. It is 30 centimeters. The yield is 1000 kg per mu, thus ensuring that the control plan is within the physical boundaries of safe production; S3-2 Proxy Model-Assisted Non-Dominated Sorting Genetic Algorithm (NSGA-II) Optimization; This step aims to discover the optimal solution set through heuristic search; First, initialize the size to... The population is divided into two groups, each individual representing a randomly generated combination of decision variables. Next, a global optimization based on a surrogate model is performed. During each generation of evolution, the algorithm batches individuals from the population into the surrogate model mapping operator, utilizing the forward inference capability of the neural network to replace physical field experiments and quickly calculate the objective function value for each combination of decision variables. Then, genetic operator operations are performed, configuring simulated binary crossover and polynomial mutation operators, with the crossover probability set to [value missing]. The mutation probability is set to By performing non-dominated sorting and crowding distance calculation, dominant individuals with good distribution and convergence are retained for the next generation; finally, iterative evolution is performed. After this, the algorithm converges and generates a sequence of algorithms. The Pareto frontier solution set consisting of several non-dominant candidate control schemes ; S3-3 Pareto Front Multi-Criterion Decision Making and Parameter Screening Based on the Approximation Ideal Solution Ranking Method (TOPSIS); This step aims to scientifically screen out the final implementation scheme from the Pareto front solution set; First, the Pareto front solution set is... The bi-objective prediction results are constructed as a dimension of The initial evaluation matrix is ​​then transformed into a standardized matrix by performing vector normalization. ; Secondly, based on the standardized matrix Extract the numerical extrema of each target column and construct positive ideal solution vectors for each. With negative ideal solution vector Specifically, since both the soil organic matter steady-state deviation and pollutant concentration are optimization objectives that are minimization indicators, the positive ideal solution vector... From the standardized matrix The minimum value elements of each objective column in the negative ideal solution vector represent the theoretical optimal boundary in the current solution space. From the standardized matrix The maximum value element in each objective column represents the theoretical worst boundary in the current solution space; We then calculate the Euclidean distances of each decision variable to the positive and negative ideal solutions within the standardized feature space. and The relative similarity of each scheme is calculated according to the following formula. : ; The Pareto front solution set Relative closeness The set of decision variables with the largest values ​​is taken as the optimal combination of control parameters. ; It includes the optimal amount of soil conditioner, the optimal amount of irrigation water, the optimal tillage depth, and the optimal amount of straw returned to the field.

[0033] S4. Dynamic Feedback of Closed-Loop Control Parameter Execution and Prediction Model S4-1 Issuance and Execution of Control Parameters; First, the optimal combination of control parameters obtained from S3 is selected... The data is transmitted to the field intelligent execution terminal via the Internet of Things communication module; then, the field intelligent execution terminal automatically adjusts the application rate of the fertilizer application equipment, the opening degree of the solenoid valve of the irrigation system, and the working depth of the tillage machinery according to the received parameters; finally, it completes the application of a specific dose of soil conditioner, straw return to the field, quantitative water and fertilizer integrated irrigation, and tillage operations. S4-2 Monitoring and Feedback Data Processing of Remediation Effectiveness: First, in the next monitoring cycle after the implementation of the control measures, multidimensional raw monitoring data are re-collected using the soil sensor array, mobile sampling terminal, and micro automatic weather station described in S1. Second, the collected data undergoes quality verification and outlier removal, and feature extraction is performed according to the same preset sliding window logic as in S1. Finally, the processed data is organized into closed-loop feedback time-series data according to the time series to quantitatively characterize the true response state of the soil environment after control. S4-3 Online fine-tuning and dynamic deployment of the prediction model; first, the mean square error between the predicted value output by the soil carbon cycle-pollutant migration coupled prediction model deployed in S2 and the newly collected true values ​​obtained in S4-2 is calculated; second, when the mean square error exceeds the preset deviation threshold... At that time, the online fine-tuning mechanism is triggered, and the most recent... Closed-loop feedback time-series data at time steps to fine-tune the learning rate. Execute on the current running model The network weights are updated and trained in rounds; finally, the fine-tuned model is redeployed and overlaid on the original decision system to ensure that the prediction core module can effectively overcome the long-term non-stationary drift of the soil environment and maintain the long-term operating accuracy of the closed-loop control system.

[0034] Experimental verification and analysis: To verify the effectiveness of the closed-loop optimization method for soil organic carbon cycle regulation and synergistic pollutant removal proposed in this invention, this experiment used a typical cadmium-contaminated farmland as the research object, and the experimental period covered the entire crop growing season. The experiment aimed to demonstrate the accuracy of the system's joint prediction in complex soil environments and the synergistic optimization effect of the multi-objective regulation scheme in practical implementation through measured data.

[0035] 1. Experimental Design and Dataset Construction Thirty-six monitoring points were set up in a 10-meter by 10-meter grid within the experimental area. From two weeks before sowing to one week after harvest, soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and human-controlled parameters were continuously collected, with a sampling frequency of once every seven days. The experiment obtained raw monitoring data for 28 time steps. After being divided into four-week sliding windows, a carbon-pollutant coupled dataset was constructed, with the training set accounting for 70%, and the validation and test sets each accounting for 15%.

[0036] 2. Summary table of experimental results Table 1 shows the model's predictive performance on the independent test set, as well as the changes in key soil remediation indicators after 4 weeks of implementing the closed-loop control strategy.

[0037] Table 1 Summary of Experimental Results

[0038] Figure 5The model demonstrates the fitting relationship between measured and model-predicted values ​​of soil organic matter content within the test set. Because the model incorporates physical constraints based on first-order kinetic equations, the predicted points are highly concentrated near the 1:1 ideal prediction line, with a coefficient of determination reaching 0.89. This indicates that the model can not only capture the seasonal trend of organic matter changes over time using LSTM, but also effectively correct numerical drift that easily occurs in purely data-driven models under small sample conditions using physical operators, accurately reflecting the slow evolution logic of the soil carbon pool.

[0039] Figure 6 This reflects the model's predictive ability for heavy metal and organic pollutant concentrations. Experimental results show that the goodness of fit between the predicted and measured values ​​is 0.85, and the mean square error is only 0.031. Although soil pollutants exhibit strong spatial heterogeneity due to external disturbances such as irrigation and tillage, this invention extracts the influence characteristics of carbon cycle on pollutant adsorption sites through interactive attention coupling units. This allows the model to sensitively perceive changes in pollutant migration rates caused by fluctuations in organic matter, thereby maintaining high-precision joint prediction performance.

[0040] Figure 7 This demonstrates the comprehensive effectiveness of implementing closed-loop optimization and control. The left vertical axis compares the changes in pollutant concentrations before and after optimization. After four weeks of precise control, the concentrations of cadmium (Cd) and polycyclic aromatic hydrocarbons (PAHs) in the soil significantly decreased by 18.7% and 14.2%, respectively. Meanwhile, the right vertical axis shows the deviation of soil organic matter from the target steady state. The diamond-shaped data points show that the SOM content of all samples was strictly maintained within the range of 95% to 105% of the target value. This proves that the multi-objective synergistic optimization strategy in step S3 can effectively balance the contradictory objectives of pollution removal and carbon retention, achieving efficient pollutant removal without sacrificing soil fertility and carbon pool stability.

[0041] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0042] While the above description illustrates specific embodiments of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A closed-loop optimization method for regulating soil organic carbon cycle and synergistically removing pollutants, characterized in that, Includes the following processes: S1. Obtain time-series data of soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and human-controlled parameters in the target area to construct a carbon-pollutant coupled dataset. S2. Construct a soil carbon cycle-pollutant migration coupled prediction model based on physical information. Input four types of time-series data into the time-series feature encoding module. Use a long short-term memory network and a feedforward attention network to perform feature selection and global time-series context weight calculation to obtain time-series attention features. The temporal attention features are input into the coupled prediction output module to obtain the predicted values ​​of soil organic matter content and pollutant concentration for future time steps; S3. Establish the set of decision variables and set the dual objective function and hard constraints. Use the trained soil carbon cycle-pollutant migration coupled prediction model as a surrogate model and input it into the non-dominated sorting genetic algorithm for global optimization to obtain the Pareto front solution set. Construct the Pareto front solution set into the initial evaluation matrix and obtain the optimal combination of control parameters by approximating the ideal solution sorting method.

2. The closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic pollutant removal as described in claim 1, characterized in that: The time-series data of soil physicochemical indicators include soil organic matter content, total carbon content, microbial biomass carbon, pH value, moisture content, and electrical conductivity. The soil pollutant concentration time series data includes soil heavy metal concentration and polycyclic aromatic hydrocarbon concentration. The heavy metals specifically include three elements: cadmium, lead, and arsenic. The polycyclic aromatic hydrocarbons specifically include three components: phenanthrene, pyrene, and benzo[a]pyrene. The time-series data of the environmental meteorological parameters include air temperature, air humidity, precipitation, and light intensity data; The time-series data of the artificially controlled parameters include the amount of soil conditioner added, irrigation water volume, tillage depth, and straw return to the field.

3. The closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic pollutant removal as described in claim 1, characterized in that: The soil carbon cycle-pollutant migration coupled prediction model first inputs four types of time-series data into the time-series feature encoding module, and then performs feature selection and global time-series context weight calculation through a long short-term memory network and a feedforward attention network to obtain time-series attention features. The temporal attention features are input into a coupled prediction output module containing two parallel branches and interactive attention units to obtain predicted values ​​of soil organic matter content and pollutant concentration for future time steps. A soil carbon cycle-pollutant migration coupled prediction model was trained using a carbon-pollutant coupled dataset. During the training process, soil carbon dynamics residuals and pollutant migration-transformation residuals were embedded as constraints, and a joint loss function was constructed with the data loss.

4. The closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic removal of pollutants as described in claim 3, characterized in that: The temporal feature encoding module comprises two layers, and each layer has... Long Short-Term Memory (LSTM) networks with hidden units and containing The system consists of a feedforward attention network with 10 neurons. First, four types of soil time-series data—soil physicochemical indicators, soil pollutant concentrations, environmental meteorological parameters, and human-controlled parameters—are concatenated according to their feature dimensions to form a multi-dimensional input vector, which is then fed into the Long Short-Term Memory (LSTM) network. Information is filtered through a triple-gating mechanism consisting of a forget gate, an input gate, and an output gate. The forget gate automatically discards redundant environmental noise information by comparing the current input with the previous hidden state, while the input gate captures sudden changes in water content caused by rainfall and concentration pulses resulting from human control, updating these features in the cell state. Next, the low-level spatiotemporal feature sequence extracted by the first-layer LSTM is directly fed into the second-layer LSTM to capture long-range dependencies across time periods at a higher semantic dimension. Finally, a feedforward attention network is introduced above the LSM output, using an additive attention mechanism to calculate the weight contributions of the hidden states at different times in the LSM network. The final output of the LSM network contains global temporal context information. Dimensional temporal attention feature vector.

5. The closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic removal of pollutants as described in claim 4, characterized in that: The coupled prediction output module consists of two parallel multilayer perceptrons and an interactive attention unit embedded between them; the first branch is the soil organic matter prediction branch, which is composed of... The network consists of cascaded fully connected layers, each using the Tanh activation function; the second branch is for pollutant concentration prediction, and its network topology is completely symmetrical with the first branch; the interactive attention unit is deployed between the penultimate layer of the two branches, and includes a query matrix mapping operator, a key-value matrix mapping operator, and cross-weight calculation; the output of the temporal feature encoding module is... The time-series attention feature vector is simultaneously fed into two parallel branches. In the first branch, a nonlinear mapping is used to extract a feature vector representing the carbon cycle evolution potential, while in the second branch, a feature vector representing the pollutant migration trajectory is extracted. When the data stream reaches the penultimate layer, the interactive attention unit uses a mapping operator to calculate the similarity between the two feature vectors, generating a cross-weight score. This dynamically represents the nonlinear correlation between soil organic matter state and the effectiveness of pollutant adsorption sites within the latent space. The weighted feature stream enters the final output layer, where the first branch directly outputs the future... The predicted soil organic matter content C for each time step is directly output by the second branch for the future time step. Predicted pollutant concentration values ​​at a time step .

6. The closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic removal of pollutants as described in claim 3, characterized in that: The specific steps involved embedding soil carbon kinetic residuals and pollutant migration-transformation residual constraints during the training process are as follows: First, the partial derivatives of the predicted values ​​output by the coupled prediction output module with respect to time and spatial coordinates are obtained using an automatic differential engine, and the soil carbon decomposition kinetic equation and the pollutant convection and diffusion equation are transformed into physical residuals; the soil carbon kinetic residuals are defined. as follows: ； in The predicted value of soil organic matter content output by the coupled prediction output module; The carbon input flux is calculated by linearly weighting the collected straw return to the field and the soil conditioner application amount using a preset carbon conversion coefficient. Based on the decomposition rate, The temperature and moisture correction function is based on the collected air temperature. With moisture content As independent variables, the product of the Lloyd-Taylor temperature correction term and the normalized moisture correction term is used to obtain the nonlinear mapping. Define pollutant migration-transformation residuals as follows: ； in The predicted pollutant concentration output by the coupled prediction output module; The dispersion coefficient; The longitudinal velocity of soil pore water is obtained by dynamically solving a one-dimensional water infiltration model by combining the collected precipitation and irrigation water volume with the real-time soil moisture content. and For the predicted value of soil organic matter content Nonlinear functions that are controlled; Among them, the blocking factor The degradation rate was calculated using a nonlinear physical mapping method based on the linear isothermal adsorption theory. The calculations were performed using Monod Kinetics nonlinear equations.

7. The closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic pollutant removal as described in claim 1, characterized in that: In step S3, the set of decision variables is established and a dual objective function and hard constraints are set, specifically as follows: First, define the set of decision variables. ,in This refers to the dosage of soil conditioner. For irrigation water volume, For tillage depth, This refers to the amount of straw returned to the field; Secondly, a surrogate model mapping operator is established, and the trained soil carbon cycle-pollutant migration coupling prediction model is used as the surrogate model. By inputting decision variables In addition, the collected time-series data of soil physicochemical indicators, soil pollutant concentrations, and environmental meteorological parameters are directly mapped and output for the future... Predicted values ​​of soil organic matter content at each time step Compared with the predicted pollutant concentration ; Redefine the dual objective function: Objective 1 is set as the predicted value of soil organic matter content and the target steady-state value. Minimize the absolute deviation between them, i.e. Objective two is set as minimizing the predicted pollutant concentration, i.e. ; Finally, hard constraints on the decision variables are set based on actual agronomic standards, namely... , , as well as ,in It is 500 kg per mu. It is 100 cubic meters per mu. It is 30 centimeters. The target is 1000 kg per mu, ensuring that the control measures are within the physical boundaries of safe production.

8. The closed-loop optimization method for soil organic matter carbon cycle regulation and synergistic removal of pollutants as described in claim 1, characterized in that: The sorting method for approximating the ideal solution in S3 is as follows: First, the Pareto front solution set is... The bi-objective prediction results are constructed as a dimension of The initial evaluation matrix is ​​then transformed into a standardized matrix by performing vector normalization. ; Secondly, based on the standardized matrix Extract the numerical extrema of each target column and construct positive ideal solution vectors for each. With negative ideal solution vector The positive ideal solution vector From the standardized matrix The minimum value elements of each objective column in the negative ideal solution vector represent the theoretical optimal boundary in the current solution space. From the standardized matrix The maximum value element in each objective column represents the theoretical worst boundary in the current solution space; We then calculate the Euclidean distances of each decision variable to the positive and negative ideal solutions within the standardized feature space. and Calculate the relative similarity of each scheme. , The Pareto front solution set Relative closeness The set of decision variables with the largest values ​​is taken as the optimal combination of control parameters. .

Images