Dynamic evaluation, regulation and control method for small watershed non-point source pollution load

Abstract

The invention discloses a small watershed non-point source pollution load dynamic evaluation and regulation method, relates to the technical field of data processing, and aims to solve the technical problem that a model-based regulation scheme lacks accuracy and effectiveness due to evaluation distortion caused by spatial-temporal variation of model key environment behavior parameters. Comprising the following steps: S1, constructing a digital twinborn body of a small watershed, wherein the digital twinborn body is used for integrating geospatial data, real-time monitoring data and a controllable water conservancy facility state; s2, performing dynamic evaluation based on a digital twinborn body, an operation mechanism and a data dual-driven non-point source pollution process model; and S3, based on the corrected model state and the pollution load control target, generating a self-adaptive regulation and control scheme in the digital twin through a hierarchical decision-making architecture. According to the method, dynamic and high-precision inversion of key pollutant environment behavior parameters is realized by constructing a physically constrained transfer learning model, and the problem of evaluation distortion caused by spatio-temporal variation of model parameters is fundamentally solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method for dynamic assessment and control of non-point source pollution load in small watersheds. Background Technology

[0002] Non-point source pollution, especially nitrogen, phosphorus, and pesticide pollutants from agricultural activities, spreads with runoff and has become a major cause of the deterioration of water quality in small watersheds. Its occurrence is random, widespread, and delayed, making load assessment and effective regulation a persistent challenge in environmental management.

[0003] Currently, technological development in this field mainly relies on distributed hydrological and water quality models based on physical mechanisms (such as SWAT and HSPF) for simulation and evaluation, combined with engineering and non-engineering measures for regulation and planning. However, in practical applications, especially for specific pesticides with strong adsorption and complex environmental behavior (such as triazoles and sulfonylureas), the existing technology system has revealed a long-standing core bottleneck that limits its accuracy: the key environmental behavior parameters of the model (such as the adsorption-desorption coefficient and degradation rate of pollutants) have high spatiotemporal variability. Fixed parameters measured under laboratory conditions are difficult to accurately reflect their dynamic changes in different fields, soil moisture, temperature, pH, and microbial environments. This fundamental contradiction between "static parameters" and "dynamic environment" leads to inherent biases in the model when simulating pollutant migration and transformation processes, resulting in insufficient reliability of load assessment results.

[0004] Designing control schemes based on this, whether it's site selection for engineering facilities or operation scheduling, is like planning a path on a "distorted map," significantly reducing its scientific rigor and effectiveness. Therefore, how to achieve dynamic and adaptive identification of key model parameters in the real environment has become a crucial scientific and technological problem that must be overcome to improve the accuracy of non-point source pollution assessment and efficient control capabilities in small watersheds. Existing technologies either rely on frequent and costly large-scale field sampling and laboratory analysis to update parameters, lacking timeliness; or they rely entirely on historical data for data-driven modeling, lacking physical mechanism constraints, resulting in weak model extrapolation capabilities and poor interpretability, neither of which fundamentally solves the aforementioned problems. In view of this, we propose a method for dynamic assessment and control of non-point source pollution loads in small watersheds. Summary of the Invention

[0005] The purpose of this invention is to provide a method for dynamic assessment and control of non-point source pollution load in small watersheds, in order to solve the technical problem that the spatiotemporal variation of key environmental behavior parameters of the model leads to assessment distortion, which in turn makes the model-based control scheme lack accuracy and effectiveness.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for dynamic assessment and control of non-point source pollution load in small watersheds, comprising the following steps: S1: Construct a digital twin of a small watershed, wherein the digital twin is used to integrate geospatial data, real-time monitoring data and controllable water conservancy facility status; S2: Based on the aforementioned digital twin, a dynamic evaluation is performed using a non-point source pollution process model driven by both operational mechanisms and data, which includes: Based on online monitoring time-series data and laboratory simulation measurement data, the environmental behavior parameters of key pollutants are dynamically inverted through a physically constrained transfer learning model. Data assimilation technology is used to assimilate real-time monitoring data into the non-point source pollution process model in order to continuously correct the model state; S3: Based on the corrected model state and pollution load control target, an adaptive control scheme is generated in the digital twin through a hierarchical decision architecture. The hierarchical decision architecture includes: a strategic layer based on scenario simulation for long-term strategy selection, a tactical layer based on rolling optimization for medium-term instruction generation, and an execution layer based on real-time feedback for instantaneous adjustment. S4: The adaptive control scheme is sent to the corresponding controllable water conservancy facilities for execution, and the execution effect data is fed back to the digital twin to form a closed loop of evaluation and control.

[0007] This invention achieves dynamic and high-precision inversion of environmental behavior parameters of key pollutants by constructing a physically constrained transfer learning model. This fundamentally solves the problem of assessment distortion caused by spatiotemporal variations in model parameters, significantly improving the mechanistic rationality and predictive accuracy of dynamic assessments. This invention combines precisely measured data obtained under controlled laboratory conditions (as anchor points or truth labels) with time-series data from online field monitoring to train a transfer learning model incorporating physical constraints (such as the law of conservation of mass). This model not only learns the complex mapping relationship from online monitoring data (such as concentration, pH, and water temperature sequences) to environmental behavior parameters (such as dynamic adsorption coefficients), but also enforces that the inverted parameters must satisfy basic physicochemical process equations through a physical constraint penalty term. This makes the inverted parameters no longer simple mathematical fitting results, but rather living parameters with clear physical meaning that can evolve in real time with environmental conditions. This method effectively integrates the reliability of mechanistic models with the flexibility of data-driven approaches, overcoming overfitting or counterintuitive predictions under data sparsity or noise. It enables the non-point source pollution process model in the digital twin to have an adaptive kernel, resulting in a qualitative leap in the physical consistency and spatiotemporal accuracy of the evaluation results, laying a solid and reliable data foundation for subsequent precise regulation.

[0008] Preferably, in step S1, constructing the digital twin specifically includes: Access to multi-source heterogeneous data, including geospatial data, meteorological data, online monitoring sensor data, and facility operation status data; The multi-source heterogeneous data is standardized, fused, and spatiotemporally aligned to form a unified spatiotemporal dataset that drives the digital twin; The online monitoring sensor data includes at least flow rate, water level, turbidity, pH, and specific pollutant concentration indicators from key cross-sections.

[0009] Preferably, the environmental behavior parameters of the key pollutants dynamically retrieved are specifically: The target label is the laboratory simulation measurement data obtained periodically, and the input feature is the online monitoring time series data of the corresponding time period. The transfer learning model for the physical constraints is trained, and a dynamic mapping relationship from online monitoring data to environmental behavior parameters is established by minimizing the joint loss function; the joint loss function is: ; In the formula, The task loss function measures the model's prediction results. With real labels The differences between them; This is a physical constraint penalty function; The input feature vector; and These are weight hyperparameters; Using the trained model, dynamically updated environmental behavior parameters are calculated and output based on real-time monitoring data.

[0010] Preferably, the physical constraint penalty term Based on the law of conservation of mass, its calculation formula is as follows: ; In the formula, For the concentration scalar field of pollutants; For time; This represents the water flow velocity vector field. For divergence operators; For predicted environmental behavior parameters Other input features The pollutant source and sink terms are jointly determined; this constraint forces the inversion parameters to satisfy the convection-diffusion reaction equation framework, ensuring its physical consistency.

[0011] Preferably, in step S2, the mechanism- and data-driven non-point source pollution process model is a multi-scale coupled model, including: Microscale models used to simulate the microscopic migration and transformation processes of pollutants; Land parcel and gully scale models used to simulate land parcel runoff, soil erosion, pollutant transport with sediment, and biochemical processes within gullies; A small-watershed-scale distributed hydrological and water quality model used to simulate the spatial transport and accumulation of pollution loads from the source area to the receiving water body; Among them, models of different scales are coupled through the transmission and feedback of key state variables.

[0012] Preferably, the continuous correction of the model state using data assimilation technology specifically employs an ensemble Kalman filter algorithm, the model state update formula of which is: ; In the formula, The state vector represents the predicted state from the model; This represents the optimal estimated state vector obtained after data assimilation analysis. Represents the observation vector; For observation operators; This is the Kalman gain matrix; this step improves short-term forecast accuracy by fusing real-time observation data and dynamically calibrating the model trajectory.

[0013] Preferably, in step S3, the rolling optimization-based instruction generation of the tactical layer employs a model predictive control algorithm, which solves the following optimization problem in each control cycle: ; Constraints: , , ; In the formula, This represents the objective function that needs to be minimized. For discrete time step index; The sequence of control instructions to be optimized; To control the time domain; The future predicted based on the model state corrected by the data assimilation technique The state variables of the step; For the desired pollution load or water quality status target; and It is a diagonal weight matrix; The state transition function represents the mechanism- and data-driven area source pollution process model after data assimilation and correction. Input for predicted future disturbances; and These represent the upper and lower limits of the amplitude of the control command, respectively. and These represent the upper and lower limits of the rate of change of the control command, respectively; the algorithm continuously solves for the optimal control command over a future period.

[0014] Preferably, in step S3, the real-time feedback-based fine-tuning of the execution layer is implemented using a reinforcement learning agent, which learns the fine-tuning strategy by maximizing cumulative rewards, and the action-value function... The update follows the Bellman optimality equation: ; In the formula, For a moment Environmental state, state Including the model predictive control algorithm in time Predicted state Compared with the actual observation state Deviations, and the real-time status of controllable facilities; For a moment The actions taken by the agent are the baseline control commands output by the model predictive control algorithm. Fine-tuning amount; Indicates the state Next action The immediate reward signal received afterward; Discount factor; The mathematical expectation operator is used to enable the intelligent agent to adaptively compensate for model errors and unknown disturbances online.

[0015] Preferably, in step S4, the timing control instructions in the adaptive control scheme are specific parameters for the opening and closing timing, opening degree, or operating water level of ponds, dams, gates, or adjustable interception dams; when the control object is multiple interception dams arranged in a distributed manner, the scheme is a set of coordinated control instructions generated according to the spatiotemporal distribution of pollution load and coordinated with each other in terms of water storage and discharge timing.

[0016] A dynamic assessment and control system for non-point source pollution load in a small watershed includes: The data sensing and acquisition module is used to acquire multi-source data required to build and drive the digital twin; The digital twin construction and maintenance module is used to integrate the multi-source data and establish and maintain a virtual mapping model of the small watershed; The dynamic evaluation model engine incorporates a non-point source pollution process model driven by both the mechanism and data, a transfer learning parameter inversion unit, and a data assimilation unit. The intelligent control and decision-making module, which incorporates the hierarchical decision-making architecture, is used to generate adaptive control schemes. The control command issuance and execution feedback module is used to convert the control scheme into control commands and issue them to the field facilities, while collecting execution status data and feeding it back to the system. The dynamic evaluation model engine and the intelligent control and decision-making module operate in the digital twin environment, forming a closed loop.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves dynamic and high-precision inversion of environmental behavior parameters of key pollutants by constructing a physically constrained transfer learning model. This fundamentally solves the problem of assessment distortion caused by spatiotemporal variations in model parameters, significantly improving the mechanistic rationality and predictive accuracy of dynamic assessments. This invention combines precisely measured data obtained under controlled laboratory conditions (as anchor points or truth labels) with time-series data from online field monitoring to train a transfer learning model incorporating physical constraints (such as the law of conservation of mass). This model not only learns the complex mapping relationship from online monitoring data (such as concentration, pH, and water temperature sequences) to environmental behavior parameters (such as dynamic adsorption coefficients), but also enforces that the inverted parameters must satisfy basic physicochemical process equations through physical constraint penalty terms. This makes the inverted parameters no longer simple mathematical fitting results, but rather living parameters with clear physical meaning that can evolve in real time with environmental conditions. This method effectively integrates the reliability of mechanistic models with the flexibility of data-driven approaches, overcoming overfitting or counterintuitive predictions under data sparsity or noise. It enables the non-point source pollution process model in the digital twin to have an adaptive kernel, resulting in a qualitative leap in the physical consistency and spatiotemporal accuracy of the evaluation results, laying a solid and reliable data foundation for subsequent precise regulation.

[0018] 2. This invention also achieves seamless connection and dynamic optimization from precise assessment to scientific regulation through the deep integration of data assimilation technology and hierarchical intelligent decision-making architecture, solving the problem of how to efficiently transform assessment results into executable and adaptive control strategies. This invention constructs a closed-loop technology chain of assessment-decision-execution. First, through data assimilation technology (such as ensemble Kalman filtering), real-time monitoring data is continuously injected and the state of the dynamic assessment model is corrected, ensuring that the digital twin remains synchronized with the real watershed, guaranteeing the timeliness and accuracy of the decision-making basis. Furthermore, a three-layer decision-making architecture of strategy-tactics-execution is designed: the strategy layer selects robust strategies based on multi-scenario simulation; the tactical layer uses model predictive control (MPC) for rolling optimization within a finite time domain, unifying the solution of pollution control objectives and engineering constraints (such as equipment operating limits) to generate a forward-looking and precise sequence of operating instructions; the execution layer utilizes a reinforcement learning agent to fine-tune the instructions online based on real-time feedback to address disturbances not covered by the model. This hierarchical architecture organically combines long-term planning, medium-term optimization, and instantaneous response, enabling the generated control schemes to be scientific, forward-looking, executable, and robust, truly achieving intelligent transformation and dynamic adaptation from assessment to control.

[0019] 3. This invention also constructs a closed-loop feedback, continuously evolving digital twin system for small watersheds, achieving knowledge accumulation and system self-evolution throughout the entire assessment and control process. This addresses the limitations of single-event decision-making and the system's insufficient long-term adaptability. This invention places dynamic assessment, intelligent decision-making, and physical facility execution within a unified digital twin environment, forming a complete closed loop of perception-analysis-decision-execution-learning. The execution effect data of each control action is collected in real time and fed back to the digital twin for model verification and correction, enabling continuous system calibration. More importantly, based on historical simulation data and successful / failed control cases, the system can automatically construct and update a knowledge graph of control strategies and a digital contingency plan library, transforming practical experience into reusable structured knowledge. This allows the system not only to handle current scenarios but also to quickly provide experience references and assist decision-making when facing similar or new scenarios through case analogy and reasoning. Furthermore, products derived from the system, such as electronic fences for high-risk areas, can directly link with agricultural machinery to achieve source avoidance. This closed-loop and learning mechanism endows the entire method with the ability to self-evolve, accumulate knowledge, and manage long-term adaptation to complex and ever-changing environments, elevating it from a one-off technical tool to a sustainable intelligent decision support system. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the construction process of the digital twin of this invention. Figure 2 This is a flowchart of the dynamic evaluation process of the present invention; Figure 3 This is a flowchart of the intelligent control and decision-making process of the present invention; Figure 4 This is a flowchart illustrating the execution feedback closed-loop process of the present invention. Detailed Implementation

[0021] Example 1: As Figures 1 to 4 As shown, the present invention relates to a method for dynamic assessment and control of non-point source pollution load in a small watershed, comprising the following steps: S1: Construct a digital twin of a small watershed, which integrates geographic information, real-time hydrological and water quality monitoring data, and the status of controllable water conservancy facilities; In an embodiment of the present invention, step S1, the step of constructing the digital twin, specifically includes: Access to multi-source heterogeneous data, including geospatial data, meteorological forecast data, real-time sensor data uploaded by online monitoring equipment, manual inspection data, and operational status data of the controllable water conservancy facilities; The multi-source heterogeneous data is standardized, fused, and spatiotemporally aligned to form a unified spatiotemporal dataset that drives the digital twin.

[0022] The online monitoring equipment includes water quality multi-parameter sensors, hydrological sensors, and microfluidic chip-type rapid pollutant detectors deployed at key hydrological sections and pollution source sink nodes; the sensor data includes at least flow rate, water level, turbidity, pH, oxidation-reduction potential, ultraviolet absorbance, and specific pollutant concentration indicators.

[0023] S2: Based on the digital twin, a non-point source pollution process model driven by both operating mechanism and data. The non-point source pollution process model uses online monitoring time series data and parallel laboratory simulation measurement data to dynamically invert the environmental behavior parameters of key pollutants through the built-in transfer learning module, and continuously corrects the model state using data assimilation technology. In an embodiment of the present invention, the laboratory simulation measurement data is obtained by conducting simulation experiments under controlled conditions based on the collection of on-site environmental media samples in a small watershed using a mobile experimental platform. The laboratory simulation measurement data includes at least the pollutant adsorption-desorption kinetic parameters, degradation rate constant, and soil erodibility parameters under the current environmental conditions.

[0024] In an embodiment of the present invention, the dynamic inversion of key pollutant environmental behavior parameters through the built-in transfer learning module specifically includes: The laboratory simulation measurement data acquired periodically is used as the target label, and the online monitoring time series data within the corresponding time period is used as the input feature; Train a physically constrained deep transfer learning model to establish a dynamic mapping relationship from the online monitoring time series data to the environmental behavior parameters of the key pollutants; Using a trained deep transfer learning model, based on the real-time acquired online monitoring time series data, dynamically updated environmental behavior parameters are calculated and output in real time, which are then provided to the non-point source pollution process model for use. Specifically, the deep transfer learning model for the physical constraints minimizes the following loss function. Conduct training: ; In the formula: The task loss function measures the model's prediction results. With real labels The differences between them, such as mean squared error; Key environmental behavior parameters of pollutants (such as dynamic adsorption coefficient) predicted by deep transfer learning models. This provides a true label for environmental behavior parameters obtained through laboratory simulation measurements; The physical constraint penalty function quantifies the prediction results. and input features The degree of violation of fundamental physical laws, and Weighted summation can only be performed after normalization or the introduction of dimensional coefficients to ensure they are of the same dimension. The input feature vector is the corresponding online monitoring time series data (such as concentration, pH, and temperature sequences). and The hyperparameters are non-negative and are used to adjust the weights of the task loss term and the physical constraint term in the total loss, respectively. They can be determined by grid search combined with the inversion accuracy on the validation set, and their search range is typically [value missing]. And satisfy ; Explanation of the operational logic: This formula represents the training objective (loss function) of a physically constrained deep transfer learning model. Its core logic is joint optimization: minimizing the model's predicted value ( ) and laboratory measured labels ( Error between (task loss) At the same time, it penalizes predictions that violate pre-set physical laws (such as the law of conservation of mass) (physical constraint penalty term). Hyperparameters and This is used to weigh the relative importance of data fitting accuracy and physical consistency. The training process adjusts the model's internal parameters to optimize the total loss. Minimization is achieved to obtain a parameter inversion model that is both faithful to the observed data and conforms to the physical mechanisms. This formulaic training objective embeds prior physical knowledge into the data-driven learning process as computable constraints, effectively overcoming the "contrary to common sense" prediction problems that may occur in purely data-driven models under data sparsity or noise interference. It guides the model to learn patterns not only by relying on sample matching but also by being rooted in physical principles, thereby significantly improving the physical rationality and generalization ability of inversion parameters in complex and variable natural environments, making the mechanistic basis of dynamic evaluation more reliable.

[0025] The deep transfer learning model for physical constraints incorporates physical constraint terms based on the laws of mass conservation, thermodynamics, or chemical reaction kinetics into its loss function. Specifically, the physical constraint penalty term This can be specifically expressed as a constraint on the law of conservation of mass, and its calculation formula is as follows: ; In the formula: This is a physical constraint penalty term based on the law of conservation of mass; For the concentration scalar field of pollutants; For time; This represents the water flow velocity vector field. To be based on the concentration scalar field and water flow velocity vector field The calculated local rate of change, , and All have the dimension of concentration / time; It is a divergence operator used to calculate the net outflow flux of a vector field at a point; For predicted environmental behavior parameters Other input features The pollutant source and sink terms are jointly determined, and this constraint forces the model to learn parameters that predict the results in accordance with the convection-diffusion reaction equation framework; Explanation of the operational logic: This formula represents the physical constraint penalty term. One specific implementation. Its operational logic is mandatory consistency: calculating the parameters derived from the model inversion. Other inputs The determined pollutant source and sink items And check its consistency with the concentration scalar field. and water flow velocity vector field Calculated local rate of change ( ) and convective-diffusion flux divergence ( Whether the mass conservation equation is satisfied. The value of the formula (the square of the norm) is the violation of the mass conservation law. The larger the value, the heavier the penalty, thus forcing the model to inversely derive parameters during training. The entire system's mass transfer and transformation process must be self-consistent. This specific constraint transforms high-order physical principles (mass conservation) into differentiable computational terms that can directly participate in model training. It ensures that the environmental behavior parameters dynamically derived through transfer learning are not isolated data fitting results, but rather key components that can be embedded into the complete physical framework of pollutant transport and transformation, maintaining its self-consistency. This fundamentally enhances the inherent physical consistency and credibility of the dynamic evaluation model, enabling subsequent model-state-based predictions and control decisions to be grounded in a solid physical foundation, exhibiting stronger stability, especially in highly dynamic and strongly nonlinear processes.

[0026] The deep transfer learning model for the physical constraints employs a three-layer fully connected neural network. The number of nodes in the input layer is... ,correspond The system monitors several online features (such as the mean and standard deviation of tebuconazole concentration, pH, water temperature, and turbidity over the past 24 hours); the two hidden layers have 128 and 64 nodes respectively, using the ReLU activation function; the output layer has [number of nodes missing]. ,correspond Environmental behavior parameters to be retrieved (such as adsorption coefficient) Degradation rate constant ).

[0027] The training data was prepared as follows: taking the date of each laboratory measurement as the center, online monitoring data for a total of 7 days before and after were collected, and their statistical characteristics were calculated as input features. Laboratory measurements were used as target labels. 80% of all samples were used as the training set, 10% as the validation set, and 10% as the test set.

[0028] The concentration field in the physical constraint penalty term and flow velocity field During the training phase, the simulated output of the mechanism model under historical scenarios can be used as an approximation for calculation; during the real-time application phase, the model prediction field corrected by data assimilation technology is used.

[0029] In an embodiment of the present invention, the mechanism- and data-driven non-point source pollution process model is a multi-scale coupled model, including: Microscale models are used to simulate the microscopic migration and transformation processes of pollutants in soil pores or sediment interfaces. Plot and channel-scale models are used to simulate runoff generation, soil erosion, pollutant transport with sediment, and biochemical processes in channel water. Small watershed-scale distributed hydrological and water quality models are used to simulate the spatial transport and accumulation of pollution loads from the source area to the receiving water body. Among them, models of different scales are coupled through the transmission and feedback of key state variables.

[0030] The multi-scale coupled model operates using an asynchronous coupling method of 'top-down driving and bottom-up feedback'. Specifically: The small watershed-scale distributed model provides the hydrological response (runoff, erosion) and initial pollutant concentrations for each computational grid, driving the downstream plot and channel-scale models.

[0031] The plot and channel-scale model receives upstream water and sediment conditions, simulates more detailed pollutant transport with sediment and endogenous biochemical processes in the channel, and feeds back the calculated pollutant form ratios (such as dissolved and particulate adsorbed forms) at the channel outlet to the small watershed-scale model to correct the pollutant distribution coefficients in its confluence calculation.

[0032] The microscale model does not run in real time; its pre-calibrated parameterized relationships (such as adsorption isotherm equations) are directly embedded into the pollutant migration module of the plot and channel scale model.

[0033] The models at each scale exchange the aforementioned key state variables at each coupled time step (e.g., 1 hour) via a pre-defined text file or memory array.

[0034] Specifically, the data assimilation technique is used to continuously correct the model state. This involves using ensemble Kalman filtering, variational assimilation, or a hybrid algorithm thereof to assimilate the sensor data uploaded in real time by the online monitoring equipment into the plot and ditch scale model and the small watershed scale distributed hydrological and water quality model, so as to correct the model's predicted state variables and some key parameters. Specifically, when using the ensemble Kalman filter algorithm, the update of the model state follows the formula: ; In the formula: The state vector represents the model's prediction and contains all the key variables predicted by the multi-scale coupled model (such as pollutant concentration, water volume, and sediment volume in each unit). This represents the optimal estimated state vector (corrected state) obtained after data assimilation analysis. The representative observation vector consists of actual measurement data directly acquired by online monitoring equipment. For observation operators, a matrix or function maps the model state space to the observation space. The concentration values ​​at the corresponding monitoring points are extracted from the state vector containing the concentrations of all grid points across the entire watershed. The Kalman gain matrix is ​​the optimal weight matrix under the combined effect of the model prediction error covariance and the observation error covariance. Explanation of the operational logic: This formula describes the core analysis steps of the Ensemble Kalman Filter (EnKF). Its operational logic is optimal linear correction: adjusting the predicted state of the model... Compared with actual observations from online monitoring By comparing, the observation increment is obtained ( Then, using the Kalman gain matrix... (This matrix is ​​determined by both the uncertainty of the model forecast and the observation error.) This increment is weighted, and the weighted correction is added to the forecast state to obtain the analytical state. . The calculations determine whether to place more trust in model forecasts or actual observations, and aim to refine the analytical state. The estimation error covariance is minimized. This assimilation update formula is the mathematical core for achieving real-time fusion of the model and data. Through a Bayesian inference framework, it dynamically and quantitatively utilizes real-time observation data to correct the model's predicted trajectory and key states, effectively overcoming forecast drift caused by initial model condition errors, parameter uncertainties, and input data errors. This enables the digital twin to continuously maintain "synchronization" with the real physical watershed, greatly improving the real-time accuracy of dynamic assessments and the reliability of short-term forecasts, providing high-quality, high-confidence input for subsequent precise control based on model states.

[0035] S3: Based on the corrected model state and the preset pollution load control target, multi-scenario simulation is performed in the digital twin, and an adaptive control scheme containing time-series control instructions is generated through a hierarchical decision architecture. The hierarchical decision architecture includes: a strategic layer based on scenario simulation for long-term strategy selection, a tactical layer based on rolling optimization for mid-term instruction generation, and an execution layer based on real-time feedback for instantaneous adjustment. In an embodiment of the present invention, step S3, the strategic strategy selection based on scenario simulation at the strategic layer specifically includes: loading multiple representative historical or predicted meteorological scenarios into the digital twin, simulating the pollution load reduction effect and engineering operation cost under different combinations of control strategies for each scenario, and selecting the benchmark control strategy set with optimal robustness based on a multi-objective evaluation algorithm.

[0036] In an embodiment of the present invention, step S3, the generation of tactical instructions based on rolling optimization of the tactical layer specifically includes: based on short-term high-precision weather forecasts and real-time monitoring data, guided by the benchmark control strategy, using a model predictive control algorithm, with the goal of achieving the optimal pollution load control effect within a future control cycle, rollingly solving and outputting a detailed sequence of operation instructions for the controllable water conservancy facilities for the next few hours to days; The model predictive control algorithm solves the following optimization problem in each control cycle to obtain the optimal sequence of operating instructions. : ; Constraints: System dynamic model: ; Control amplitude constraints: ; Control quantity change rate constraint: ; In the formula: This represents the objective function (cost function) that needs to be minimized. For discrete time step index; For the control command sequence to be optimized, such as the gate opening sequence, , To control the time domain; The future predicted based on the model state corrected by the data assimilation technique The state variables of the step (prediction time domain); For the desired pollution load or water quality status target; and Let be a diagonal weight matrix, representing the penalty weights for the state tracking error and the drastic changes in the control input, respectively. and The state variables and control variables need to be normalized or weighted according to their actual dimensions to ensure the objective function. Since the two quantities have the same dimensions, they can be weighted and summed. They can be first initialized by diagonal normalization based on the magnitudes of the state variables and control variables, and then tuned by trial and error or automatic parameter tuning algorithms (such as Bayesian optimization) so that the system can achieve the desired control performance in the simulation. The state transition function represents the mechanism- and data-driven area source pollution process model after data assimilation and correction. For predicted future disturbance inputs, such as rainfall forecasts; and These represent the upper and lower limits of the amplitude of the control command, respectively. and These represent the upper and lower limits of the rate of change of the control command, respectively. Explanation of the operational logic: This formula defines a rolling optimization problem for Model Predictive Control (MPC). Its operational logic is multi-step look-ahead and constrained optimization: In each control cycle, starting from the current model state after data assimilation and correction, the solution is obtained for a future prediction time domain (…). A series of control commands within the step Optimization Objective It consists of two parts: the first part requires the model to predict the state trajectory. As close as possible to the desired target state (e.g., water quality standards); Part Two requires changes to control instructions. The optimization should be as gradual as possible to ensure the stable operation of the engineering equipment. The entire optimization process is implemented within a dynamic model. (Representing the modified pollution process model) and various engineering safety constraints are considered. After obtaining the optimal control sequence, only the first control command is implemented, and this process is repeated in the next cycle. This optimization problem formally defines the mathematical connotation of "precise regulation." It not only considers instantaneous optimality but also performs multi-step simulation and optimization based on the model, demonstrating foresight. Simultaneously, it unifies the pollution load control target (state tracking term) with the safety and stability requirements of engineering operation (control cost term) within a unified framework, and incorporates all physical and equipment constraints into the solution process. This ensures that the generated regulation scheme is not only environmentally effective but also engineering-feasible and robust, achieving a scientific, automatic, and dynamic transformation from environmental goals to specific executable operational commands, significantly improving the systematicness and precision of regulation.

[0037] In an embodiment of the present invention, step S3, the execution fine-tuning based on real-time feedback of the execution layer specifically includes: using a reinforcement learning agent to adaptively adjust the detailed operation instruction sequence online based on the deviation between the actual and expected operating states of the controllable water conservancy facilities and the real-time monitored changes in pollution load, in order to handle model prediction errors and unforeseen disturbances; Specifically, the goal of the reinforcement learning agent is to learn a policy. To maximize cumulative rewards, its action value function The update follows the Bellman equation: ; Among them, actions The reference control command output by the model predictive control algorithm Fine-tuning amount; instant rewards It is negatively correlated with the effectiveness of pollution load control and operational stability. This is the discount factor.

[0038] In the formula: For a moment Environmental state, state Including the model predictive control algorithm in time Predicted state Compared with the actual observation state Deviations, and the real-time status of controllable facilities; For a moment The actions taken by the agent are the baseline control commands output by the model predictive control algorithm. Fine-tuning amount; Indicates the state Next action The immediate reward signals obtained afterward are typically designed to be positively correlated with pollution control effectiveness and negatively correlated with excessive or frequent manipulation. Designed as a dimensionless, instantaneous evaluation signal, or after normalization, and compared with the action-value function. With consistent dimensions, the immediate reward function for reinforcement learning is designed as follows: ,in This represents the reduction in pollution load concentration relative to the target value. To control the adjustment range of the command, and As a weight, used to balance environmental benefits and operating costs; Discount factor ( This is used to weigh the importance of immediate rewards versus future rewards. For mathematical expectation operators; Explanation of the operational logic: This formula is the Bellman optimal equation, which describes how a reinforcement learning agent evaluates a state. Take action below long-term value ( (Value). Its operational logic is temporal difference and value iteration: current state-action pair. value( (), equal to the immediate reward After discounting ( (as a discount factor), in the next state The maximum possible value among all possible actions The expected value. The agent updates the value through continuous interaction with the environment (trying actions, observing new states and rewards). Value table or The value network eventually learns a strategy that maximizes long-term cumulative rewards. In this scheme, the agent learns how to fine-tune the baseline instructions of MPC. This equation forms the theoretical basis for the adaptive fine-tuning of the reinforcement learning agent. Through online learning, the agent can handle prediction errors caused by simplified or unmodeled dynamics in the MPC model, as well as sudden, unforeseen disturbances. It endows the system with real-time learning and adaptive capabilities at the execution layer, enabling it to adapt to the latest environmental feedback (rewards). The system dynamically adjusts its fine-tuning strategy to compensate for the shortcomings of the upper-level optimization model. This hierarchical architecture of "model optimization + learning adaptation" significantly enhances the robustness and flexibility of the entire control system in the face of uncertainty, nonlinearity, and time-varying characteristics, making the control behavior more intelligent and closely aligned with actual dynamics.

[0039] S4: The adaptive control scheme is sent to the corresponding controllable water conservancy facilities for execution, and the execution effect data is fed back to the digital twin to form a closed loop of dynamic evaluation and control.

[0040] In an embodiment of the present invention, the timing control command in the adaptive control scheme is a specific control parameter generated for the opening and closing timing, opening degree, or operating water level of a pond, dam, gate, ecological filter, or adjustable interception dam.

[0041] When the controllable water conservancy facilities are multiple interception dams arranged in a distributed manner, the adaptive control scheme is a collaborative control scheme. Based on the spatiotemporal distribution of pollution load simulated by the non-point source pollution process model, it generates a set of differentiated control instructions for the upstream, midstream, and downstream interception dams to cooperate with each other in terms of water storage, discharge, and sediment discharge timing.

[0042] The method further includes: in the digital twin, based on historical simulation data and successful control cases, constructing and updating a control strategy knowledge graph and / or digital contingency plan library to assist the decision-making process of the hierarchical decision-making architecture.

[0043] The knowledge graph of the regulation strategy is stored in a graph structure. Nodes represent 'meteorological scenarios', 'soil moisture', 'regulation facility status', 'implementation actions', and 'effect evaluation', while edges represent causal or temporal relationships between nodes. Its construction method is as follows: after each successful regulation case, the key feature vectors of this event (such as rainfall characteristics, previous pollution load, the sequence of strategy instructions used, and the final reduction rate) are automatically extracted as a new knowledge record, and the graph is updated using a graph embedding algorithm.

[0044] The digital contingency plan database is a relational database. Each contingency plan contains the following fields: scenario number, meteorological condition feature code, recommended control strategy ID, expected effect index, and confidence level. The generation method is as follows: cluster analysis is performed on historical scenario simulation results to group similar scenarios into one category, and the control strategy with the best effect in that category is used as the recommended contingency plan for that category of scenarios.

[0045] The method further includes: dynamically generating a spatial distribution map of high-risk areas of non-point source pollution within a small watershed based on the simulation results of the digital twin, and linking the high-risk areas with the agricultural machinery intelligent scheduling platform in the form of electronic fences to guide agricultural activities to avoid pollution.

[0046] Example 2: A dynamic assessment and control system for non-point source pollution load in a small watershed, comprising: The data sensing and acquisition module is used to acquire multi-source data required to build and drive the digital twin; In another embodiment of the present invention, the data sensing and acquisition module includes an online monitoring sensor network deployed in the watershed, a mobile experimental platform, and a data interface for receiving meteorological and remote sensing data; the control command issuance and execution feedback module includes an edge computing gateway and a field programmable logic controller.

[0047] The digital twin construction and maintenance module is used to integrate the multi-source data and establish and maintain a virtual mapping model of the small watershed; The dynamic assessment model engine has a built-in mechanism- and data-driven non-point source pollution process model, a transfer learning parameter inversion unit, and a data assimilation unit, which are used to perform dynamic simulation and assessment of pollution load. The intelligent control and decision-making module has a built-in hierarchical decision-making architecture for generating adaptive control schemes. The control command issuance and execution feedback module is used to convert the control scheme into control commands and issue them to the controllable water conservancy facilities on site, while collecting execution status data and feeding it back to the system. The dynamic evaluation model engine and the intelligent control and decision-making module run in the digital twin environment.

[0048] Example 3: Dynamic assessment and control of non-point source pollution from tebuconazole in a small watershed in a hilly area; Taking a typical red soil hilly area small watershed (watershed area of ​​about 15.6 km², mainly tea gardens and paddy fields) as the application object, this invention is used to address the non-point source pollution problem of the triazole fungicide tebuconazole, which is widely used in this watershed, and demonstrates the specific implementation process and effects of this invention.

[0049] 1. Digital twin construction and data foundation; 1.1 Data Acquisition Network Deployment: An online monitoring sensor network comprising 8 nodes was deployed at key locations within the watershed, specifically including: Three comprehensive hydrological and water quality monitoring stations (sections S1, S2, and S3): equipped with multi-parameter water quality sensors (measuring flow rate, water level, turbidity, pH, water temperature, and conductivity) and a microfluidic chip-based rapid pesticide detector (detection frequency: 15 minutes / time, detection limit: 0.05). ).

[0050] Five farmland runoff monitoring points (P1-P5): simple runoff collection devices and water quality sensors are deployed to focus on monitoring farmland runoff during rainfall.

[0051] 1.2 Operation of the Mobile Laboratory Experimental Platform: The mobile experimental platform was operated synchronously, and topsoil (0-20cm) and ditch sediment samples from different land use types (tea gardens, paddy fields) within the watershed were collected monthly. Under controlled conditions (simulating on-site temperature, humidity, and pH), batch experiments were conducted to determine the adsorption-desorption kinetic parameters (Freundlich coefficients Kf and n) and degradation half-life (DT50) of tebuconazole. Table 1. Environmental behavior parameters of tebuconazole determined by laboratory simulation.

[0052] 2. Implementation of dynamic evaluation and parameter inversion; 2.1 Training and dynamic parameter inversion of the physical constraint transfer learning model; The following example uses data from a 30-day period in May 2024 (early rainy season). Weekly laboratory measurements (as shown in Table 1) are used as the target label. The online monitoring time series data (including tebuconazole concentration, pH, water temperature, and turbidity) within the corresponding time window are used as input features. Training a deep transfer learning model with physical constraints, key hyperparameter settings: , (Focusing more on data fitting), the network structure is a three-layer fully connected neural network (128-64-32 neurons).

[0053] The dynamic adsorption coefficients obtained by the model inversion on the validation set were compared with the laboratory measurements. The results show that the dynamic trend of the inverted values ​​can effectively capture the influence of soil moisture and temperature changes caused by rainfall events on adsorption behavior.

[0054] 2.2 Data assimilation and model state correction; A typical rainstorm event (24-hour rainfall of 85 mm) on June 15, 2024, was selected to demonstrate the data assimilation effect. The online monitoring data of tebuconazole concentration at section S2 was assimilated into a mechanistic model with embedded dynamic inversion parameters. An ensemble Kalman filter algorithm was used, with the ensemble membership set to 50.

[0055] Table 2 Comparison of model prediction performance before and after data assimilation (S2 section)

[0056] As can be seen from Table 2, data assimilation effectively corrects the biases caused by the initial conditions and input uncertainties in the model, and significantly improves the prediction accuracy of the peak time and magnitude of the pollution load.

[0057] 3. Generation of hierarchical intelligent decision-making and control schemes; In response to the rainstorm event on June 15, 2024, the digital twin system generated an adaptive control scheme based on the assimilated and corrected model state.

[0058] 3.1 Strategic level tactical selection; The system loaded five historical similar rainfall scenarios and simulated three control strategies: A (only the downstream main reservoir is activated), B (the main reservoir and ecological ditch #2 are activated), and C (the main reservoir, ecological ditch #2, and upstream decentralized reservoirs are activated). Based on multi-objective evaluation (maximizing load reduction and minimizing operational energy consumption), strategy B was selected as the baseline strategy.

[0059] 3.2 Tactical Layer Scrolling Optimization; Guided by strategy B, the model predictive control (MPC) algorithm is used to generate the next 12 hours (control time domain) in a rolling manner. Predicting the time domain Detailed operating instructions (control cycle 1.5 hours) are provided. The optimization objective is to control the tebuconazole concentration at section S2 at 10... The following steps aim to minimize the gate operation frequency.

[0060] Table 3 shows a partial sequence of control commands generated by MPC.

[0061] 3.3 Real-time fine-tuning at the execution layer; During command execution, the reinforcement learning agent fine-tuned the gate opening based on the deviation between the real-time monitored concentration and the MPC prediction value. At 17:45, when the actual concentration decreased faster than predicted, the agent fine-tuned the main dam gate opening to 55% (lower than the 60% MPC command), reducing unnecessary water release while ensuring control effectiveness.

[0062] 4. Evaluation and feedback of the regulation effect; After the control measures for this rainstorm event concluded, the system collected data on the effectiveness of the measures and fed it back to the digital twin.

[0063] Table 4 Comparison of the effects of different regulation modes

[0064] Improved assessment accuracy: By employing dynamic parameter inversion and data assimilation, the model's predicted NSE was improved from 0.62 to 0.88, laying the foundation for accurate decision-making.

[0065] Significant control effect: The intelligent control scheme reduced the peak concentration of tebuconazole at the target section to 9.5%. This is superior to traditional experience-based scheduling (15.8). Furthermore, the total discharge volume was reduced by 26.2%, achieving a balance between environmental benefits and water conservation.

[0066] System Adaptability: The fine-tuning of the reinforcement learning agent effectively compensates for model prediction errors (for example, the instruction at 17:30 in Table 3 is better after fine-tuning), demonstrating the system's online learning and adaptability.

[0067] The case data, model parameters, and control strategies from this successful regulation were automatically stored in the regulation strategy knowledge graph and digital contingency plan library. The system extracted a rule from it: "In scenarios similar to the early stage of soil dryness and short-term heavy rainfall, adopting strategy B (main pond dam + ecological ditch combination) and increasing the water diversion of the ecological ditch 0.5 hours in advance can more effectively intercept the initial high concentration of runoff."

[0068] Meanwhile, based on the simulation results of this event, the digital twin updated the distribution map of high-risk areas of tebuconazole non-point source pollution in the watershed and generated electronic fences, which were sent to the agricultural machinery dispatch platform of the local agricultural cooperative to guide subsequent agricultural activities (avoid soil tillage and other disturbing operations in high-risk areas within 3 days after rainfall).

[0069] The embodiments disclosed in this invention are preferred embodiments, but are not limited thereto. Those skilled in the art can easily understand the spirit of this invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of this invention, they are all within the protection scope of this invention.

Claims

1. A method for dynamic assessment and control of non-point source pollution load in a small watershed, characterized in that, Includes the following steps: S1: Construct a digital twin of a small watershed, wherein the digital twin is used to integrate geospatial data, real-time monitoring data and controllable water conservancy facility status; S2: Based on the aforementioned digital twin, a dynamic evaluation is performed using a non-point source pollution process model driven by both operational mechanisms and data, which includes: Based on online monitoring time-series data and laboratory simulation measurement data, the environmental behavior parameters of key pollutants are dynamically inverted through a physically constrained transfer learning model. Data assimilation technology is used to assimilate real-time monitoring data into the non-point source pollution process model in order to continuously correct the model state; S3: Based on the corrected model state and pollution load control target, an adaptive control scheme is generated in the digital twin through a hierarchical decision architecture. The hierarchical decision architecture includes: a strategic layer based on scenario simulation for long-term strategy selection, a tactical layer based on rolling optimization for medium-term instruction generation, and an execution layer based on real-time feedback for instantaneous adjustment. S4: The adaptive control scheme is sent to the corresponding controllable water conservancy facilities for execution, and the execution effect data is fed back to the digital twin to form a closed loop of evaluation and control.

2. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 1, characterized in that, In step S1, constructing the digital twin specifically includes: Access to multi-source heterogeneous data, including geospatial data, meteorological data, online monitoring sensor data, and facility operation status data; The multi-source heterogeneous data is standardized, fused, and spatiotemporally aligned to form a unified spatiotemporal dataset that drives the digital twin; The online monitoring sensor data includes at least flow rate, water level, turbidity, pH, and specific pollutant concentration indicators from key cross-sections.

3. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 1, characterized in that, The environmental behavior parameters of the key pollutants dynamically retrieved are specifically as follows: The target label is the laboratory simulation measurement data obtained periodically, and the input feature is the online monitoring time series data of the corresponding time period. The transfer learning model for the physical constraints is trained, and a dynamic mapping relationship from online monitoring data to environmental behavior parameters is established by minimizing the joint loss function; the joint loss function is: ； In the formula, The task loss function measures the model's prediction results. With real labels The differences between them; This is a physical constraint penalty function; The input feature vector; and These are weight hyperparameters; Using the trained model, dynamically updated environmental behavior parameters are calculated and output based on real-time monitoring data.

4. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 3, characterized in that, The physical constraint penalty item Based on the law of conservation of mass, its calculation formula is as follows: ； In the formula, For the concentration scalar field of pollutants; For time; This represents the water flow velocity vector field. For divergence operators; For predicted environmental behavior parameters Other input features The pollutant source and sink terms are jointly determined; this constraint forces the inversion parameters to satisfy the convection-diffusion reaction equation framework, ensuring its physical consistency.

5. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 1, characterized in that, In step S2, the mechanism- and data-driven non-point source pollution process model is a multi-scale coupled model, including: Microscale models used to simulate the microscopic migration and transformation processes of pollutants; Land parcel and gully scale models used to simulate land parcel runoff, soil erosion, pollutant transport with sediment, and biochemical processes within gullies; A small-watershed-scale distributed hydrological and water quality model used to simulate the spatial transport and accumulation of pollution loads from the source area to the receiving water body; Among them, models of different scales are coupled through the transmission and feedback of key state variables.

6. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 1, characterized in that, The continuous correction of the model state using data assimilation technology specifically employs an ensemble Kalman filter algorithm, whose model state update formula is as follows: ； In the formula, The state vector represents the predicted state from the model; This represents the optimal estimated state vector obtained after data assimilation analysis. Represents the observation vector; For observation operators; This is the Kalman gain matrix; this step improves short-term forecast accuracy by fusing real-time observation data and dynamically calibrating the model trajectory.

7. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 1, characterized in that, In step S3, the tactical layer's rolling optimization-based instruction generation employs a model predictive control algorithm, which solves the following optimization problem in each control cycle: ； Constraints: , , ; In the formula, This represents the objective function that needs to be minimized. For discrete time step index; The sequence of control instructions to be optimized; To control the time domain; The future predicted based on the model state corrected by the data assimilation technique The state variables of the step; For the desired pollution load or water quality status target; and It is a diagonal weight matrix; The state transition function represents the mechanism- and data-driven area source pollution process model after data assimilation and correction. Input for predicted future disturbances; and These represent the upper and lower limits of the amplitude of the control command, respectively. and These represent the upper and lower limits of the rate of change of the control command, respectively; the algorithm continuously solves for the optimal control command over a future period.

8. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 7, characterized in that, In step S3, the real-time feedback-based fine-tuning of the execution layer is implemented using a reinforcement learning agent, which learns the fine-tuning strategy by maximizing cumulative rewards, and the action-value function... The update follows the Bellman optimality equation: ； In the formula, For a moment Environmental state, state Including the model predictive control algorithm in time Predicted state Compared with the actual observation state The deviation, and the real-time status of controllable facilities; For a moment The actions taken by the agent are the baseline control commands output by the model predictive control algorithm. Fine-tuning amount; Indicates the state Next action The immediate reward signal received afterward; Discount factor; The mathematical expectation operator is used to enable the intelligent agent to adaptively compensate for model errors and unknown disturbances online.

9. The method for dynamic assessment and control of non-point source pollution load in a small watershed according to claim 1, characterized in that, In step S4, the timing control instructions in the adaptive control scheme are specific parameters for the opening and closing timing, opening degree, or operating water level of ponds, dams, gates, or adjustable interception dams; when the control object is multiple interception dams arranged in a distributed manner, the adaptive control scheme is a set of coordinated control instructions generated according to the spatiotemporal distribution of pollution load and coordinated with each other in terms of water storage and discharge timing.

10. A system for dynamic assessment and control of non-point source pollution load in a small watershed, used to implement the method for dynamic assessment and control of non-point source pollution load in a small watershed as described in any one of claims 1-9, characterized in that, include: The data sensing and acquisition module is used to acquire multi-source data required to build and drive the digital twin; The digital twin construction and maintenance module is used to integrate the multi-source data and establish and maintain a virtual mapping model of the small watershed; The dynamic evaluation model engine incorporates a non-point source pollution process model driven by both the mechanism and data, a transfer learning parameter inversion unit, and a data assimilation unit. The intelligent control and decision-making module, which incorporates the hierarchical decision-making architecture, is used to generate adaptive control schemes. The control command issuance and execution feedback module is used to convert the control scheme into control commands and issue them to the field facilities, while collecting execution status data and feeding it back to the system. The dynamic evaluation model engine and the intelligent control and decision-making module operate in the digital twin environment, forming a closed loop.

Images