Basin water environment regulation method and system based on uncertainty and time-varying source analysis

Abstract

The method and system for regulating water environment of a basin based on uncertainty and time-varying source analysis relate to the technical field of regulating water environment of a basin.The method comprises the following steps: collecting multi-source monitoring data to form an observation vector; inputting a Bayesian physically constrained neural network to predict component mean, uncertainty and over-standard risk, and combining conservation residual to generate a sampling strategy; updating data according to the strategy to build a continuous water quality field constrained by water dynamics, water quality transport, water color and conservation; establishing a water quality sample-index matrix and a bias state, combining near-bottom flow velocity, bottom sediment consolidation state and critical start shear stress to calculate bottom sediment resuspension and internal source phosphorus release flux; generating external source load according to rainfall, runoff and pollutant concentration, decomposing and correcting a source contribution matrix to obtain internal and external source contribution probability; and finally taking the bias state as an object and device control input as a variable to build a predictive control problem with a control Lyapunov function decay constraint and output optimal control input.

Description

Technical Field

[0001] This invention relates to the field of watershed water environment regulation technology, and more specifically, to a watershed water environment regulation method and system based on uncertainty and time-varying source analysis. Background Technology

[0002] Watershed water environment management and water quality control are important issues in ecological and environmental protection. With the development of IoT and AI technologies, smart water systems are increasingly being used in watershed monitoring and regulation. Related fields have raised clear demands for high-frequency monitoring and accurate inversion of water quality parameters, accurate analysis of pollution sources, and stable control of treatment measures to achieve scientific management of watershed water quality and effective response to pollution risks.

[0003] Existing smart water management systems typically employ a technical approach combining fixed monitoring stations with physicochemical sensors, remote sensing imagery, and data-driven models. In practice, high-frequency monitoring is conducted using one-dimensional sensor data such as pH, dissolved oxygen, turbidity, chlorophyll a, total phosphorus, and total nitrogen. Models such as LSTM, CNN, and random forests are used for short-term water quality trend prediction. PMF or multivariate statistical models are employed for static pollution source analysis. Water quality parameters are retrieved from hyperspectral remote sensing data using empirical databases or deep learning models. These methods constitute the main technical pathways for current watershed water quality monitoring and control.

[0004] However, existing technologies have several shortcomings in practical applications. Edge-end adaptive sampling lacks a closed-loop feedback mechanism. When the error between the current predicted and measured values ​​is used as the basis for adjusting the sampling frequency, the true measured value cannot be obtained during the sensor's non-sampling intervals, making online calculation of the true error infeasible. The water quality inversion process lacks physical consistency guarantees. Purely data-driven models are prone to producing results that do not conform to mass conservation, hydrodynamic transport laws, or water color radiation mechanisms in scenarios with small samples, strong disturbances, and cross-seasonal and cross-water areas. Endogenous sediment analysis ignores the time-varying nature of the consolidation state. Traditional models treat the critical initiation shear stress of sediment as a constant, failing to reflect the dynamic impact of low-velocity static periods, sediment consolidation, water content changes, and biological disturbances on the sediment's scour resistance. Water body treatment and control suffer from large time lags and oscillation risks. The effects of aeration, flow propulsion, algae control, and chemical dosing measures have lags of several hours to several days. The lack of stability constraints easily leads to frequent start-ups and shutdowns of implementing equipment, increased energy consumption, and even problems such as enhanced sediment resuspension caused by bottom disturbances. Summary of the Invention

[0005] This invention provides a watershed water environment regulation method and system based on uncertainty and time-varying source analysis, in order to improve at least one of the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides a watershed water environment regulation method based on uncertainty and time-varying source analysis, which includes steps S1 to S6.

[0007] S1. Collect multi-source monitoring data of the watershed to form a multi-source observation vector.

[0008] S2. Input the multi-source observation vector into the Bayesian physical constraint neural network to obtain the predicted mean, cognitive uncertainty and probability of exceeding the standard of water quality components, and then combine the finite volume conservation residual to generate an active sampling strategy.

[0009] S3. Update the observation data based on the active sampling strategy and construct a continuous water quality field that is jointly constrained by hydrodynamic constraints, water quality transport constraints, water color constraints and conservation constraints.

[0010] S4. Construct a water quality sample-index matrix and water quality deviation state based on the continuous water quality field, and obtain the sediment resuspension flux and endogenous phosphorus release flux based on the near-bottom flow velocity, sediment consolidation state and time-varying critical starting shear stress.

[0011] S5. Generate external runoff loads based on rainfall, runoff volume, and surface pollutant concentrations. Decompose the water quality sample-index matrix into a source contribution matrix and a factor load matrix. Combine the sediment resuspension flux and external runoff loads to correct the source contribution matrix and obtain the endogenous contribution probability and the exogenous contribution probability.

[0012] S6. Taking the water quality deviation state as the prediction object and the control input of the treatment execution equipment as the optimization variable, the future water quality deviation state under the action of the control input is predicted using a neural network prediction model. Based on the internal contribution probability and the external contribution probability, a predictive control problem with control Lyapunov function decay constraint is constructed. The optimal control input for the current control step is solved and output.

[0013] As a preferred aspect of the present invention, the multi-source monitoring data includes water body hyperspectral remote sensing reflectance, water quality physicochemical parameters, hydrodynamic parameters, water level, water temperature, and rainfall.

[0014] The multi-source observation vector is:

[0015] .

[0016] In the formula, express Multi-source observation vector at time step. Indicates wavelength Place Water body remote sensing reflectance at any given time. express pH value of the water at any given time. express Dissolved oxygen concentration at any given time. express Turbidity at any given moment. express Chlorophyll a concentration at time. express Total phosphorus concentration at time. express Total nitrogen concentration at time t. express The flow velocity or flow velocity profile at any given time. express The water level at any given time. express The water temperature at any given time. express Rainfall at any given moment.

[0017] The water quality sample-index matrix is ​​formed by the water quality physicochemical parameters in the multi-source observation vector and the concentration of water quality components corresponding to the sampling time and space location of the continuous water quality field.

[0018] As a preferred aspect of the present invention, the Bayesian physical constraint neural network in S2 obtains the predicted mean and cognitive uncertainty of water quality components through multiple forward samplings.

[0019] .

[0020] .

[0021] In the formula, Indicates sampling position ,time Predicted mean values ​​of water quality components. Indicates sampling position ,time Cognitive uncertainty at the point. This indicates the number of forward samples taken in Monte Carlo. This represents the Monte Carlo sampling index. Indicates the first Second sampling parameters The predicted output of the network. Indicates the first The parameters are then sampled from the variational posterior. Representing network parameters The variational posterior distribution. The parameter represents the variational posterior distribution. This represents the parameters of a Bayesian physics-constrained neural network.

[0022] Pollution indicators exceeded the limit The probability of risk is: .

[0023] In the formula, Indicators of pollution In position ,time The probability of exceeding the limit. Indicates the first Pollution indicators for each water quality component. Indicators of pollution The standard limit. Indicators of pollution In position ,time The predicted mean. Indicators of pollution In position ,time Cognitive uncertainty. Indicators of pollution In position ,time The random uncertainty. This indicates a positive number used to avoid a denominator of zero. The cumulative distribution function represents the standard normal distribution.

[0024] As a preferred aspect of the present invention, the active sampling strategy in S2 is as follows: first, the active sampling score is calculated based on cognitive uncertainty, probability of exceeding the standard risk, physical equation residual, and finite volume conservation residual. .

[0025] In the formula, Indicates position ,time Active sampling score. This represents the cognitive uncertainty score, which is the aggregated and normalized cognitive uncertainty of each pollution indicator. This represents the score item for the risk probability of exceeding the standard, which is the aggregated and normalized risk probability of each pollution indicator. This represents the residual of the physical equation after normalization. This represents the normalized finite volume conservation residual. This represents the energy cost of sampling after normalization. for The weighting coefficients. for The weighting coefficients. for The weighting coefficients. for The weighting coefficients. for The weighting coefficients.

[0026] The sampling period for fixed data collection stations is: .

[0027] In the formula, express The sampling period of a fixed station is determined at all times. Indicates the reference sampling period. This represents the truncation function. This represents the natural exponential function. This represents the sensitivity coefficient of the sampling score to the period adjustment. Indicates fixed station location Place Active sampling score at any given time. This indicates the sampling location of the fixed site. This indicates the lower limit of the sampling period. This indicates the upper limit of the sampling period.

[0028] The next sampling location for the mobile sampling platform is: .

[0029] In the formula, This indicates the next optimal sampling position for the mobile sampling platform. This represents the independent variable that maximizes the objective function. Indicates the sampling location. This represents the spatial domain that can be sampled. This represents the movement cost weighting coefficient that converts the movement distance to the same scale as the active sampling score. Indicates candidate position With current location The distance traveled between them. This indicates the current location of the mobile sampling platform.

[0030] As a preferred aspect of the present invention, the hydrodynamic constraint is formed by a two-dimensional shallow water equation satisfied by the water depth and flow velocity: .

[0031] .

[0032] In the formula, Indicates water depth. Indicates time. This represents the velocity vector. This indicates the source and sink of water. It represents the acceleration due to gravity. Indicates the water surface elevation. This represents the eddy viscosity coefficient. This indicates wind stress. This represents the shear stress on the bed surface. This indicates the density of the water. This represents a partial differential operator. This represents the gradient operator. This represents the divergence operator. Represents the tensor product.

[0033] The water transport constraints are formed by the convection-diffusion-reaction equations satisfied by each water component: .

[0034] In the formula, Indicates the first The residuals of the convection-diffusion-reaction equations for various water quality components. This indicates an index of water quality components. Indicates the first Network-predicted concentration fields of various water quality components. Indicates the first The equivalent diffusion coefficient of a water quality component. Indicates the first The reaction sources and sinks of various water quality components. Indicates the sampling location. This represents the set of network predicted concentration fields for each water quality component. This represents the reaction parameter vector.

[0035] The water color constraint is formed by a semi-analytical bio-optical model of the water body and theoretical remote sensing reflectance. The semi-analytical bio-optical model of the water body includes the absorption coefficient and the backscattering coefficient.

[0036] .

[0037] .

[0038] In the formula, Indicates wavelength The total absorption coefficient at that location. Indicates the wavelength of pure water The absorption coefficient at that location. This represents the set of water quality component indices that participate in hyperspectral inversion. Indicates the first Water quality components at wavelength The specific absorption coefficient at that location. Indicates wavelength The total backscattering coefficient at that location. Indicates the wavelength of pure water The backscattering coefficient at that location. Indicates the first Water quality components at wavelength The backscattering coefficient at that location.

[0039] The theoretical remote sensing reflectance is: .

[0040] In the formula, Indicates wavelength The theoretical remote sensing reflectance at that location. This represents the first empirical coefficient of the Gordon model. This represents the second empirical coefficient of the Gordon model. Indicates wavelength The backscattering ratio at that location.

[0041] The conservation constraints are formed by the finite volume conservation residuals of each water quality component on the control volume: .

[0042] In the formula, Indicates control body In the The first step of time Finite volume conservation residuals of various water quality components. Indicates the control body. Indicates control body The area or volume. Indicates the time step index. Indicates the first Time step Water quality components in control system The average concentration within. Indicates the first Time step Water quality components in control system The average concentration within. Indicates the time step. Indicates the control volume boundary surface index. Indicates control body The set of boundary surfaces. Indicates the first Time step Water quality components are controlled through the body boundary surface The flux. Indicates the first Time step control of the body's first Sources and sinks of various water quality components.

[0043] As a preferred aspect of the present invention, the sediment resuspension flux and endogenous phosphorus release flux of S4 are determined by a time-varying critical shear stress model coupled with the sediment consolidation state.

[0044] The bed shear stress is: .

[0045] In the formula, express The bed shear stress at any given moment. This indicates the density of the water. This represents the coefficient of friction of the bed surface. express The near-bottom velocity vector at time t. express The magnitude of the near-bottom velocity vector at any given time. It represents the acceleration due to gravity. Indicates the hydraulic radius. This indicates the frictional gradient.

[0046] The evolution model of the consolidation state variables of the sediment is as follows: .

[0047] In the formula, Represents the differential symbol. express The consolidation state variable of the bottom mud at time t. Indicates time. This represents the consolidation rate constant. This represents the natural exponential function. This represents the activation energy of the consolidation process. This represents the universal gas constant. express The water temperature at any given time. express When the bed shear stress is less than the static stability threshold, the value is taken as follows: Otherwise, the value is... Indicator functions. This represents the static steady-state threshold shear stress. This represents the rate constant of consolidation degradation caused by erosion. express The time-varying critical starting shear stress at time t. This represents the positive value truncation function. This represents the rate constant of the disruption of consolidation caused by biological disturbance. express The intensity of biological disturbance at any given moment.

[0048] The time-varying critical starting shear stress is: .

[0049] In the formula, This represents the reference critical shear stress. This represents the volume fraction of clay. express Moisture content or pore water state index at any given time. This represents the exponential coefficient corresponding to the volume fraction of clay. This represents the exponential coefficient corresponding to the consolidation state variable. This represents the exponential coefficient corresponding to the water content or pore water state index. This represents the exponential coefficient corresponding to the intensity of biological disturbance.

[0050] Real-time re-floating is: .

[0051] In the formula, express The resuspension criterion ratio at time.

[0052] The sediment resuspension flux and endogenous phosphorus release flux are respectively: .

[0053] .

[0054] In the formula, express Momentary sediment resuspension flux. This represents the erosion coefficient. This represents the power of erosion. express The endogenous phosphorus release flux at any given time. This indicates the phosphorus content in the sediment. This represents the diffusion release rate constant. This indicates the phosphorus concentration in pore water. This indicates the phosphorus concentration in the overlying water. This represents the regulatory function of dissolved oxygen, pH, and water temperature on phosphorus release. express Dissolved oxygen concentration at any given time. express pH value of the water at any given time.

[0055] As a preferred aspect of the present invention, S5 specifically includes: External runoff loads are generated based on rainfall, runoff volume, and surface pollutant concentrations.

[0056] .

[0057] In the formula, express The external runoff load proxy variable at time. express The intensity of rainfall at any given moment. express Runoff volume at any given time. express Agent for the concentration of surface pollutants at any given time.

[0058] Perform nonnegative matrix decomposition on the water quality sample-index matrix: .

[0059] In the formula, This represents the sample-index matrix. Indicates the number of samples. Indicates the number of indicators. This represents the contribution matrix. This represents the number of source factors. This represents the factor loading matrix. This represents the residual matrix.

[0060] The objective function of the fundamental positive matrix factorization of nonnegative matrix factorization is: .

[0061] In the formula, This represents the objective function value of the fundamental positive matrix factorization. Indicates the sample index. Indicates the index of the indicator. Indicates the first The first sample The observed values ​​of each indicator. This indicates the product of the contribution matrix and the factor loading matrix, corresponding to the first... The first sample The reconstructed values ​​of each indicator. Indicates the first The first sample The uncertainty of each indicator.

[0062] The source apportionment objective function value after adding internal hydrodynamic constraints and external runoff constraints is: .

[0063] In the formula, This represents the source analysis objective function value after incorporating internal hydrodynamic constraints and external runoff constraints. This represents the weight of the internal constraint regularization. This represents the external constraint regularization weight. This represents the set of time indices used in dynamic pollution source analysis. Indicates time. express The contribution value of the source factor within a given time step. This represents the intrinsic factor index. This represents the scale matching coefficient of the intrinsic constraints. This represents the resuspension flux estimated by the model. express The contribution value of exogenous factors at time step. This indicates the exogenous factor index. This represents the scale matching coefficient of the external constraint.

[0064] The contribution values ​​of intrinsic and extrinsic factors obtained from source analysis are normalized to obtain the intrinsic contribution probability and extrinsic contribution probability.

[0065] As a preferred aspect of the present invention, the water quality deviation state vector of S4 is: .

[0066] In the formula, Indicates the first The water quality deviation state vector for each step, with positive values ​​indicating either exceeding or failing to meet standards. Indicates the control step index. This indicates the target value for dissolved oxygen safety. Indicates the first Step-by-step actual dissolved oxygen concentration. Indicates the first The actual chlorophyll concentration. This indicates the safety target value for chlorophyll. Indicates the first Step-by-step actual total phosphorus concentration. This indicates the safety target value for total phosphorus. Indicates the first Step-by-step actual ammonia nitrogen concentration. This indicates the safety target value for ammonia nitrogen. This indicates transpose.

[0067] As a preferred aspect of the present invention, S6 includes: constructing a delayed embedded state with the water quality deviation state as the prediction object and the control input of the treatment execution device as the optimization variable, and inputting the delayed embedded state, the control input, and the disturbance estimate into a neural network prediction model to predict the future water quality deviation state under the action of the control input.

[0068] The delayed embedding state is: .

[0069] In the formula, Indicates the first The delayed embedding state vector of the step. Indicates the first The water quality deviation state vector of the step. Indicates the first The water quality deviation state vector of the step. This indicates the state delay order. Indicates the first Step-by-step historical control input. Indicates the first Step-by-step historical control input. This indicates the delay order of the control input. Indicates the first The endogenous phosphorus release flux of the step. Indicates the first External runoff load of the step.

[0070] The prediction model and the disturbance observer are as follows: .

[0071] .

[0072] In the formula, Indicates the first The water quality deviation state vector of the step. The parameter is The neural network prediction model. Indicates the first Step control input. Indicates the first The perturbation estimate of the step. Indicates the first The perturbation estimate of the step. This represents the gain matrix of the perturbation observer. Indicates the first The actual observation output of the step. Indicates the first The model predicts the output step by step.

[0073] As a preferred aspect of the present invention, S6 further includes: Construct a control Lyapunov function to constrain the convergence of the water quality deviation state: .

[0074] In the formula, Indicates the first The step controls the Lyapunov function value. This represents a positive definite symmetric weight matrix. This indicates that the matrix is ​​positive definite.

[0075] When the probability of intrinsic contribution increases, the control input penalty corresponding to the underlying disturbance device is increased, and the additional bed shear stress caused by the governance execution device is limited by the anti-resuspending shear constraint.

[0076] The control input weight matrix and the anti-resuspending shear constraint are as follows: .

[0077] .

[0078] In the formula, Indicates the first The control input weight matrix for each step. This represents the baseline control weight matrix. This represents the intrinsic penalty gain. This represents the endogenous contribution probability obtained from dynamic pollution source analysis based on positive matrix factorization. This represents a diagonal matrix used to impose additional penalties on the flow generator and aeration layer disturbance equipment. Indicates the first The bed surface shear stress is caused by hydrodynamic forces. Indicates the first Step by control input Additional bed shear stress caused by the corresponding actuator. This represents the safety margin coefficient. Indicates the first The time-varying critical starting shear stress of the step.

[0079] When the probability of external contribution increases, the dynamic Lyapunov decay rate used to limit the convergence speed of water quality deviation is increased.

[0080] The dynamic Lyapunov decay rate of the current control step is: .

[0081] In the formula, Indicates the first The dynamic Lyapunov decay rate of the step. This represents the truncation function. This represents the baseline attenuation rate. This represents the gain coefficient of the external contribution to the attenuation rate. This represents the external contribution probability obtained from dynamic pollution source analysis based on positive matrix factorization. This represents the inhibition coefficient of the intrinsic contribution on the decay rate. This indicates the lower limit of the decay rate. This indicates the upper limit of the decay rate.

[0082] A predictive control optimization problem is constructed based on the control Lyapunov function, the control input weight matrix, and the dynamic Lyapunov decay rate. The control Lyapunov function characterizes the convergence degree of water quality deviation in the prediction time domain, the dynamic Lyapunov decay rate limits the decay rate of the control Lyapunov function, and the control input weight matrix limits the penalty intensity of the control inputs for different governance execution devices.

[0083] The objective function of the predictive control optimization problem is: .

[0084] In the formula, This represents the set of optimization variables. This represents the control step index within the prediction time domain. This indicates the length of the prediction time domain. Indicates the first The water quality deviation state vector of the step. This represents the state deviation weight matrix. Indicates the first Step control input. Indicates the first Step relative to the first The control increment of the step. This represents the control increment weight matrix. This represents the penalty coefficient for slack variables. The first Lyapunov decay constraint acting on the third Slack off variables step by step. This represents the Lyapunov function value for terminal control.

[0085] The predictive control optimization problem is solved in a rolling manner, and the control input of the current control step obtained by the solution is used as the optimal control input of the current control step.

[0086] Secondly, the present invention provides a watershed water environment regulation system based on uncertainty and time-varying source analysis, which includes a multi-dimensional heterogeneous sensing terminal, an edge control layer, a cloud computing center, and governance execution equipment.

[0087] The multidimensional heterogeneous sensing terminal is used to collect hyperspectral, water quality physicochemical, hydrodynamic, water level, water temperature, rainfall and equipment operation status data.

[0088] The edge control layer is used to clean, synchronize, cache anomalies, and register spatial coordinates of the data uploaded by the multidimensional heterogeneous sensing terminal, and adjust the sampling cycle of fixed sites or the next sampling position of the mobile sampling platform according to the active sampling strategy.

[0089] The cloud computing center is used to run Bayesian physical constraint neural networks, time-varying critical shear stress models coupled with the consolidation state of sediment, dynamic pollution source analysis models based on positive matrix factorization, and model predictive controllers with constraints on controlling the decay of Lyapunov functions, in order to perform the methods described above.

[0090] The treatment execution device is used to receive control quantities issued by the edge master control layer and perform aeration, algae control, flow propulsion or chemical dosing operations.

[0091] By adopting the above technical solution, the present invention can achieve the following technical effects: This invention uses uncertainty, exceedance probability, physical equation residuals, and finite volume conservation residuals as sampling feedback quantities. It dynamically adjusts the sampling strategy without relying on real-time measured values, effectively solving the problem of being unable to calculate the true prediction error during unsampled periods, thus improving the relevance and effectiveness of monitoring data. Simultaneously, this invention embeds hydrodynamic constraints, water quality transport constraints, water color constraints, and conservation constraints into the water quality field inversion network, significantly enhancing the physical rationality of water quality inversion results under sparse samples and complex hydrodynamic conditions. This avoids problems such as violations of mass conservation and hydrodynamic laws that are prone to occur in purely data-driven models.

[0092] This invention further introduces a time-varying critical shear stress model coupled with the consolidation state of sediment, transforming the determination of endogenous resuspension from a fixed threshold to a physical criterion that dynamically evolves with environmental factors such as consolidation degree, water content, and biological disturbance, significantly improving the accuracy of endogenous release identification. Furthermore, this invention couples the contribution matrix from positive matrix factorization with the weights of resuspension flux, runoff load, and seasonal characteristics, effectively enhancing the ability to distinguish between endogenous sediment release and exogenous runoff input. Simultaneously, through source apportionment-driven Lyapunov stability constraints and anti-resuspension shear constraints, it reduces control oscillations, equipment wear, and the risk of secondary sediment release in the treatment of large-time-delay water bodies, achieving more stable and efficient water environment regulation. Attached Figure Description

[0093] Figure 1 This is a flowchart illustrating a watershed water environment regulation method based on uncertainty and time-varying source analysis.

[0094] Figure 2 This is a system architecture diagram of a watershed water environment regulation method based on uncertainty and time-varying source analysis. Detailed Implementation

[0095] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0096] Example 1, please refer to Figures 1 to 2 The first embodiment of the present invention provides a watershed water environment control method based on uncertainty and time-varying source analysis. It can be executed by a watershed water environment control device based on uncertainty and time-varying source analysis (hereinafter referred to as: watershed water environment control device). Specifically, it is executed by one or more processors in the watershed water environment control device to implement steps S1 to S6.

[0097] S1. Collect multi-source monitoring data of the watershed to form a multi-source observation vector.

[0098] In this embodiment, a multi-dimensional heterogeneous sensing terminal collects multi-source monitoring data of the watershed. The multi-source monitoring data includes water body hyperspectral remote sensing reflectance, water quality physicochemical parameters, hydrodynamic parameters, water level, water temperature, and rainfall. Water quality physicochemical parameters include at least one of pH, dissolved oxygen (DO), turbidity, chlorophyll a (Chl-a), total phosphorus (TP), and total nitrogen (TN). Hydrodynamic parameters include flow velocity or flow velocity profiles. The edge master control layer performs timestamp unification, outlier marking, missing value encoding, and spatial coordinate registration on the multi-frequency data to form a multi-source observation vector.

[0099] The multi-source observation vector is: .

[0100] In the formula, express Multi-source observation vector at time step. Indicates wavelength Place Water body remote sensing reflectance at any given time. express pH value of the water at any given time. express Dissolved oxygen concentration at any given time. express Turbidity at any given moment. express Chlorophyll a concentration at time. express Total phosphorus concentration at time. express Total nitrogen concentration at time t. express The flow velocity or flow velocity profile at any given time. express The water level at any given time. express The water temperature at any given time. express Rainfall at any given moment.

[0101] S2. Input the multi-source observation vector into the Bayesian physical constraint neural network to obtain the predicted mean, cognitive uncertainty and probability of exceeding the standard of water quality components, and then combine the finite volume conservation residual to generate an active sampling strategy.

[0102] In this embodiment, the cloud computing center uses a Bayesian physics-constrained neural network to process multi-source observation vectors. Let the network parameters be... The variational posterior distribution is The Bayesian physical constraint neural network obtains the predicted mean and cognitive uncertainty of water quality components through multiple forward sampling.

[0103] .

[0104] .

[0105] In the formula, Indicates sampling position ,time Predicted mean values ​​of water quality components. Indicates sampling position ,time Cognitive uncertainty at the point. This indicates the number of forward samples taken in Monte Carlo. This represents the Monte Carlo sampling index. Indicates the first Second sampling parameters The predicted output of the network. Indicates the first The parameters are then sampled from the variational posterior. Representing network parameters The variational posterior distribution. The parameter represents the variational posterior distribution. This represents the parameters of a Bayesian physics-constrained neural network. .

[0106] Pollution indicators exceeded the limit The probability of risk is: .

[0107] In the formula, Indicators of pollution In position ,time The probability of exceeding the limit. Indicates the first Pollution indicators for each water quality component. Indicators of pollution The standard limit. Indicators of pollution In position ,time The predicted mean. Indicators of pollution In position ,time Cognitive uncertainty. Indicators of pollution In position ,time The random uncertainty. This indicates a positive number used to avoid a denominator of zero. The cumulative distribution function represents the standard normal distribution.

[0108] The active sampling strategy is as follows: First, calculate the active sampling score based on cognitive uncertainty, probability of exceeding the limit, physical equation residuals, and finite volume conservation residuals. .

[0109] In the formula, Indicates position ,time Active sampling score. This represents the cognitive uncertainty score, which is the aggregated and normalized cognitive uncertainty of each pollution indicator. This represents the score item for the risk probability of exceeding the standard, which is the aggregated and normalized risk probability of each pollution indicator. This represents the residual of the physical equation after normalization. This represents the normalized finite volume conservation residual. This represents the energy cost of sampling after normalization. for The weighting coefficients. for The weighting coefficients. for The weighting coefficients. for The weighting coefficients. for The weighting coefficients.

[0110] Specifically, all input items are normalized to the same scoring scale before being included in the active sampling score: The residuals of the physical equations are obtained as follows: water depth is determined based on water level data collected by a water level gauge and river cross-section or bottom shape data; the velocity vector is determined based on the velocity profile; and the increase or decrease in water volume is determined based on rainfall, tributary inflow, gate discharge, or pumping station drainage. The water depth, velocity vector, and the increase or decrease in water volume are substituted into the two-dimensional shallow water equations, and both sides of the two-dimensional shallow water equations are shifted to the same side to calculate the unbalance of the two-dimensional shallow water equations. The velocity profile is obtained by an acoustic Doppler current profiler (ADCP).

[0111] Substitute the water quality component concentration field predicted by the network, the flow velocity vector, the equivalent diffusion coefficient, and the increase or decrease in concentration of water quality components due to sedimentation, release, consumption, or transformation into the convection-diffusion-reaction equation, and shift both sides of the convection-diffusion-reaction equation to the same side to calculate the unbalance quantity of the convection-diffusion-reaction equation.

[0112] After processing the unbalance quantities of the two-dimensional shallow water equation and the convection-diffusion-reaction equation by absolute or square value processing, they are aggregated according to preset weights and normalized to the same scoring scale of the active sampling score to obtain the normalized physical equation residuals.

[0113] The finite volume conservation residual is used to characterize the mass conservation deviation at the control volume scale and participates in the calculation as an independent input item in the active sampling score, without being repeatedly included in the physical equation residual.

[0114] The energy consumption cost of sampling is determined based on the distance the mobile sampling platform travels from its current location to the sampling location, the energy consumption per unit distance, the power of the sampling equipment, the duration of a single sampling, the power of communication upload, and the duration of communication upload.

[0115] The sampling period for fixed data collection stations is: .

[0116] In the formula, express The sampling period of a fixed station is determined at all times. Indicates the reference sampling period. This represents the truncation function. This represents the natural exponential function. This represents the sensitivity coefficient of the sampling score to the period adjustment. Indicates fixed station location Place Active sampling score at any given time. This indicates the sampling location of the fixed site. This indicates the lower limit of the sampling period. This indicates the upper limit of the sampling period.

[0117] The next sampling location for the mobile sampling platform is: .

[0118] In the formula, This indicates the next optimal sampling position for the mobile sampling platform. This represents the independent variable that maximizes the objective function. Indicates the sampling location. This represents the spatial domain that can be sampled. This represents the movement cost weighting coefficient that converts the movement distance to the same scale as the active sampling score. Indicates candidate position With current location The distance traveled between them. This indicates the current location of the mobile sampling platform.

[0119] Therefore, the system can proactively adjust its sampling behavior based on model confidence and physical consistency without waiting for new measured true values.

[0120] S3. Update the observation data based on the active sampling strategy and construct a continuous water quality field that is jointly constrained by hydrodynamic constraints, water quality transport constraints, water color constraints and conservation constraints.

[0121] In this embodiment, the edge master control layer adjusts the sampling period of fixed stations or the next sampling location of the mobile sampling platform according to the active sampling strategy, and uploads the updated observation data to the cloud computing center. The cloud computing center constructs a continuous water quality field based on the updated observation data. Specifically, the measured concentrations are only discrete data at certain monitoring points and at certain times. Based on these discrete data, the Bayesian physical constraint neural network inverts or completes the water quality component concentration field of the entire watershed in continuous space and continuous time.

[0122] The hydrodynamic constraints are formed by the two-dimensional shallow water equations satisfied by the water depth and flow velocity: .

[0123] .

[0124] In the formula, Indicates water depth. Indicates time. This represents the velocity vector. This indicates the source and sink of water. It represents the acceleration due to gravity. Indicates the water surface elevation. This represents the eddy viscosity coefficient. This indicates wind stress. This represents the shear stress on the bed surface. This indicates the density of the water. This represents a partial differential operator. This represents the gradient operator. This represents the divergence operator. Represents the tensor product.

[0125] The water source and sink terms represent the changes in water volume that cause an increase or decrease in water depth per unit time, and .

[0126] The water quality transport constraints are formed by the convection-diffusion-reaction equations satisfied by each water quality component. The residuals of these convection-diffusion-reaction equations are also referred to as the water quality transport-reaction equation residuals.

[0127] .

[0128] In the formula, Indicates the first The residuals of the convection-diffusion-reaction equations for various water quality components. This indicates an index of water quality components. Indicates the first Network-predicted concentration fields of various water quality components. Indicates the first The equivalent diffusion coefficient of a water quality component. Indicates the first The reaction sources and sinks of various water quality components. Indicates the sampling location. This represents the set of network predicted concentration fields for each water quality component. This represents the reaction parameter vector.

[0129] The equivalent diffusion coefficient is expressed in a nonnegative parameterized form: .

[0130] In the formula, Indicates the first Water quality components at sampling locations ,time The equivalent diffusion coefficient. This represents the minimum value of the diffusion coefficient, and This represents the diffusion coefficient scaling factor, and This represents a smooth, non-negative activation function. Indicates the network output of the first Dimensionless unconstrained diffusion field of various water quality components.

[0131] The water color constraint is formed by a semi-analytical bio-optical model of the water body and theoretical remote sensing reflectance. The semi-analytical bio-optical model of the water body includes the absorption coefficient and the backscattering coefficient.

[0132] The water color constraint consists of two parts: First, based on the bio-optical semi-analysis model of the water body, the absorption coefficient and backscattering coefficient of the water body are calculated according to the concentration of water quality components predicted by the network. Then, the theoretical remote sensing reflectance is calculated based on the absorption coefficient and backscattering coefficient, and the deviation between the theoretical remote sensing reflectance and the measured hyperspectral remote sensing reflectance is taken as the water color reflectance loss.

[0133] The absorption coefficient and backscattering coefficient are respectively: .

[0134] .

[0135] In the formula, Indicates wavelength The total absorption coefficient at that location. Indicates the wavelength of pure water The absorption coefficient at that location. This represents the set of water quality component indices that participate in hyperspectral inversion. Indicates the first Water quality components at wavelength The specific absorption coefficient at that location. Indicates wavelength The total backscattering coefficient at that location. Indicates the wavelength of pure water The backscattering coefficient at that location. Indicates the first Water quality components at wavelength The backscattering coefficient at that location.

[0136] The theoretical remote sensing reflectance is: .

[0137] In the formula, Indicates wavelength The theoretical remote sensing reflectance at that location. This represents the first empirical coefficient of the Gordon model. This represents the second empirical coefficient of the Gordon model. Indicates wavelength The backscattering ratio at that location. .

[0138] To avoid local mass non-conservation caused by ordinary PINN relying solely on point residuals, this invention further introduces finite volume conservation residuals.

[0139] The conservation constraints are formed by the finite volume conservation residuals of each water quality component on the control volume: .

[0140] In the formula, Indicates control body In the The first step of time Finite volume conservation residuals of various water quality components. Indicates the control body. Indicates control body The area or volume. Indicates the time step index. Indicates the first Time step Water quality components in control system The average concentration within. Indicates the first Time step Water quality components in control system The average concentration within. Indicates the time step. Indicates the control volume boundary surface index. Indicates control body The set of boundary surfaces. Indicates the first Time step Water quality components are controlled through the body boundary surface The flux. Indicates the first Time step control of the body's first Sources and sinks of various water quality components. .

[0141] The overall training objective of the network is: .

[0142] In the formula, The parameter represents the variational posterior distribution. It represents the set consisting of the logarithmic variance parameters of each term. Indicates the parameter Follows variational posterior distribution The expectation. This represents the parameters of a Bayesian physics-constrained neural network. Representing network parameters The variational posterior distribution. This represents the index of the loss term, with values ​​ranging from all loss terms in the network training objective. Indicates the first The logarithmic variance parameter of the loss term. Indicates the first Item of loss. Indicates the first The variance scaling term corresponding to the loss term. This represents the KL divergence. Indicates parameters The prior distribution of .

[0143] This includes data loss, shallow water equation residuals, convection-diffusion-reaction equation residuals, water color reflectance loss, finite volume conservation residuals, initial value loss, and boundary loss.

[0144] S4. Construct a water quality sample-index matrix and water quality deviation state based on the continuous water quality field, and obtain the sediment resuspension flux and endogenous phosphorus release flux based on the near-bottom flow velocity, sediment consolidation state and time-varying critical starting shear stress.

[0145] The water quality sample-index matrix is ​​formed by the water quality physicochemical parameters in the multi-source observation vector and the concentration of water quality components corresponding to the sampling time and space location of the continuous water quality field.

[0146] Specifically, before conducting dynamic pollution source analysis, based on the continuous water quality field... In addition to the water quality physicochemical parameters in the multi-source observation vectors, a water quality sample-index matrix is ​​constructed. .in, Indicates the number of samples. This represents the number of water quality indicators involved in source apportionment. Each sampling spatiotemporal unit is considered a sample, and the sampling spatiotemporal unit is defined by the sampling location. and sampling time Sure, The water quality indicators used in pollution source analysis are used as matrix columns. These water quality indicators include at least one of dissolved oxygen, turbidity, chlorophyll a, total phosphorus, total nitrogen, and ammonia nitrogen. .

[0147] For the The first sample and the first Several water quality indicators are used to determine whether there are any correlations between the multi-source observation vectors and the sampling location. and sampling time The corresponding measured water quality physicochemical parameters. If they exist, the measured values, after time synchronization, outlier removal, and dimensional unification, will be used as candidate observations. If they do not exist, the continuous water quality field will be used. At the sampling location and sampling time field value Generate candidate supplementary values. For water quality indicators that have both measured values ​​and continuous water quality field values, the measured values ​​and continuous water quality field values ​​are weighted according to a preset fusion weight. Weighted fusion is performed to obtain matrix elements. For water quality indicators that only have continuous water quality field values, As matrix elements This results in a water quality sample-index matrix. .

[0148] In constructing the water quality sample-indicator matrix At the same time, water quality indicators of different dimensions are subjected to uniform scaling, and matrix elements used for nonnegative matrix factorization are nonnegated to achieve the desired effect. This satisfies the input requirements for positive matrix factorization.

[0149] The water quality deviation state vector is: .

[0150] In the formula, Indicates the first The water quality deviation state vector for each step, with positive values ​​indicating either exceeding or failing to meet standards. Indicates the control step index. This indicates the target value for dissolved oxygen safety. Indicates the first Step-by-step actual dissolved oxygen concentration. Indicates the first The actual chlorophyll concentration. This indicates the safety target value for chlorophyll. Indicates the first Step-by-step actual total phosphorus concentration. This indicates the safety target value for total phosphorus. Indicates the first Step-by-step actual ammonia nitrogen concentration. This indicates the safety target value for ammonia nitrogen. This indicates transpose.

[0151] The sediment resuspension flux and endogenous phosphorus release flux were determined by a time-varying critical shear stress model coupled with the sediment consolidation state.

[0152] The bed shear stress is: .

[0153] In the formula, express The bed shear stress at any given moment. This indicates the density of the water. This represents the coefficient of friction of the bed surface. express The near-bottom velocity vector at time t. express The magnitude of the near-bottom velocity vector at any given time. It represents the acceleration due to gravity. Indicates the hydraulic radius. This indicates the frictional gradient. .

[0154] Near-bottom velocity vector The velocity profile is obtained from flow profile data. Specifically, the flow profile is acquired by an acoustic Doppler velocity profiler, and after time synchronization and outlier removal, the lowest laminar velocity measurement point close to the sediment surface and within the effective measurement range is selected as the near-bottom velocity. When there is a distance deviation between the lowest laminar velocity measurement point and the sediment surface, interpolation is performed based on two or more adjacent effective velocity measurement points to obtain the near-bottom velocity vector at a preset near-bottom height. When a complete velocity profile is not obtained, the near-bottom velocity vector at a preset near-bottom height is estimated based on the measured average velocity, water depth, and channel cross-sectional parameters. .

[0155] The evolution model of the consolidation state variables of the sediment is as follows: .

[0156] In the formula, Represents the differential symbol. express The consolidation state variable of the bottom mud at time t. Indicates time. This represents the consolidation rate constant. This represents the natural exponential function. This represents the activation energy of the consolidation process. This represents the universal gas constant. express The water temperature at any given time. express When the bed shear stress is less than the static stability threshold, the value is taken as follows: Otherwise, the value is... Indicator functions. This represents the static steady-state threshold shear stress. This represents the rate constant of consolidation degradation caused by erosion. express The time-varying critical starting shear stress at time t. This represents the positive value truncation function. This represents the rate constant of the disruption of consolidation caused by biological disturbance. express The intensity of biological disturbance at any given moment.

[0157] The temperature in the observation vector is obtained through time synchronization. The range of values ​​is .

[0158] The time-varying critical starting shear stress is: .

[0159] In the formula, This represents the reference critical shear stress. This represents the volume fraction of clay. express Moisture content or pore water state index at any given time. This represents the exponential coefficient corresponding to the volume fraction of clay. This represents the exponential coefficient corresponding to the consolidation state variable. This represents the exponential coefficient corresponding to the water content or pore water state index. This represents the exponential coefficient corresponding to the intensity of biological disturbance.

[0160] Increased clay content and degree of consolidation raise the critical shear stress, while increased water content and bioturbation reduce the sediment's resistance to erosion.

[0161] Real-time re-floating is: .

[0162] In the formula, express The resuspension criterion ratio at time.

[0163] The sediment resuspension flux and endogenous phosphorus release flux are respectively: .

[0164] .

[0165] In the formula, express Momentary sediment resuspension flux. This represents the erosion coefficient. This represents the power of erosion. express The endogenous phosphorus release flux at any given time. This indicates the phosphorus content in the sediment. This represents the diffusion release rate constant. This indicates the phosphorus concentration in pore water. This indicates the phosphorus concentration in the overlying water. This represents the regulatory function of dissolved oxygen, pH, and water temperature on phosphorus release. express Dissolved oxygen concentration at any given time. express pH value of the water at any given time.

[0166] S5. Generate external runoff loads based on rainfall, runoff volume, and surface pollutant concentrations. Decompose the water quality sample-index matrix into a source contribution matrix and a factor load matrix. Combine the sediment resuspension flux and external runoff loads to correct the source contribution matrix and obtain the endogenous contribution probability and the exogenous contribution probability.

[0167] Specifically, external runoff loads are generated on behalf of the government based on rainfall, runoff volume, and surface pollutant concentrations.

[0168] .

[0169] In the formula, express The external runoff load proxy variable at time. express The intensity of rainfall at any given moment. express Runoff volume at any given time. express Agent for the concentration of surface pollutants at any given time.

[0170] Agent for surface pollutant concentration The determination of surface pollution contribution factors is based on land use type, non-point source pollution intensity coefficient, number of days without rainfall in the preceding period, historical water quality monitoring data, and rainfall erosion intensity. Specifically, the catchment area is divided into several surface pollution contribution units, and the land use type, area, runoff coefficient, and pollutant output coefficient of each surface pollution contribution unit are determined. The number of days without rainfall in the preceding period characterizes the degree of surface pollutant accumulation, and the rainfall intensity characterizes the degree of surface pollutant erosion. The results are then calculated for each surface pollution contribution unit. The pollutant concentration at any given time is then represented. A weighted sum is then calculated based on the runoff contribution weights of each surface pollution contribution unit to obtain the final result. Agent of surface pollutant concentration at any time .

[0171] When surface runoff or outfall water quality monitoring points are set up, the surface pollutant concentration proxy The concentrations of total phosphorus, total nitrogen, ammonia nitrogen, chemical oxygen demand, or suspended solids are primarily determined based on data collected from surface runoff or outfall water quality monitoring stations. When no corresponding monitoring stations are established, the surface pollutant concentrations are used as a proxy. It was estimated based on land use type, pollutant output coefficient, runoff coefficient, number of rainless days in the previous period, and historical water quality monitoring data.

[0172] Perform nonnegative matrix decomposition on the water quality sample-index matrix: .

[0173] In the formula, This represents the sample-index matrix. Indicates the number of samples. Indicates the number of indicators. This represents the contribution matrix. This represents the number of source factors. This represents the factor loading matrix. This represents the residual matrix.

[0174] The objective function of the fundamental positive matrix factorization of nonnegative matrix factorization is: .

[0175] In the formula, This represents the objective function value of the fundamental positive matrix factorization. Indicates the sample index. Indicates the index of the indicator. Indicates the first The first sample The observed values ​​of each indicator. This indicates the product of the contribution matrix and the factor loading matrix, corresponding to the first... The first sample The reconstructed values ​​of each indicator. Indicates the first The first sample The uncertainty of each indicator. .

[0176] The source apportionment objective function value after adding internal hydrodynamic constraints and external runoff constraints is: .

[0177] In the formula, This represents the source analysis objective function value after incorporating internal hydrodynamic constraints and external runoff constraints. This represents the weight of the internal constraint regularization. This represents the external constraint regularization weight. This represents the set of time indices used in dynamic pollution source analysis. Indicates time. express The contribution value of the source factor within a given time step. This represents the intrinsic factor index. This represents the scale matching coefficient of the intrinsic constraints. This represents the resuspension flux estimated by the model. express The contribution value of exogenous factors at time step. This indicates the exogenous factor index. This represents the scale matching coefficient of the external constraint.

[0178] To adapt to seasonal changes, seasonal weights based on feature importance can be further introduced: .

[0179] .

[0180] In the formula, Indicates season Next The weight of each indicator. Indicates the index of the indicator. This indicates a seasonal variable. This represents the natural exponential function. Indicates season Next The dimensionless SHAP feature importance value of each indicator. This represents the index of the index in the normalized summation. Indicates the number of indicators. Indicates season Next The dimensionless SHAP feature importance value of each indicator. This represents the objective function after incorporating seasonal weights. Indicates the sample index. Indicates the number of samples. Indicates the first The season of each sample Next The weight of each indicator. Indicates the first The season to which each sample belongs. Indicates the first The first sample The observed values ​​of each indicator. Representing the contribution matrix With factor loading matrix After multiplication, the corresponding number is The first sample The reconstructed values ​​of each indicator. Indicates the first The first sample The uncertainty of each indicator. .

[0181] If a source factor simultaneously satisfies The factor loadings of total phosphorus (TP), iron (Fe), and suspended solids (SS) are higher than the threshold. ,and This is then marked as an internal source of resuscitation power. Indicates the first The correlation coefficient between the contribution sequence of individual source factors and the sediment resuspension flux sequence. Indicates the first The contribution sequence of each source factor in each time sample. This represents the index of the source factor to be determined. Indicates by The resulting sediment resuspension flux sequence. express The resuspension criterion ratio at time. This represents the correlation threshold that is on the same scale as the correlation coefficient.

[0182] If a certain source factor is Highly correlated, rainfall intensity exceeding the threshold, and This is then marked as external surface runoff input. express The external runoff load proxy variable at time.

[0183] The intrinsic and extrinsic factor contribution values ​​obtained from source analysis are normalized to obtain the intrinsic contribution probability. and external contribution probability and All values ​​are taken from .

[0184] .

[0185] .

[0186] In the formula, express The probability of intrinsic contribution at time step. express The probability of external contribution at time t. express Time of the first The contribution value of each source factor. This indicates a positive number used to avoid a denominator of zero.

[0187] S6. Taking the water quality deviation state as the prediction object and the control input of the treatment execution equipment as the optimization variable, the future water quality deviation state under the action of the control input is predicted using a neural network prediction model. Based on the internal contribution probability and the external contribution probability, a predictive control problem with control Lyapunov function decay constraint is constructed. The optimal control input for the current control step is solved and output.

[0188] First, a delayed embedded state is constructed using the water quality deviation state as the prediction object and the control input of the treatment execution equipment as the optimization variable. The delayed embedded state, control input, and disturbance estimate are then input into a neural network prediction model to predict the future water quality deviation state under the action of the control input.

[0189] This embodiment takes into account the large time delay of the water body and the inertia of the equipment, and constructs a delayed embedding state.

[0190] The delayed embedding state is: .

[0191] In the formula, Indicates the first The delayed embedding state vector of the step. Indicates the first The water quality deviation state vector of the step. Indicates the first The water quality deviation state vector of the step. This indicates the state delay order. Indicates the first Step-by-step historical control input. Indicates the first Step-by-step historical control input. This indicates the delay order of the control input. Indicates the first The endogenous phosphorus release flux of the step. Indicates the first External runoff load of the step.

[0192] The prediction model and the disturbance observer are as follows: .

[0193] .

[0194] In the formula, Indicates the first The water quality deviation state vector of the step. The parameter is The neural network prediction model. Indicates the first Step control input. Indicates the first The perturbation estimate of the step. Indicates the first The perturbation estimate of the step. This represents the gain matrix of the perturbation observer. Indicates the first The actual observation output of the step. Indicates the first The model predicts the output step by step.

[0195] The neural network prediction model is implemented using existing nonlinear autoregressive exogenous input neural network prediction models or equivalent feedforward neural network prediction models. This type of model is a commonly used nonlinear dynamic prediction model in water quality prediction, process control, and time series modeling. It can learn the nonlinear mapping relationship between system input and future state by utilizing current state, historical state, historical control variables, and external disturbance information.

[0196] In this method, the neural network prediction model is used to predict the next control step or the water quality deviation state in the time domain. Its inputs include information such as the water quality deviation state, historical control parameters of treatment equipment, endogenous phosphorus release flux, exogenous runoff load, and disturbance estimates. Its output is the future water quality deviation state. After training, this model serves as the state prediction module in the model predictive controller, used to evaluate the impact of different treatment control parameters on the water quality change trend.

[0197] The disturbance estimation is obtained using an existing output error feedback disturbance observer. This disturbance observer does not require rebuilding a complex model, but rather comprehensively estimates the external disturbances, model errors, and time delay effects that are not fully represented by the predicted model, based on the deviation between the actual water quality observation results of the current control step and the neural network prediction results.

[0198] Specifically, in each control step, the system first uses a neural network prediction model to obtain the water quality prediction result for the current or next step, and then compares it with the sensor measured value or the fused value of the continuous water quality field. The deviation between the two is corrected by the disturbance observer to form the current disturbance estimate. This disturbance estimate is then used as the input to the prediction model and the model prediction controller to compensate for the impact of factors such as sudden rainfall changes, external pollution input, equipment execution deviations, and unmodeled hydrodynamic changes on the water quality state prediction.

[0199] Then, a control Lyapunov function is constructed based on the water quality deviation state to constrain the convergence of the water quality deviation state.

[0200] .

[0201] In the formula, Indicates the first The step controls the Lyapunov function value. This represents a positive definite symmetric weight matrix. This indicates that the matrix is ​​positive definite.

[0202] When the probability of intrinsic contribution increases, the control input penalty corresponding to the underlying disturbance device is increased, and the additional bed shear stress caused by the governance execution device is limited by the anti-resuspending shear constraint.

[0203] The control input weight matrix and the anti-resuspending shear constraint are as follows: .

[0204] .

[0205] In the formula, Indicates the first The control input weight matrix for each step. This represents the baseline control weight matrix. This represents the intrinsic penalty gain. This represents the endogenous contribution probability obtained from dynamic pollution source analysis based on positive matrix factorization. This represents a diagonal matrix used to impose additional penalties on the flow generator and aeration layer disturbance equipment. Indicates the first The bed surface shear stress is caused by hydrodynamic forces. Indicates the first Step by control input Additional bed shear stress caused by the corresponding actuator. This represents the safety margin coefficient. Indicates the first The time-varying critical starting shear stress of the step. .

[0206] When the probability of external contribution increases, the dynamic Lyapunov decay rate used to limit the convergence speed of water quality deviation is increased.

[0207] The dynamic Lyapunov decay rate of the current control step is: .

[0208] In the formula, Indicates the first The dynamic Lyapunov decay rate of the step. This represents the truncation function. This represents the baseline attenuation rate. This represents the gain coefficient of the external contribution to the attenuation rate. This represents the external contribution probability obtained from dynamic pollution source analysis based on positive matrix factorization. This represents the inhibition coefficient of the intrinsic contribution on the decay rate. This indicates the lower limit of the decay rate. This indicates the upper limit of the decay rate. and satisfy .

[0209] Finally, a predictive control optimization problem is constructed based on the control Lyapunov function, the control input weight matrix, and the dynamic Lyapunov decay rate. The control Lyapunov function characterizes the convergence degree of water quality deviation in the prediction time domain, the dynamic Lyapunov decay rate limits the decay rate of the control Lyapunov function, and the control input weight matrix limits the penalty intensity of the control inputs for different treatment execution devices.

[0210] The objective function of the predictive control optimization problem is: .

[0211] In the formula, This represents the set of optimization variables. This represents the control step index within the prediction time domain. This indicates the length of the prediction time domain. Indicates the first The water quality deviation state vector of the step. This represents the state deviation weight matrix. Indicates the first Step control input. Indicates the first Step relative to the first The control increment of the step. This represents the control increment weight matrix. This represents the penalty coefficient for slack variables. The first Lyapunov decay constraint acting on the third Slack off variables step by step. This represents the Lyapunov function value for terminal control.

[0212] The constraints include: .

[0213] .

[0214] .

[0215] In the formula, Indicates the first The water quality deviation state vector of the step. The parameter is The neural network prediction model. Indicates the first The delayed embedding state vector of the step. Indicates the first Step control input. Indicates the first The perturbation estimate of the step. Indicates the first The step controls the Lyapunov function value. Indicates the first The step controls the Lyapunov function value. Indicates the first Lyapunov decay rate of the step. The first Lyapunov decay constraint acting on the third Slack variables step by step, and This indicates the lower limit of the control input. This indicates the upper limit of the input limit. Indicates the first Step relative to the first The control increment of the step. This represents the maximum permissible value for the control increment. The upper and lower limits of the control input and the control increment constraints are applied item by item according to each component of the control input.

[0216] The predictive control optimization problem is solved in a rolling manner, and the control input of the current control step obtained by the solution is used as the optimal control input of the current control step.

[0217] Therefore, in scenarios involving sudden external pollution, the system rapidly initiates measures such as interception, oxygenation, and chemical dosing. In scenarios dominated by endogenous sediment release, the system limits strong disturbances in the bottom layer and prioritizes surface microporous aeration, low-shear plug flow, or zoned slow-release treatment.

[0218] In one embodiment, fixed monitoring stations are deployed along the main channel and tributaries of the watershed to be treated. These shore-based stations are equipped with miniature hyperspectral imagers, physicochemical sensor arrays, flow velocity profilers, water level gauges, water thermometers, and rain gauges. The sensors are connected to an edge gateway via RS485, Ethernet, or wireless links. The edge gateway uploads the cleaned observation data to the cloud. The cloud performs rolling calculations on the water quality field, risk probability, internal and external source contribution rates, and control laws, and sends updated network parameter packets, sampling periods, and control commands to the edge gateway.

[0219] When a continuous low flow rate condition persists for an extended period of time, the consolidation state variables Increase, critical shear stress This increases accordingly. When sudden rainfall or gate operation causes shear stress on the bed surface... Suddenly rise and exceed At that time, resuspension flux Rapidly rising. If the PMF source factor simultaneously exhibits high loadings of TP, Fe, and SS and is related to... If the correlation is high, the system determines that the contribution of endogenous resuspension is increased and limits the bottom flow or strong aeration power.

[0220] When rainfall intensity and runoff load are proxy variables rise and At this point, the system determines that external surface runoff input is dominant. In this case, the Lyapunov-MPC system increases the attenuation coefficient. Under the premise of meeting the constraints of equipment power, start-stop frequency and stability, increase the output of interception, oxygenation, chemical dosing or algae control equipment.

[0221] This invention uses uncertainty, exceedance probability, physical equation residuals, and finite volume conservation residuals as sampling feedback quantities. It dynamically adjusts the sampling strategy without relying on real-time measured values, effectively solving the problem of being unable to calculate the true prediction error during unsampled periods, thus improving the relevance and effectiveness of monitoring data. Simultaneously, this invention embeds hydrodynamic constraints, water quality transport constraints, water color constraints, and conservation constraints into the water quality field inversion network, significantly enhancing the physical rationality of water quality inversion results under sparse samples and complex hydrodynamic conditions. This avoids problems such as violations of mass conservation and hydrodynamic laws that are prone to occur in purely data-driven models.

[0222] This invention further introduces a time-varying critical shear stress model coupled with the consolidation state of sediment, transforming the determination of endogenous resuspension from a fixed threshold to a physical criterion that dynamically evolves with environmental factors such as consolidation degree, water content, and biological disturbance, significantly improving the accuracy of endogenous release identification. Furthermore, this invention couples the contribution matrix from positive matrix factorization with the weights of resuspension flux, runoff load, and seasonal characteristics, effectively enhancing the ability to distinguish between endogenous sediment release and exogenous runoff input. Simultaneously, through source apportionment-driven Lyapunov stability constraints and anti-resuspension shear constraints, it reduces control oscillations, equipment wear, and the risk of secondary sediment release in the treatment of large-time-delay water bodies, achieving more stable and efficient water environment regulation.

[0223] The hyperspectral imager in this invention can be replaced by a multispectral camera, UAV remote sensing imagery, or satellite remote sensing imagery. The flow velocity profile can be replaced by a radar flow meter, an acoustic point flow meter, or flow velocity data calculated from the water level-flow relationship. The edge gateway can use ARM, x86, or heterogeneous acceleration hardware. The communication method can be 4G, 5G, LoRa, NB-IoT, Ethernet, or private network communication.

[0224] The Uncertainty Quantification Bayesian Physics-Informed Neural Network (UQ-BPINN) in this invention can output cognitive uncertainty using methods such as variational inference, deep ensemble, Monte Carlo Dropout, or Laplace approximation. Positive Matrix Factorization (PMF) can be replaced with Non-negative Matrix Factorization (NMF), Bayesian Positive Matrix Factorization (Bayesian PMF), or Non-negative Tensor Decomposition. The Lyapunov-based Model Predictive Controller (Lyapunov-MPC) with constraints controlling Lyapunov function decay can be implemented using quadratic programming, sequential quadratic programming, or a neural network approximation solver. These substitutions do not affect the core technical essence of this invention, which focuses on active uncertainty sampling, physical constraint inversion, time-varying sediment consolidation source analysis, and stability prediction control.

[0225] It is understood that the watershed water environment control equipment can be an electronic device with computing power, such as a portable laptop computer, desktop computer, server, smartphone or tablet computer.

[0226] Example 2, please refer to Figure 2 This embodiment provides a watershed water environment control system based on uncertainty and time-varying source analysis. The system includes a multi-dimensional heterogeneous sensing terminal, an edge control layer, a cloud computing center, and governance execution devices.

[0227] The multidimensional heterogeneous sensing terminal is used to collect hyperspectral, water quality physicochemical, hydrodynamic, water level, water temperature, rainfall and equipment operation status data.

[0228] The edge control layer is used to clean, synchronize, cache anomalies, and register spatial coordinates of the data uploaded by the multidimensional heterogeneous sensing terminal, and adjust the sampling cycle of fixed sites or the next sampling position of the mobile sampling platform according to the active sampling strategy.

[0229] The cloud computing center is used to run Bayesian physical constraint neural networks, time-varying critical shear stress models coupled with the consolidation state of sediment, dynamic pollution source analysis models based on positive matrix factorization, and model predictive controllers with control over Lyapunov function decay constraints to execute the above methods.

[0230] The treatment execution device is used to receive control quantities issued by the edge master control layer and perform aeration, algae control, flow propulsion or chemical dosing operations.

[0231] Specifically, the system architecture of this invention adopts an edge-cloud collaborative architecture, including a multi-dimensional heterogeneous sensing terminal, an edge control layer, a cloud computing center, and treatment execution devices. The multi-dimensional heterogeneous sensing terminal is used to acquire hyperspectral, water quality physicochemical, hydrodynamic, water level, water temperature, rainfall, and equipment status data. The edge control layer is used for data cleaning, time synchronization, anomaly caching, active sampling scheduling, and local security control. The cloud computing center is used to run uncertain quantification physical constraint networks, time-varying sediment consolidation models, hydrodynamically constrained PMF source apportionment models, and Lyapunov-MPC controllers. Treatment execution devices include aerators, algae control pumps, flow promoters, dosing valves, or other water body control equipment.

[0232] Figure 2 The system architecture diagram illustrates the data flow, parameter flow, and control flow between the multi-dimensional heterogeneous sensing end, the edge control layer, the cloud computing center, and the governance execution devices. The multi-dimensional heterogeneous sensing end uploads observation data to the edge control layer via RTSP, RS485, Modbus RTU, Ethernet, or other industrial communication methods. The edge control layer interacts with the cloud computing center via a secure MQTT channel, sharing model parameters, risk assessment results, sampling strategies, and control laws, and drives the governance execution devices through PLC, GPIO, or PWM interfaces.

Claims

1. A watershed water environment regulation method based on uncertainty and time-varying source analysis, characterized in that, Include: S1. Collect multi-source monitoring data of the watershed to form a multi-source observation vector; S2. Input the multi-source observation vector into the Bayesian physical constraint neural network to obtain the predicted mean, cognitive uncertainty and probability of exceeding the standard of water quality components, and then combine the finite volume conservation residual to generate an active sampling strategy. S3. Update the observation data according to the active sampling strategy and construct a continuous water quality field that is jointly constrained by hydrodynamic constraints, water quality transport constraints, water color constraints and conservation constraints; S4. Construct a water quality sample-index matrix and water quality deviation state based on the continuous water quality field, and obtain the sediment resuspension flux and endogenous phosphorus release flux based on the near-bottom flow velocity, sediment consolidation state and time-varying critical initiation shear stress. S5. Based on rainfall, runoff and surface pollutant concentration, generate external runoff loads, decompose the water quality sample-index matrix into source contribution matrix and factor load matrix, and combine sediment resuspension flux and external runoff load to correct the source contribution matrix, and obtain the internal contribution probability and external contribution probability. S6. Taking the water quality deviation state as the prediction object and the control input of the treatment execution equipment as the optimization variable, the future water quality deviation state under the action of the control input is predicted using a neural network prediction model. Based on the internal contribution probability and the external contribution probability, a predictive control problem with control Lyapunov function decay constraint is constructed. The optimal control input for the current control step is solved and output.

2. The watershed water environment regulation method based on uncertainty and time-varying source analysis according to claim 1, characterized in that, In S2, the Bayesian physical constraint neural network obtains the predicted mean and cognitive uncertainty of water quality components through multiple forward samplings. ； ； In the formula, Indicates sampling position ,time Predicted mean values ​​of water quality components; Indicates sampling position ,time Cognitive uncertainty at the point of view; Indicates the number of forward samples in Monte Carlo; Indicates the Monte Carlo sampling index; Indicates the first Second sampling parameters The predicted output of the lower network; Indicates the first The parameters are then sampled from the variational posterior. Representing network parameters The variational posterior distribution; The parameters representing the variational posterior distribution; These represent the parameters of a Bayesian physics-constrained neural network. Pollution indicators exceeded the limit The probability of risk is: ； In the formula, Indicators of pollution In position ,time The probability of exceeding the limit; Indicates the first Pollution indicators of each water quality component; Indicators of pollution Standard limits; Indicators of pollution In position ,time The predicted mean; Indicators of pollution In position ,time Cognitive uncertainty; Indicators of pollution In position ,time The random uncertainty; This indicates a positive number used to avoid a denominator of zero; The cumulative distribution function represents the standard normal distribution.

3. The watershed water environment regulation method based on uncertainty and time-varying source analysis according to claim 1, characterized in that, The active sampling strategy in S2 first calculates the active sampling score based on cognitive uncertainty, probability of exceeding the limit risk, physical equation residuals, and finite volume conservation residuals. ； In the formula, Indicates position ,time Active sampling score; This represents the cognitive uncertainty score item, which is the aggregated and normalized cognitive uncertainty of each pollution indicator. This represents the score item for the risk probability of exceeding the standard, which is the aggregated and normalized risk probability of each pollution indicator. This represents the normalized physical equation residuals; This represents the normalized finite volume conservation residual; This represents the energy cost of sampling after normalization; for Weighting coefficients; for Weighting coefficients; for Weighting coefficients; for Weighting coefficients; for Weighting coefficients; The sampling period for fixed data collection stations is: ； In the formula, express The sampling period of a fixed station is determined at all times; Indicates the reference sampling period; This represents the truncation function; Represents the natural exponential function; This represents the sensitivity coefficient of the sampling score to the adjustment of the period; Indicates fixed station location Place Active sampling score at any given moment; Indicates the sampling location of the fixed site; Indicates the lower limit of the sampling period; Indicates the upper limit of the sampling period; The next sampling location for the mobile sampling platform is: ； In the formula, This indicates the next optimal sampling position for the mobile sampling platform; This represents the independent variable that maximizes the objective function; Indicates the sampling location; Represents the sampleable spatial domain; This represents the movement cost weighting coefficient that converts the movement distance to the same scale as the active sampling score; Indicates candidate position With current location The distance between them; This indicates the current location of the mobile sampling platform.

4. The watershed water environment regulation method based on uncertainty and time-varying source analysis according to claim 1, characterized in that, The hydrodynamic constraints are formed by the two-dimensional shallow water equations satisfied by the water depth and flow velocity: ； ； In the formula, Indicates water depth; Indicates time; Represents the velocity vector; Indicates water source and sink items; Represents gravitational acceleration; Indicates water surface elevation; Indicates the eddy viscosity coefficient; Indicates wind stress; Indicates the shear stress on the bed surface; Indicates the density of water; Represents partial differential operators; Represents the gradient operator; Represents the divergence operator; Represents the tensor product; The water transport constraints are formed by the convection-diffusion-reaction equations satisfied by each water component: ； In the formula, Indicates the first The convection-diffusion-reaction equation residuals for various water quality components; Indicates a water quality component index; Indicates the first Network-predicted concentration fields of various water quality components; Indicates the first Equivalent diffusion coefficients of various water quality components; Indicates the first The reaction source and sink of various water quality components; Indicates the sampling location; This represents the set of network-predicted concentration fields for each water quality component. Represents a vector of reaction parameters; The water color constraint is formed by a semi-analytical bio-optical model of the water body and theoretical remote sensing reflectance; the semi-analytical bio-optical model of the water body includes the absorption coefficient and the backscattering coefficient. ； ； In the formula, Indicates wavelength The overall absorption coefficient at the location; Indicates the wavelength of pure water Absorption coefficient at the location; This represents the set of water quality component indices that participated in the hyperspectral inversion; Indicates the first Water quality components at wavelength Specific absorption coefficient at the location; Indicates wavelength The total backscattering coefficient at that location; Indicates the wavelength of pure water Backscattering coefficient at; Indicates the first Water quality components at wavelength The backscattering coefficient at that location; The theoretical remote sensing reflectance is: ； In the formula, Indicates wavelength Theoretical remote sensing reflectance at that location; Represents the first empirical coefficient of the Gordon model; This represents the second empirical coefficient of the Gordon model; Indicates wavelength Backscattering ratio at the location; The conservation constraints are formed by the finite volume conservation residuals of each water quality component on the control volume: ； In the formula, Indicates control body In the The first step of time Finite volume conservation residuals of various water quality components; Indicates the control entity; Indicates control body The area or volume; Indicates the time step index; Indicates the first Time step Water quality components in control system The average concentration within; Indicates the first Time step Water quality components in control system The average concentration within; Indicates the time step; Indicates the control volume boundary surface index; Indicates control body The set of boundary surfaces; Indicates the first Time step Water quality components are controlled through the body boundary surface flux; Indicates the first Time step control of the body's first Sources and sinks of various water quality components.

5. The watershed water environment regulation method based on uncertainty and time-varying source analysis according to claim 1, characterized in that, The sediment resuspension flux and endogenous phosphorus release flux of S4 were determined by a time-varying critical shear stress model coupled with the sediment consolidation state. The bed shear stress is: ； In the formula, express The bed shear stress at any given moment; Indicates the density of water; Indicates the coefficient of friction of the bed surface; express The near-bottom velocity vector at time; express The magnitude of the near-bottom velocity vector at any given moment; Represents gravitational acceleration; Indicates the hydraulic radius; Indicates the friction gradient; The evolution model of the consolidation state variables of the sediment is as follows: ； In the formula, Represents the differential symbol; express The consolidation state variables of the sediment at any given time; Indicates time; This represents the consolidation rate constant; Represents the natural exponential function; This represents the activation energy of the consolidation process; Represents the universal gas constant; express The water temperature at any given time; express When the bed shear stress is less than the static stability threshold, the value is taken as follows: Otherwise, the value is... Indicator functions; This represents the statically stable threshold shear stress. This represents the rate constant of consolidation degradation caused by erosion; express The time-varying critical starting shear stress at a given moment; This represents a function that truncates positive values. This represents the rate constant of consolidation disruption caused by biological disturbance; express The intensity of biological disturbance at any given moment; The time-varying critical starting shear stress is: ； In the formula, Indicates the reference critical shear stress; This represents the volume fraction of clay. express Moisture content or pore water state index at any given time; This represents the exponential coefficient corresponding to the volume fraction of clay. Represents the exponential coefficient corresponding to the consolidation state variable; The exponential coefficient representing the water content or pore water state index; The exponential coefficient representing the intensity of biological disturbance; Real-time re-floating is: ； In the formula, express The resuspension criterion ratio at time; The sediment resuspension flux and endogenous phosphorus release flux are respectively: ； ； In the formula, express Momentary sediment resuspension flux; Indicates the erosion coefficient; Indicates the power of erosion; express Endogenous phosphorus release flux at time t; This indicates the phosphorus content in the sediment; Represents the diffusion release rate constant; This indicates the phosphorus concentration in pore water; This indicates the phosphorus concentration in the overlying water; This represents the regulatory function of dissolved oxygen, pH, and water temperature on phosphorus release; express Dissolved oxygen concentration at any given time; express pH value of the water at any given time.

6. The watershed water environment regulation method based on uncertainty and time-varying source analysis according to claim 1, characterized in that, S5 specifically includes: External runoff loads are generated on behalf of the government based on rainfall, runoff volume, and surface pollutant concentrations. ； In the formula, express The external runoff load proxy variable at time; express The intensity of rainfall at any given moment; express Runoff volume at any given time; express Agents for surface pollutant concentrations at any given time; Perform nonnegative matrix decomposition on the water quality sample-index matrix: ； In the formula, Represents the sample-index matrix; Indicates the number of samples; Indicates the number of indicators; Represents the contribution matrix; Indicates the number of source factors; Represents the factor loading matrix; Represents the residual matrix; The objective function of the fundamental positive matrix factorization of nonnegative matrix factorization is: ； In the formula, This represents the objective function value of the fundamental positive matrix factorization; Indicates the sample index; Indicates the index of indicators; Indicates the first The first sample Observed values ​​of each indicator; This indicates the product of the contribution matrix and the factor loading matrix, corresponding to the first... The first sample The reconstructed values ​​of each indicator; Indicates the first The first sample The uncertainty of each indicator; The source apportionment objective function value after adding internal hydrodynamic constraints and external runoff constraints is: ； In the formula, This represents the source analysis objective function value after incorporating internal hydrodynamic constraints and external runoff constraints. Indicates the weights for regularization of intrinsic constraints; This represents the regularization weight of the external constraint; This represents the set of time indices used in dynamic pollution source analysis; Indicates time; express The contribution value of the source factor within a given time step; Indicates the intrinsic factor index; The scaling coefficient represents the intrinsic constraint. This represents the resuspension flux estimated by the model; express The contribution value of exogenous factors at time step; Indicates the exogenous factor index; The scale matching coefficient represents the external constraint. The contribution values ​​of intrinsic and extrinsic factors obtained from source analysis are normalized to obtain the intrinsic contribution probability and extrinsic contribution probability.

7. A watershed water environment regulation method based on uncertainty and time-varying source analysis according to claim 1, characterized in that, The water quality deviation state vector for S4 is: ； In the formula, Indicates the first The water quality deviation state vector for each step, with positive values ​​indicating either exceeding or failing to meet standards; Indicates the control step index; This indicates the target value for dissolved oxygen safety. Indicates the first Step to actual dissolved oxygen concentration; Indicates the first Actual chlorophyll concentration; Indicates the safety target value for chlorophyll; Indicates the first Actual total phosphorus concentration; This indicates the safety target value for total phosphorus; Indicates the first Step-by-step actual ammonia nitrogen concentration; This indicates the target value for ammonia nitrogen safety. Indicates transpose; S6 includes: using the water quality deviation state as the prediction object and the control input of the treatment execution equipment as the optimization variable, constructing a delayed embedded state, and inputting the delayed embedded state, control input, and disturbance estimate into a neural network prediction model to predict the future water quality deviation state under the action of the control input; The delayed embedding state is: ； In the formula, Indicates the first The delayed embedding state vector of the step; Indicates the first The water quality deviation state vector at each step; Indicates the first The water quality deviation state vector at each step; Indicates the state delay order; Indicates the first Step-by-step historical control input; Indicates the first Step-by-step historical control input; Indicates the delay order of the control input; Indicates the first The endogenous phosphorus release flux of the step; Indicates the first Step-by-step external runoff load; The prediction model and the disturbance observer are as follows: ； ； In the formula, Indicates the first The water quality deviation state vector at each step; The parameter is Neural network prediction models; Indicates the first Step control input; Indicates the first The perturbation estimate of the step; Indicates the first The perturbation estimate of the step; This represents the gain matrix of the perturbation observer; Indicates the first The actual observation output of the step; Indicates the first The model predicts the output step by step.

8. A watershed water environment regulation method based on uncertainty and time-varying source analysis according to claim 7, characterized in that, S6 also includes: Construct a control Lyapunov function to constrain the convergence of the water quality deviation state: ； In the formula, Indicates the first The step controls the Lyapunov function value; This represents a positive definite symmetric weight matrix; Indicates that the matrix is ​​positive definite; When the probability of intrinsic contribution increases, the control input penalty corresponding to the underlying disturbance device is increased, and the additional bed shear stress caused by the governance execution device is limited by the anti-resuspending shear constraint. The control input weight matrix and the anti-resuspending shear constraint are as follows: ； ； In the formula, Indicates the first The control input weight matrix for the step; Represents the baseline control weight matrix; Indicates the intrinsic penalty gain; This represents the endogenous contribution probability obtained from dynamic pollution source analysis based on positive matrix factorization; This represents a diagonal matrix used to impose additional penalties on the flow generator and aeration substrate disturbance equipment; Indicates the first The bed shear stress caused by hydrodynamics; Indicates the first Step by control input Additional bed shear stress caused by the corresponding actuator; Indicates the safety margin coefficient; Indicates the first The time-varying critical starting shear stress of the step; When the probability of external contribution increases, the dynamic Lyapunov decay rate used to limit the convergence speed of water quality deviation is increased. The dynamic Lyapunov decay rate of the current control step is: ； In the formula, Indicates the first The dynamic Lyapunov decay rate of the step; This represents the truncation function; Indicates the reference attenuation rate; This represents the gain coefficient of the external contribution to the attenuation rate; This represents the external contribution probability obtained from dynamic pollution source analysis based on positive matrix factorization; This represents the inhibition coefficient of the intrinsic contribution on the decay rate; Indicates the lower limit of the decay rate; Indicates the upper limit of the decay rate; Based on the control Lyapunov function, the control input weight matrix, and the dynamic Lyapunov decay rate, a predictive control optimization problem is constructed; wherein, the control Lyapunov function is used to characterize the convergence degree of the water quality deviation state in the prediction time domain, the dynamic Lyapunov decay rate is used to limit the decay rate of the control Lyapunov function, and the control input weight matrix is ​​used to limit the penalty intensity of the control input of different governance execution devices; The objective function of the predictive control optimization problem is: ； In the formula, Represents the set of optimization variables; Indicates the control step index within the prediction time domain; Indicates the length of the prediction time domain; Indicates the first The water quality deviation state vector at each step; Represents the state deviation weight matrix; Indicates the first Step control input; Indicates the first Step relative to the first Step control increment; This represents the control increment weight matrix; This represents the penalty coefficient for slack variables; The first Lyapunov decay constraint acting on the third Step slack variables; This indicates the value of the Lyapunov function used to control the terminal. The predictive control optimization problem is solved in a rolling manner, and the control input of the current control step obtained by the solution is used as the optimal control input of the current control step.

9. A watershed water environment regulation method based on uncertainty and time-varying source analysis according to any one of claims 1 to 8, characterized in that, The multi-source monitoring data includes water body hyperspectral remote sensing reflectance, water quality physicochemical parameters, hydrodynamic parameters, water level, water temperature, and rainfall; The multi-source observation vector is: ； In the formula, express Multi-source observation vector at time; Indicates wavelength Place Water body remote sensing reflectance at any given time; express pH value of the water at any given time; express Dissolved oxygen concentration at any given time; express Turbidity at any given moment; express Chlorophyll a concentration at time; express Total phosphorus concentration at time; express Total nitrogen concentration at time; express The velocity or velocity profile at any given moment; express The water level at that moment; express The water temperature at any given time; express Rainfall at any given moment; The water quality sample-index matrix is ​​formed by the water quality physicochemical parameters in the multi-source observation vector and the concentration of water quality components corresponding to the sampling time and space location of the continuous water quality field.

10. A watershed water environment control system based on uncertainty and time-varying source analysis, characterized in that, Include: Multidimensional heterogeneous sensing terminals are used to collect hyperspectral, water quality physicochemical, hydrodynamic, water level, water temperature, rainfall and equipment operating status data; The edge master control layer is used to clean, synchronize, cache anomalies, and register spatial coordinates of the data uploaded by the multi-dimensional heterogeneous sensing terminal, and adjust the sampling cycle of fixed sites or the next sampling position of the mobile sampling platform according to the active sampling strategy. A cloud computing center is used to run a Bayesian physical constraint neural network, a time-varying critical shear stress model coupled with the consolidation state of sediment, a dynamic pollution source analytical model based on positive matrix factorization, and a model predictive controller with control over Lyapunov function decay constraints, to perform the method described in any one of claims 1 to 9. The treatment execution device is used to receive control quantities issued by the edge master control layer and perform aeration, algae control, flow propulsion or chemical dosing operations.

Images