Plateau lake agricultural non-point source pollution treatment method

Abstract

The invention provides a plateau lake agricultural non-point source pollution treatment method which comprises the following steps: acquiring vegetation indexes and surface temperature field data by using a remote sensing satellite, and generating a pollution source space thermodynamic diagram in combination with water quality and soil parameters of ground sampling points; performing space-time alignment and data fusion on the thermodynamic diagram and real-time runoff and soil permeability data acquired by the hydrological sensor network, and constructing a structured pollution migration database; on the basis of the database, a hybrid neural network model embedded with physical constraints is utilized to predict pollutant concentration distribution within 72 hours in the future; inputting the predicted value into a multi-stage optimization controller, and generating a control parameter set comprising treatment intensity, engineering parameters and a fertilization ratio; generating a treatment strategy map covering the drainage basin through a GIS; and deploying an Internet of Things monitoring node to collect the treated water quality data to form a closed-loop control link. The treatment efficiency and effect can be improved, the treatment cost is reduced, and the negative influence on the ecological environment is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural non-point source pollution control, and more specifically, to a method for controlling agricultural non-point source pollution in plateau lakes. Background Art

[0002] With the rapid development of the social economy, agricultural activities around plateau lakes are becoming increasingly frequent. The use of large amounts of chemical fertilizers and pesticides, as well as the discharge of livestock and poultry manure, have led to increasingly serious agricultural non-point source pollution in the plateau lake basin. Traditional methods for controlling agricultural non-point source pollution mainly rely on single physical, chemical, or biological means, such as the construction of sedimentation / oxidation ponds, wetland purification, and ecological ditches. Although these methods can reduce pollutant emissions to a certain extent, they have problems such as insignificant control effects, high costs, and difficulty adapting to complex and changing watershed environments. In addition, existing control methods often lack accurate identification and dynamic monitoring of pollution sources, making it difficult to achieve real-time regulation and long-term control of pollution.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: it is impossible to accurately identify and locate the source of pollution, it is difficult to achieve dynamic monitoring and real-time regulation of pollution, the treatment effect is not significant and the cost is high, and it is difficult to adapt to the complex and changeable plateau lake basin environment. Summary of the Invention

[0004] The present invention provides a method for controlling agricultural non-point source pollution in plateau lakes, comprising:

[0005] The NDVI vegetation index and surface temperature field data of the target area are obtained through remote sensing satellites. Combined with the water quality and soil parameters collected at ground sampling points, the pollution source intensity distribution is calculated and a spatial heat map of the pollution source with geographic coordinates is generated.

[0006] The spatial heat map of pollution sources is spatiotemporally aligned and data fused with real-time runoff data and soil permeability data collected by the hydrological sensor network to construct a structured pollution migration database containing hydrological dynamic parameters, soil migration parameters, and pollutant diffusion parameters;

[0007] Based on the structured pollution migration database, a hybrid neural network model embedded with physical constraints is used to predict the pollutant concentration distribution at different time and space nodes, and output dynamic concentration prediction values within the next 72 hours;

[0008] The dynamic concentration prediction value is input into a multi-stage optimization controller, which performs parameter sensitivity analysis, spatial correlation constraint optimization, and time rolling optimization in sequence to generate a control parameter set including treatment intensity, engineering parameters, and fertilization ratio;

[0009] Based on the control parameter set, a management strategy map covering the watershed is generated through a GIS geographic information system, the map including the spatial coordinates of the vegetation restoration belt, the three-dimensional design parameters of the intercepting ditch, and the spatiotemporal distribution plan of the fertilization operation;

[0010] The Internet of Things monitoring nodes are deployed to collect water quality data after treatment, and the weight parameters of the hybrid neural network model and the constraints of the multi-stage optimization controller are dynamically adjusted through the feedback evaluation system to form a closed-loop control link.

[0011] Furthermore, the NDVI vegetation index matrix and the surface temperature field matrix with a resolution of 10m×10m are obtained through remote sensing satellites, wherein the NDVI vegetation index matrix represents the normalized vegetation index value of each coordinate point, and the surface temperature field matrix represents the surface temperature measurement value of each coordinate point;

[0012] Sampling points were arranged in a 1km×1km grid within the watershed to collect total nitrogen concentration in water samples and adsorption coefficient in soil. The total nitrogen concentration represents the concentration of water pollutants at each coordinate point, and the adsorption coefficient represents the soil adsorption capacity at each coordinate point.

[0013] The discrete sampling data are converted into a continuous distribution field using a spatial interpolation algorithm, and the pollution source intensity index is calculated based on a weighted combination of vegetation influence coefficient, temperature influence coefficient and pollution coupling coefficient, where the sum of the coefficients is 1;

[0014] The pollution source intensity index is normalized to an intensity value of 0-100, and the geographic coordinate information is superimposed to generate a rasterized heat map.

[0015] Furthermore, the construction of the structured pollution migration database in step 2 includes:

[0016] Calculating hydrological dynamic parameters based on a velocity distribution function of a water-passing section, a water depth measurement, and an integral result of a water-passing section area varying with time;

[0017] Establishing a soil pore blockage model that characterizes porosity changes over time based on an integral relationship between initial porosity, soil type-dependent attenuation coefficients, and pollutant concentrations;

[0018] A pollutant diffusion equation is constructed, which characterizes the diffusion process of pollutants based on the diffusion adjustment factor, the anisotropic conductivity coefficient matrix and the gradient relationship of the pollutant concentration.

[0019] Furthermore, the construction of the hybrid neural network model in step 3 includes:

[0020] A dual-channel input structure was designed. The first channel input contained spatial raster data of pollution source intensity index, hydrological dynamic parameters, and soil porosity, while the second channel input contained time series data of pollutant diffusion coefficient and pollutant concentration.

[0021] A three-dimensional convolutional layer and a gated recurrent unit are alternately deployed in the hidden layer, wherein the three-dimensional convolutional layer extracts features using a Gaussian kernel function defined by a spatial kernel width and a temporal kernel width;

[0022] Mass conservation constraints and concentration boundary constraints are added to the output layer to ensure that the concentration values predicted by the model conform to physical laws.

[0023] Furthermore, the execution process of the multi-stage optimization controller in step 4 includes:

[0024] Parameter pre-optimization stage: By calculating the sensitivity matrix of the control parameters to the predicted concentration values, the parameters with sensitivity higher than the threshold are screened and entered into the optimization sequence;

[0025] Spatial correlation optimization stage: construct a regional influence weight matrix based on regional centroid distance, concentration deviation and coupling influence coefficient to quantify the mutual correlation between regions;

[0026] Time rolling optimization stage: A variable step-size predictive control algorithm is used to construct the objective function with concentration tracking weights and control cost weights to optimize the control sequence within the future rolling time domain window.

[0027] Furthermore, the parameter pre-optimization stage also includes:

[0028] Establishing a dynamic feasible domain constraint condition, wherein the condition dynamically adjusts a feasible range of a control parameter based on a lower limit scaling factor, an upper limit scaling factor, and an adaptive adjustment rate;

[0029] A chaotic particle swarm optimization algorithm is used for parameter search, which updates the particle position according to the particle velocity vector and the chaotic perturbation intensity.

[0030] Furthermore, the spatial association optimization stage further includes:

[0031] Design a multi-objective optimization function that includes economic cost coefficient, ecological protection coefficient, and implementation difficulty penalty coefficient;

[0032] The distributed ADMM algorithm is used to solve the optimization problem, and the regional correlation constraint matrix is processed by the augmented Lagrangian function.

[0033] Furthermore, the time rolling optimization stage also includes:

[0034] constructing a time-varying state-space model that describes system dynamics based on an attenuation coefficient matrix, a control gain matrix, and an environmental disturbance term;

[0035] Design a predictive controller to generate the optimal control sequence by minimizing the objective function in the prediction horizon;

[0036] A feedback correction mechanism is implemented to trigger parameter reoptimization when the monitored concentration deviation exceeds the threshold.

[0037] Furthermore, the generation of the governance strategy map in step 5 includes:

[0038] The planting density of the vegetation restoration belt is calculated based on the plant purification efficiency coefficient, grid unit area, absorption per unit area and growth cycle;

[0039] Determine the depth and slope of the intercepting ditch based on peak flow, design velocity, channel bottom width, roughness coefficient, and upstream and downstream elevations;

[0040] Generate a fertilization control plan based on fertilizer conversion rate, fertilizer utilization rate and control time interval.

[0041] Furthermore, the implementation of the closed-loop control link in step 6 includes:

[0042] Construct a three-dimensional evaluation index including spatial weight, temporal accumulation weight and control cost weight;

[0043] Design parameter adaptive update law based on learning rate, Hessian matrix and parameter gradient operator;

[0044] When the residual between the model-predicted concentration value and the actual concentration exceeds a threshold, the model reconstruction process is initiated to retrain the hybrid neural network model.

[0045] The above-described embodiments of the present invention have at least the following beneficial effects: the method can achieve accurate identification and dynamic monitoring of agricultural non-point source pollution in plateau lakes. Through multi-source data fusion and hybrid neural network model prediction, it can accurately locate pollution sources and predict the concentration distribution of pollutants, providing a scientific basis for pollution control. At the same time, the method can also dynamically adjust the control strategy based on real-time monitoring data, generate an optimal set of control parameters through a multi-stage optimization controller, achieve real-time regulation and long-term control of pollution, effectively improve control efficiency, reduce control costs, and reduce negative impacts on the ecological environment.

[0046] Furthermore, this method can generate a comprehensive, systematic solution for agricultural non-point source pollution control, including the spatial coordinates of vegetation restoration zones, the three-dimensional design parameters of interception ditches, and the temporal and spatial distribution of fertilization operations. This approach also generates a comprehensive, systematic solution for agricultural non-point source pollution control. Through the design of a closed-loop control link, timely feedback on control results can be provided, further optimizing control measures and ensuring the continued effectiveness of control efforts, resulting in significant ecological and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0048] Figure 1 A schematic flow chart of a method for controlling agricultural non-point source pollution in plateau lakes provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0049] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0050] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0051] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0052] Reference below Figure 1 , Figure 1 This is a flow chart of a method for treating agricultural non-point source pollution in plateau lakes according to one embodiment of the present invention. Figure 1 As shown, a method for controlling agricultural non-point source pollution in plateau lakes includes:

[0053] Step 1: Obtain NDVI vegetation index and surface temperature field data of the target area through remote sensing satellites. Combined with water quality and soil parameters collected at ground sampling points, calculate the intensity distribution of pollution sources and generate a spatial heat map of pollution sources with geographic coordinates.

[0054] Step 2: performing spatiotemporal alignment and data fusion on the pollution source spatial heat map with the real-time runoff data and soil permeability data collected by the hydrological sensor network to construct a structured pollution migration database containing hydrological dynamic parameters, soil migration parameters, and pollutant diffusion parameters;

[0055] Step 3: Based on the structured pollution migration database, a hybrid neural network model embedded with physical constraints is used to predict the pollutant concentration distribution at different spatiotemporal nodes, and output dynamic concentration prediction values within the next 72 hours;

[0056] Step 4: Input the dynamic concentration prediction value into a multi-stage optimization controller, and perform parameter sensitivity analysis, spatial correlation constraint optimization, and time rolling optimization in sequence to generate a control parameter set including treatment intensity, engineering parameters, and fertilization ratio;

[0057] Step 5: Based on the control parameter set, a management strategy map covering the watershed is generated through a GIS geographic information system. The map includes the spatial coordinates of the vegetation restoration zone, the three-dimensional design parameters of the intercepting ditch, and the spatiotemporal distribution plan of the fertilization operation.

[0058] Step 6: deploy IoT monitoring nodes to collect water quality data after treatment, and dynamically adjust the weight parameters of the hybrid neural network model and the constraints of the multi-stage optimization controller through the feedback evaluation system to form a closed-loop control link.

[0059] It should be noted that the first step of this method is to obtain the vegetation index (NDVI) and surface temperature field data of the target area through remote sensing satellites, and combine it with the water quality and soil parameters collected at ground sampling points to calculate the intensity distribution of pollution sources and generate a spatial heat map of pollution sources with geographic coordinates. The vegetation index (NDVI) mentioned here is an indicator obtained through satellite remote sensing technology to reflect the growth status and coverage of vegetation, and the surface temperature field data can reflect the thermal distribution of the region. The water quality parameters collected at the ground sampling points are mainly the total nitrogen concentration in the water samples, and the soil parameters are mainly the adsorption coefficient of the soil. These parameters can help to more accurately assess the intensity and distribution of pollution sources.

[0060] Specifically, the NDVI vegetation index matrix and the surface temperature field matrix obtained by remote sensing satellites have a resolution of 10m×10m, which means that relatively detailed vegetation and surface temperature information can be obtained. Sampling points are arranged in a 1km×1km grid within the watershed to collect the total nitrogen concentration in water samples and the adsorption coefficient in the soil. The discrete sampling data is converted into a continuous distribution field through a spatial interpolation algorithm, and then the pollution source intensity index is calculated. This index is calculated by comprehensively considering the vegetation influence coefficient, temperature influence coefficient, and pollution coupling coefficient, where the sum of these three coefficients is 1. Finally, the pollution source intensity index is normalized to an intensity value of 0-100, and the geographic coordinate information is superimposed to generate a rasterized heat map. This heat map can intuitively display the spatial distribution of pollution sources, providing an important basis for subsequent pollution control.

[0061] Preferably, in order to improve the accuracy of the spatial heat map of pollution sources, higher-resolution remote sensing satellite data, such as a resolution of 5m×5m, can be used when acquiring data to obtain more detailed vegetation and surface temperature information. At the same time, the density of ground sampling points can be increased, such as reducing the sampling grid to 500m×500m, to obtain richer water quality and soil parameter data. In addition, more environmental parameters, such as soil moisture and wind speed, can be introduced to further improve the calculation model of the pollution source intensity index. For example, soil moisture can reflect the water holding capacity of the soil, and wind speed will affect the diffusion rate of pollutants. Through these improvements, the distribution of pollution sources can be assessed more comprehensively and accurately, providing more scientific decision-making support for pollution control.

[0062] In some embodiments, the generation of the pollution source spatial heat map in step 1 includes:

[0063] The NDVI vegetation index matrix T(x,y) and the surface temperature field matrix H(x,y) with a resolution of 10m×10m are obtained through remote sensing satellites, where T(x,y) represents the normalized vegetation index value at the coordinates (x,y), and H(x,y) represents the surface temperature measurement value at the coordinates (x,y);

[0064] Sampling points were arranged in a 1km×1km grid in the watershed to collect the total nitrogen concentration C in the water samples. w (x,y) and the adsorption coefficient K in the soil s (x,y), where C w (x,y) represents the concentration of water pollutants at the coordinates (x,y), K s (x,y) represents the soil adsorption coefficient at coordinates (x,y);

[0065] The spatial interpolation algorithm is used to convert discrete sampling data into a continuous distribution field and calculate the pollution source intensity index.

[0066] PI(x,y)=α·T(x,y)+β·H(x,y)+γ·ln[C w (x,y)·K s (x,y)]

[0067] Among them, α is the vegetation influence coefficient, β is the temperature influence coefficient, γ is the pollution coupling coefficient, α+β+γ=1;

[0068] Normalize PI(x,y) to an intensity value of 0-100 and overlay geographic coordinate information to generate a rasterized heat map.

[0069] It's important to note that the core of building a structured pollution migration database lies in integrating multiple data sources to form a comprehensive database encompassing hydrological dynamic parameters, soil transport parameters, and pollutant diffusion parameters. Hydrological dynamic parameters primarily describe water flow characteristics such as velocity and flow rate, soil transport parameters address the migration of pollutants within the soil, and pollutant diffusion parameters describe the spatial diffusion of pollutants. By spatially and temporally aligning these parameters with spatial heat maps of pollution sources and integrating these data, we can more comprehensively understand and predict pollution migration processes, providing data support for subsequent pollution control efforts.

[0070] Specifically, the process of building a structured pollution migration database includes the calculation of multiple key parameters. For example, the calculation of hydrological dynamic parameters needs to consider the velocity distribution function of the water section, the water depth measurement value, and the water section area that changes with time. These parameters are obtained through integral calculation and can reflect the dynamic changes of water flow. The soil pore blockage model is used to describe the changes in soil pores over time, among which the initial porosity, soil type-related attenuation coefficient and pollutant concentration are key factors. The pollutant diffusion equation describes the diffusion process of pollutants in the soil by considering the diffusion adjustment factor, the anisotropic conductivity coefficient matrix and the gradient operator. The setting and calculation of these parameters need to be adjusted according to the specific geographical environment and pollution situation to ensure the accuracy and reliability of the model.

[0071] Preferably, in order to improve the accuracy and practicality of the structured pollution migration database, more sophisticated calculation methods and data collection means can be adopted. For example, when calculating hydrological dynamic parameters, higher-precision flow sensors and water depth measurement equipment can be used to obtain more accurate flow velocity distribution and water depth data. In the soil pore blockage model, the attenuation coefficient can be adjusted according to different soil types to more accurately reflect the changes in soil pores. In addition, more environmental variables, such as rainfall and evaporation, can be introduced to further improve the pollutant diffusion equation. For example, rainfall will affect the permeability of the soil and the migration speed of pollutants, while evaporation will affect the amount and quality of surface water. Through these improvements, the pollution migration process can be described more comprehensively, providing a more scientific basis for pollution control.

[0072] In some embodiments, the construction of the structured pollution migration database in step 2 includes:

[0073] Calculate the hydrological dynamic parameter Q(t) = ∫ A(t) v(s,t)·h(s,t)ds, where v(s,t) is the velocity distribution function at coordinate s on the water-passing section A(t) at time t, h(s,t) is the water depth measurement at coordinate s at time t, A(t) is the water-passing section area that changes with time, and Q(t) is the real-time runoff at time t.

[0074] Establishing a soil pore blockage model where η0 is the initial porosity, k is the soil type-dependent attenuation coefficient, C(s) is the pollutant concentration at time s, and η(t) is the soil porosity at time t.

[0075] Constructing the pollutant diffusion equation Where μ is the diffusion adjustment factor, M(x,y) is the anisotropic conductivity matrix, is a two-dimensional gradient operator, D(x,y,t) is the pollutant diffusion flux at the coordinate (x,y) at time t, and C(x,y,t) is the pollutant concentration at the coordinate (x,y) at time t.

[0076] It should be noted that a hybrid neural network model embedded with physical constraints predicts pollutant concentration distributions at different spatiotemporal nodes and outputs dynamic concentration forecasts for the next 72 hours. This advanced method combines traditional neural networks with physical models. Using a dual-channel input structure, which inputs spatial raster data and time series data, it can more accurately predict pollutant concentration changes while taking into account physical laws. The addition of physical constraints ensures the physical plausibility of the model's predictions, further improving the accuracy and reliability of the predictions.

[0077] Specifically, the construction of the hybrid neural network model includes designing a dual-channel input structure. The first channel inputs spatial raster data, including spatial information such as vegetation index, surface temperature, and soil adsorption coefficient; the second channel inputs time series data, such as hydrological dynamic parameters and soil migration parameters, which change over time. In the hidden layer, three-dimensional convolutional layers and gated recurrent units are deployed alternately. The three-dimensional convolutional layer is used to extract spatial features, and the gated recurrent unit is used to process time series data. The convolution kernel function is a Gaussian function used to define the weight distribution of the convolution operation, where the spatial kernel width and temporal kernel width are parameters that control the shape of the convolution kernel. In the output layer, physical constraints are added. These constraints include mass conservation and concentration boundary constraints to ensure that the model prediction results conform to physical laws.

[0078] Preferably, in order to further improve the prediction accuracy of the hybrid neural network model, more data types and more complex physical constraints can be introduced during the model training process. For example, meteorological data (such as wind speed, wind direction, rainfall, etc.) can be added as input data because these factors have a significant impact on the diffusion of pollutants. In the physical constraints, more detailed fluid mechanics and chemical reaction models can be introduced to more accurately describe the transmission and transformation process of pollutants. In addition, more advanced optimization algorithms, such as the stochastic gradient descent algorithm with adaptive learning rate, can be used to optimize the training process of the model. The hyperparameters of the model, such as the size and number of convolution kernels, the number of layers of gated recurrent units, etc., can also be adjusted through methods such as cross-validation to improve the generalization ability and prediction accuracy of the model.

[0079] In some embodiments, the construction of the hybrid neural network model in step 3 includes:

[0080] Design a dual-channel input structure, the first channel inputs spatial grid data [PI(x,y),Q(t),η(t)], and the second channel inputs time series data [D(x,y,t),C(x,y,t)];

[0081] Three-dimensional convolutional layers and gated recurrent units are deployed alternately in the hidden layer, and the convolution kernel function is

[0082] Among them, σ s is the spatial kernel width, σ t is the width of the time kernel, x and y are the coordinates of the convolution kernel in space,

[0083] Add physical constraints to the output layer

[0084]

[0085] Among them, λ1 is the mass conservation constraint strength, λ2 is the concentration boundary constraint strength, C pred is the concentration value predicted by the model, C min is the minimum permissible value of pollutant concentration, C max is the maximum permissible value of pollutant concentration, is the diffusion term, is the Laplace operator, It is the convection term, which indicates the convection transport of pollutants under the action of flow velocity v, and reflects the movement of pollutants with media such as water flow.

[0086] It should be noted that the multi-stage optimization controller is an optimization tool that sequentially performs parameter sensitivity analysis, spatial correlation constraint optimization, and time-rolling optimization. Parameter sensitivity analysis is used to screen key parameters that have a significant impact on pollution control effectiveness. Spatial correlation constraint optimization considers the mutual influence between different regions. And time-rolling optimization ensures that the pollution control strategy can be dynamically adjusted based on real-time data, thereby generating an optimal set of control parameters and providing a scientific basis for pollution control.

[0087] Specifically, the parameter pre-optimization stage screens sensitive parameters by calculating the sensitivity matrix of the control parameters. The calculation of the sensitivity matrix is based on the derivative of the control parameters on the governance effect, and the initial parameter value is used to determine the benchmark point for the sensitivity calculation. The screened sensitive parameters will enter the optimization sequence. The spatial correlation optimization stage constructs a regional impact weight matrix, which takes into account the centroid distance, concentration deviation and coupling influence coefficient between regions, and is used to quantify the mutual influence between different regions. The time rolling optimization stage adopts a variable step size predictive control algorithm. By setting parameters such as the rolling time domain window length, concentration tracking weight and control cost weight, the control sequence is optimized to achieve the best governance effect.

[0088] Preferably, in order to improve the performance of the multi-stage optimization controller, dynamic feasible domain constraints can be introduced in the parameter pre-optimization stage. These constraints are based on the lower limit scaling coefficient, the upper limit scaling coefficient and the adaptive adjustment rate, and can dynamically adjust the feasible range of the parameters, thereby improving the flexibility and adaptability of the optimization. In the spatial correlation optimization stage, a multi-objective optimization function can be designed, taking into account factors such as economic cost, ecological protection and implementation difficulty, and solving the optimization problem through a distributed ADMM algorithm to further improve the efficiency and effect of the optimization. In the time rolling optimization stage, a time-varying state space model can be constructed and a predictive controller can be designed. The feedback correction mechanism can immediately trigger parameter re-optimization when the concentration deviation exceeds the standard, ensuring the real-time and effectiveness of the governance strategy.

[0089] In some embodiments, the execution process of the multi-stage optimization controller in step 4 includes:

[0090] Parameter pre-optimization stage: Calculate the control parameter sensitivity matrix

[0091] Among them, u j is the jth control parameter, u0 is the initial parameter value, and the parameters with sensitivity higher than the threshold δ are selected to enter the optimization sequence;

[0092] Spatial association optimization stage: constructing regional influence weight matrix

[0093] W ij =exp[-d ij 2 / (2σ2 )+ρ·ΔC i ΔC j ], where d ij is the distance between the centroids of regions i and j, σ is the spatial correlation radius, ΔC i is the concentration deviation of region i, ρ is the coupling influence coefficient, ΔC j is the concentration deviation of region j.

[0094] Time rolling optimization stage: using variable step size predictive control algorithm, the objective function

[0095] Among them, N is the length of the rolling time domain window, w1 is the concentration tracking weight, w2 is the control cost weight, C pred (k) is the predicted pollutant concentration value at the kth time step in the rolling time window, C target is the target pollutant concentration, and Δu(k) is the change in the control parameter at the kth time step in the rolling time domain window.

[0096] It should be noted that a watershed-wide remediation strategy map was generated using a geographic information system (GIS). This map is a visualization tool used to showcase specific remediation measures and their layout. It includes the spatial coordinates of vegetation restoration zones, the three-dimensional design parameters of intercepting ditches, and the spatial and temporal distribution of fertilization operations. This information can provide detailed guidance for actual pollution control projects, ensuring the scientific and operational nature of remediation measures.

[0097] Specifically, the planting density of the vegetation restoration belt is calculated based on the plant purification efficiency coefficient, grid unit area, absorption per unit area and growth cycle. This parameter reflects the appropriate planting density of the vegetation restoration belt in different areas to achieve the best purification effect. The three-dimensional parameters of the interception ditch include depth, slope, etc. These parameters are calculated based on the peak flow rate at the location, design flow rate, channel bottom width, roughness coefficient, and upstream and downstream elevations, and are used to design ditches that can effectively intercept and transport pollutants. The fertilization control plan is formulated based on factors such as fertilizer efficiency conversion rate, fertilizer utilization rate and control time interval to achieve precise fertilization and reduce the pollution of water bodies by fertilizer loss.

[0098] Preferably, in order to further optimize the generation of the governance strategy map, more ecological parameters can be introduced into the design of vegetation restoration belts, such as the pollution tolerance and ecological adaptability of plants, so as to select plant species that are more suitable for the local environment. In the design of interception ditches, it is possible to consider introducing ecological slope protection technology, which can not only intercept pollutants but also provide ecological habitats. For fertilization control plans, the amount and time of fertilizer application can be dynamically adjusted in combination with soil fertility monitoring data to adapt to the needs of different seasons and soil conditions. In addition, the governance strategy map can be pre-evaluated and optimized by simulating the implementation effects of different governance measures to ensure its effectiveness in actual applications.

[0099] In some embodiments, the parameter pre-optimization stage further includes:

[0100] Establish dynamic feasible region constraints

[0101]

[0102] Among them, μ1 is the lower limit scaling coefficient, μ2 is the upper limit scaling coefficient, κ is the adaptive adjustment rate, ΔC(τ) is the concentration deviation at time τ, is the lower limit of the jth control parameter at time t, is the upper limit of the jth control parameter at time t, is the baseline value of the jth control parameter.

[0103] The chaotic particle swarm algorithm is used for parameter search, and the position update formula is:

[0104] Among them, v i is the particle velocity vector, η is the chaotic disturbance intensity, is the position of the i-th particle at the k+1-th iteration, is the position of the i-th particle at the k-th iteration, is the velocity vector of the i-th particle at the k+1-th iteration.

[0105] It should be noted that IoT monitoring nodes are deployed to collect post-treatment water quality data. A feedback evaluation system dynamically adjusts the weight parameters of the hybrid neural network model and the constraints of the multi-stage optimization controller, forming a closed-loop control chain. IoT monitoring nodes are devices capable of real-time data collection and transmission. They are deployed in the treatment area to monitor water quality changes. The feedback evaluation system evaluates the treatment effectiveness based on the collected data and adjusts the model and controller parameters accordingly, achieving dynamic optimization and closed-loop control of the treatment process.

[0106] Specifically, the deployment of IoT monitoring nodes needs to consider the distribution density and location of monitoring points to ensure comprehensive coverage of the governance area and obtain representative data. The water quality data collected by the monitoring nodes includes, but is not limited to, key indicators such as pollutant concentration, pH value, and dissolved oxygen. The feedback evaluation system constructs evaluation indicators based on this data, for example, by setting parameters such as spatial weights, time accumulation weights, and control cost weights to quantify the governance effect. When the model residual is detected to exceed the set threshold, the system will initiate the model reconstruction process, retraining the convolution kernel parameters and LSTM unit weights of the hybrid neural network model to improve the model's prediction accuracy and adaptability.

[0107] To improve the efficiency and accuracy of closed-loop control links, intelligent sensor networks can be introduced into the deployment of IoT monitoring nodes. These sensors can automatically calibrate and adapt to environmental changes, thereby improving the reliability of data collection. In the feedback evaluation system, more advanced machine learning algorithms, such as reinforcement learning, can be used to dynamically adjust model parameters and controller constraints. In addition, multiple threshold levels can be set, and different degrees of adjustment measures can be taken according to the degree of deviation, such as fine-tuning for minor deviations and comprehensive re-optimization for severe deviations. Expert systems can also be introduced to combine artificial intelligence algorithms with the experience and knowledge of human experts to further optimize the adjustment process of governance policies.

[0108] In some embodiments, the spatial association optimization stage further includes:

[0109] Designing multi-objective optimization functions

[0110]

[0111] Among them, α is the economic cost coefficient, β is the ecological protection coefficient, γ is the implementation difficulty penalty coefficient, and W ij is the weight value between region i and region j in the regional influence weight matrix, u j is the jth control parameter, is the maximum concentration of pollutants allowed in area i, is the predicted pollutant concentration value in area i, is the minimum value of the jth control parameter.

[0112] Distributed ADMM algorithm is used to solve the optimization problem and augment the Lagrangian function

[0113] Among them, ρ is the penalty parameter, λ is the dual variable, A is the regional association constraint matrix, L ρ is the value of the augmented Lagrangian function, F i (u i) is the multi-objective optimization function value of region i, considering the influence of the control parameter ui in region i on the multi-objectives, b is the constraint condition vector, which defines the specific requirements and restrictions of the regional association constraints, and u is the control parameter vector.

[0114] It should be noted that this method uses dynamic feasible region constraints and a chaotic particle swarm algorithm for parameter search during the parameter pre-optimization phase. Dynamic feasible region constraints are a strategy that dynamically adjusts parameter ranges based on real-time data, ensuring that parameters remain within a reasonable range during the optimization process. The chaotic particle swarm algorithm, an optimization algorithm based on swarm intelligence, searches for the optimal solution by simulating the movement of particles in a search space. This method combines the randomness of chaos theory with the global search capabilities of the particle swarm algorithm, effectively improving the efficiency and accuracy of parameter search.

[0115] Specifically, the dynamic feasible domain constraint dynamically adjusts the feasible range of the parameter by setting the lower limit scaling factor, the upper limit scaling factor, and the adaptive adjustment rate. The lower limit scaling factor and the upper limit scaling factor are used to determine the minimum and maximum values of the parameter, and the adaptive adjustment rate dynamically adjusts the parameter range according to the real-time concentration deviation. In the chaotic particle swarm algorithm, the particle velocity and position update formula introduces the chaotic perturbation intensity. This perturbation can increase the randomness of the particles and prevent the algorithm from falling into the local optimal solution. In the particle position update formula, the chaotic perturbation intensity and the particle velocity vector jointly determine the particle's motion trajectory in the search space.

[0116] To further enhance parameter optimization, additional environmental parameters, such as temperature and humidity, can be incorporated into the dynamic feasible region constraints. These parameters can influence the migration and transformation of pollutants, thereby impacting the appropriate range of parameters. Within the chaotic particle swarm optimization algorithm, various chaotic mapping methods, such as logistic mapping or tent mapping, can be incorporated to provide more random selection and enhance the algorithm's global search capabilities. Furthermore, other optimization algorithms, such as genetic algorithms or simulated annealing, can be combined to form a hybrid optimization strategy, further improving the efficiency and accuracy of parameter search.

[0117] In some embodiments, the time rolling optimization phase further includes:

[0118] Constructing a time-varying state-space model

[0119] dx / dt=A(t)x+B(t)u+ξ(t),

[0120] Where A(t)=diag[a1(t),...,a n (t)] is the attenuation coefficient matrix, a i(t) is the natural decay characteristics of different state variables in the system over time, and x is the system state vector;

[0121] B(t)=[b ij (t)] is the control gain matrix, b ij (t) is the gain of the j-th control input to the i-th state variable;

[0122] ξ(t) is the environmental disturbance term;

[0123] Designing a predictive controller

[0124] Among them, T is the prediction time domain length, E(·) is the terminal cost function, argmin u(t) In order to find the value of u(t) that minimizes the subsequent expression within the range of the control input u(t), J(x(τ), u(τ)) is the objective function used to measure the performance index under the combination of time τ, system state x(τ) and control input u(τ).

[0125] Implement feedback correction mechanism: when the concentration deviation ΔC(t)≥ΔC th , the parameter re-optimization is triggered immediately and the control sequence in the future time domain is recalculated.

[0126] It should be noted that this method uses a multi-objective optimization function and a distributed ADMM algorithm to solve the optimization problem during the spatial correlation optimization phase. A multi-objective optimization function is an optimization method that comprehensively considers multiple objectives and can balance the relationships between different objectives, such as economic costs, ecological protection, and implementation difficulty. The distributed ADMM algorithm is used to solve distributed optimization problems. By introducing augmented Lagrangian functions and dual variables, it can effectively handle multi-region correlation constraints and achieve regional collaborative governance.

[0127] Specifically, the multi-objective optimization function balances the relationships between different objectives by setting an economic cost coefficient, an ecological protection coefficient, and an implementation difficulty penalty coefficient. The economic cost coefficient is used to measure the economic feasibility of the governance measures, the ecological protection coefficient is used to assess the impact of the governance measures on the ecological environment, and the implementation difficulty penalty coefficient is used to account for the difficulty of implementing the governance measures. The distributed ADMM algorithm solves the optimization problem by constructing an augmented Lagrangian function, introducing a penalty parameter and dual variables, and a regional correlation constraint matrix. The penalty parameter is used to adjust the strictness of the constraints, the dual variables are used to process the constraints, and the regional correlation constraint matrix is used to describe the relationships between different regions.

[0128] Preferably, in order to further improve the effect of spatial correlation optimization, more objective coefficients, such as social impact coefficients or sustainable development coefficients, can be introduced into the multi-objective optimization function to more comprehensively evaluate the comprehensive benefits of governance measures. In the distributed ADMM algorithm, more advanced optimization strategies can be adopted, such as dynamically adjusting penalty parameters or introducing an adaptive dual variable update mechanism to improve the convergence speed and stability of the algorithm. In addition, geographic information system (GIS) technology can be combined to dynamically adjust the regional correlation constraint matrix according to the geographical characteristics and ecological conditions of the region to more accurately reflect the mutual influence between different regions.

[0129] In some embodiments, the generation of the governance strategy map in step 5 includes:

[0130] Calculate the planting density N of the vegetation restoration belt p (x,y)=θ·C pred (x,y)·A(x,y) / (k p ·T g ), where θ is the plant purification efficiency coefficient, A(x,y) is the grid unit area, and k p is the absorption per unit area, T g For the growth cycle;

[0131] Determine the three-dimensional parameters of the interception ditch: depth h d (x) = Q max (x) / (v c W d (x)·n), slope S(x)=arctan[(H u (x)-H d (x)) / L(x)], where Q max (x) is the peak flow rate at position x, v c is the design flow rate, W d is the channel bottom width, n is the roughness coefficient, H u 、H d are the upstream and downstream elevations respectively;

[0132] Generate fertilization control plan:

[0133] ΔF(x,y,t)=ρ f ·[C target (x,y)-C pred (x,y,t)]·A(x,y) / (k f ·Δt), where ρ f is the fertilizer conversion rate, k f is the fertilizer utilization rate, and Δt is the control time interval.

[0134] It should be noted that this method constructs a time-varying state-space model and designs a predictive controller in the time-rolling optimization phase, while also implementing a feedback correction mechanism. The time-varying state-space model is a dynamic system model used to describe the changes in system state over time, and can take into account the dynamic characteristics within the system and the interference of the external environment. Based on this model, the predictive controller predicts the system state for a period of time in the future and calculates the optimal control input. The feedback correction mechanism is used to monitor the deviation between the system output and the expected value in real time. When the deviation exceeds the set threshold, it triggers the re-optimization of the control parameters to ensure the stability and control accuracy of the system.

[0135] Specifically, the time-varying state-space model describes the dynamic behavior of the system by defining an attenuation coefficient matrix, a control gain matrix, and an environmental disturbance term. The attenuation coefficient matrix reflects the natural attenuation characteristics of the system state over time, the control gain matrix represents the impact of the control input on the system state, and the environmental disturbance term takes into account the interference of external environmental factors on the system. The design of the predictive controller is based on the prediction time domain length and the terminal cost function, and the optimal control sequence is calculated by minimizing the cost function. The feedback correction mechanism determines whether the control parameters need to be re-optimized by setting a concentration deviation threshold. When the monitored concentration deviation exceeds the threshold, the parameter re-optimization process is immediately initiated.

[0136] Preferably, in order to further improve the performance of time rolling optimization, more dynamic characteristics can be introduced into the time-varying state space model, such as considering nonlinear changes in the system state or time-varying external interference. In the design of the predictive controller, more advanced control strategies can be adopted, such as multi-step prediction and rolling optimization based on model predictive control (MPC) to improve the flexibility and adaptability of the control. In addition, an adaptive control mechanism can be introduced to dynamically adjust the prediction model and control strategy according to real-time monitoring data to better cope with changes in the dynamic characteristics of the system. In the feedback correction mechanism, multi-level thresholds can be set, and different degrees of correction measures can be taken according to the severity of the deviation, such as fine-tuning for slight deviations and comprehensive re-optimization for serious deviations.

[0137] In some embodiments, the implementation of the closed-loop control link in step 6 includes:

[0138] Constructing three-dimensional evaluation indicators Among them, ω s is the spatial weight, ω t is the time cumulative weight, ω p To control the cost weight, τ is the time window length;

[0139] Design parameter adaptive update law: Among them, η is the learning rate, M is the Hessian matrix, is the parameter gradient operator;

[0140] When the model residual R(t) is detected to be real (t)-C pred (t)||≥R th , the model reconstruction process is started to retrain the convolution kernel parameters and LSTM unit weights of the hybrid neural network model.

[0141] It should be noted that this method constructs a three-dimensional evaluation index in the implementation of the closed-loop control link, designs a parameter adaptive update law, and sets up a model reconstruction process. The three-dimensional evaluation index is a comprehensive evaluation method used to quantify multiple aspects of governance effectiveness, including spatial distribution, time accumulation, and control cost. The parameter adaptive update law is a mechanism for dynamically adjusting model parameters. It can adjust model weights based on real-time data to improve the model's adaptability and prediction accuracy. The model reconstruction process is used to retrain the model when the model residual exceeds the set threshold to ensure the accuracy and reliability of the model.

[0142] Specifically, the three-dimensional evaluation index comprehensively assesses governance effectiveness by setting spatial weights, temporal accumulation weights, and control cost weights. Spatial weights are used to measure the importance of governance effects in different regions, temporal accumulation weights are used to assess the temporal sustainability of governance effects, and control cost weights are used to consider the economic costs of governance measures. The parameter adaptive update law adjusts model parameters by setting the learning rate and Hessian matrix. The learning rate determines the step size of the parameter update, and the Hessian matrix is used to calculate the second-order derivatives of the parameters, thereby optimizing the direction of parameter updates. When the model residual is detected to exceed the set threshold, the model reconstruction process is initiated to retrain the convolution kernel parameters and LSTM unit weights of the hybrid neural network model to improve the model's prediction accuracy.

[0143] Preferably, in order to further improve the performance of the closed-loop control link, more evaluation dimensions can be introduced into the three-dimensional evaluation indicators, such as ecological restoration indicators or social impact indicators, to more comprehensively evaluate the governance effect. In the parameter adaptive update law, more advanced optimization algorithms, such as adaptive moment estimation (Adam) or variants of stochastic gradient descent (SGD), can be used to improve the efficiency and stability of parameter updates. In addition, an incremental learning mechanism can be introduced into the model reconstruction process so that the model can retain some historical information when retraining, thereby improving the generalization ability and adaptability of the model.

[0144] The above-mentioned embodiments of the present invention have the following beneficial effects: The present invention can accurately locate the distribution of pollution sources by generating a spatial heat map of pollution sources, providing an intuitive basis for subsequent governance. Constructing a structured pollution migration database can integrate multi-source data to achieve a comprehensive grasp of the pollution migration process. Using a hybrid neural network model to predict the distribution of pollutant concentrations can predict pollution trends in advance and provide forward-looking guidance for the formulation of governance measures. The use of a multi-stage optimization controller can generate a scientific and reasonable set of control parameters, optimize the allocation of governance resources, and improve governance effects. With the help of GIS to generate a governance strategy map, the visual layout of governance measures can be realized, which is convenient for implementation and management. Deploying IoT monitoring nodes and forming a closed-loop control link can provide real-time feedback on governance effects, dynamically adjust governance strategies, and ensure the continued effectiveness of governance work.

[0145] In addition, this method optimizes parameters through dynamic feasible domain constraints and chaotic particle swarm algorithm, which can improve the efficiency and accuracy of parameter search and further enhance the governance effect. The distributed ADMM algorithm is used to solve the optimization problem, which can effectively handle multi-region correlation constraints and realize regional collaborative governance. Constructing a time-varying state space model and designing a predictive controller can enhance the adaptability to dynamic changes and improve control accuracy. Implementing a feedback correction mechanism can timely adjust the governance measures when the concentration deviation exceeds the standard to ensure the achievement of the governance goals. The comprehensive application of these technical means can effectively improve the scientificity, accuracy and efficiency of the governance of agricultural non-point source pollution in plateau lakes, and provide strong support for protecting the ecological environment of plateau lakes.

[0146] At the same time, this method also enables refined management of pollution sources. By calculating the planting density of vegetation restoration belts, the three-dimensional parameters of interception ditches, and fertilization control plans, it can provide detailed design parameters for specific remediation projects, ensuring the scientific and effective nature of each remediation measure. This method not only effectively reduces pollutant emissions but also improves the sustainability of agricultural production, achieving a win-win situation for ecological protection and agricultural development. By implementing a closed-loop control link, model parameters can be further optimized, improving the model's prediction accuracy and adaptability, providing more reliable decision-making support for the control of agricultural non-point source pollution.

[0147] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, a server, a mobile phone, or a tablet.

[0148] The above descriptions are merely some preferred embodiments of the present invention and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also encompass other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by mutually replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. A method for controlling agricultural non-point source pollution in plateau lakes, characterized in that: The following steps are involved: Step 1: Obtain NDVI vegetation index and surface temperature field data of the target area through remote sensing satellites. Combined with water quality and soil parameters collected at ground sampling points, calculate the intensity distribution of pollution sources and generate a spatial heat map of pollution sources with geographic coordinates. Step 2: performing spatiotemporal alignment and data fusion on the pollution source spatial heat map with the real-time runoff data and soil permeability data collected by the hydrological sensor network to construct a structured pollution migration database containing hydrological dynamic parameters, soil migration parameters, and pollutant diffusion parameters; Step 3: Based on the structured pollution migration database, a hybrid neural network model embedded with physical constraints is used to predict the pollutant concentration distribution at different spatiotemporal nodes, and output dynamic concentration prediction values within the next 72 hours; Step 4: Input the dynamic concentration prediction value into a multi-stage optimization controller, and perform parameter sensitivity analysis, spatial correlation constraint optimization, and time rolling optimization in sequence to generate a control parameter set including treatment intensity, engineering parameters, and fertilization ratio; Step 5: Based on the control parameter set, a management strategy map covering the watershed is generated through a GIS geographic information system. The map includes the spatial coordinates of the vegetation restoration zone, the three-dimensional design parameters of the intercepting ditch, and the spatiotemporal distribution plan of the fertilization operation. Step 6: deploy IoT monitoring nodes to collect water quality data after treatment, and dynamically adjust the weight parameters of the hybrid neural network model and the constraints of the multi-stage optimization controller through the feedback evaluation system to form a closed-loop control link.

2. The method according to claim 1, wherein The generation of the pollution source spatial heat map in step 1 includes: The NDVI vegetation index matrix and the surface temperature field matrix with a resolution of 10m×10m are obtained through remote sensing satellites. The NDVI vegetation index matrix represents the normalized vegetation index value of each coordinate point, and the surface temperature field matrix represents the surface temperature measurement value of each coordinate point. Sampling points were arranged in a 1km×1km grid within the watershed to collect total nitrogen concentration in water samples and adsorption coefficient in soil. The total nitrogen concentration represents the concentration of water pollutants at each coordinate point, and the adsorption coefficient represents the soil adsorption capacity at each coordinate point. The discrete sampling data are converted into a continuous distribution field using a spatial interpolation algorithm, and the pollution source intensity index is calculated based on a weighted combination of vegetation influence coefficient, temperature influence coefficient and pollution coupling coefficient, where the sum of the coefficients is 1; The pollution source intensity index is normalized to an intensity value of 0-100, and the geographic coordinate information is superimposed to generate a rasterized heat map.

3. The method according to claim 1, wherein The construction of the structured pollution migration database in step 2 includes: Calculating hydrological dynamic parameters based on a velocity distribution function of a water-passing section, a water depth measurement, and an integral result of a water-passing section area varying with time; Establishing a soil pore blockage model that characterizes porosity changes over time based on an integral relationship between initial porosity, soil type-dependent attenuation coefficients, and pollutant concentrations; A pollutant diffusion equation is constructed, which characterizes the diffusion process of pollutants based on the diffusion adjustment factor, the anisotropic conductivity coefficient matrix and the gradient relationship of the pollutant concentration.

4. The method according to claim 1, wherein The construction of the hybrid neural network model in step 3 includes: A dual-channel input structure was designed. The first channel input contained spatial raster data of pollution source intensity index, hydrological dynamic parameters, and soil porosity, while the second channel input contained time series data of pollutant diffusion coefficient and pollutant concentration. A three-dimensional convolutional layer and a gated recurrent unit are alternately deployed in the hidden layer, wherein the three-dimensional convolutional layer extracts features using a Gaussian kernel function defined by a spatial kernel width and a temporal kernel width; Mass conservation constraints and concentration boundary constraints are added to the output layer to ensure that the concentration values predicted by the model conform to physical laws.

5. The method according to claim 1, wherein The execution process of the multi-stage optimization controller in step 4 includes: Parameter pre-optimization stage: By calculating the sensitivity matrix of the control parameters to the predicted concentration values, the parameters with sensitivity higher than the threshold are screened and entered into the optimization sequence; Spatial correlation optimization stage: construct a regional influence weight matrix based on regional centroid distance, concentration deviation and coupling influence coefficient to quantify the mutual correlation between regions; Time rolling optimization stage: A variable step-size predictive control algorithm is used to construct the objective function with concentration tracking weights and control cost weights to optimize the control sequence within the future rolling time domain window.

6. The method according to claim 5, wherein The parameter pre-optimization stage also includes: Establishing a dynamic feasible domain constraint condition, wherein the condition dynamically adjusts a feasible range of a control parameter based on a lower limit scaling factor, an upper limit scaling factor, and an adaptive adjustment rate; A chaotic particle swarm optimization algorithm is used for parameter search, which updates the particle position according to the particle velocity vector and the chaotic perturbation intensity.

7. The method according to claim 5, wherein The spatial association optimization stage also includes: Design a multi-objective optimization function that includes economic cost coefficient, ecological protection coefficient, and implementation difficulty penalty coefficient; The distributed ADMM algorithm is used to solve the optimization problem, and the regional correlation constraint matrix is processed by the augmented Lagrangian function.

8. The method according to claim 5, wherein The time rolling optimization phase also includes: constructing a time-varying state-space model that describes system dynamics based on an attenuation coefficient matrix, a control gain matrix, and an environmental disturbance term; Design a predictive controller to generate the optimal control sequence by minimizing the objective function in the prediction horizon; A feedback correction mechanism is implemented to trigger parameter reoptimization when the monitored concentration deviation exceeds the threshold.

9. The method according to claim 1, wherein The generation of the governance strategy map in step 5 includes: The planting density of the vegetation restoration belt is calculated based on the plant purification efficiency coefficient, grid unit area, absorption per unit area and growth cycle; Determine the depth and slope of the intercepting ditch based on peak flow, design velocity, channel bottom width, roughness coefficient, and upstream and downstream elevations; Generate a fertilization control plan based on fertilizer conversion rate, fertilizer utilization rate and control time interval.

10. The method according to claim 1, wherein The implementation of the closed-loop control link in step 6 includes: Construct a three-dimensional evaluation index including spatial weight, temporal accumulation weight and control cost weight; Design parameter adaptive update law based on learning rate, Hessian matrix and parameter gradient operator; When the residual between the model-predicted concentration value and the actual concentration exceeds a threshold, the model reconstruction process is initiated to retrain the hybrid neural network model.

Images