Hydrodynamic water quality rapid simulation and emergency regulation method and device integrated with artificial intelligence

Abstract

The present application belongs to the field of river and lake hydrodynamic water environment simulation and water resources regulation, and relates to a water dynamic water quality rapid simulation and emergency regulation method and device integrating artificial intelligence, the method comprising the steps of: constructing a training data set based on historical hydrological sequences; constructing a shallow lake hydrodynamic proxy model based on a physical information neural network; implementing grid scale mass conservation constraints and flow flux correction; performing multi-parameter synchronous rapid prediction of the hydrodynamic field; performing rapid solution of typical water quality element transport and overall model performance evaluation; and implementing differentiated emergency regulation for sudden pollution events during water diversion and non-water diversion periods. The present application takes into account physical reliability and prediction timeliness, realizes a magnitude leap in calculation efficiency, has differentiated emergency regulation capability, can deduce the spatio-temporal evolution characteristics of pollution groups under different working conditions, and provides scientific support and efficient tools for daily management of river and lake water environment and emergency regulation and decision-making for sudden water pollution events.

Description

Technical Field

[0001] This invention relates to the field of river and lake hydrodynamic and water environment simulation and water resource regulation technology, and in particular to a method and device for rapid simulation and emergency regulation of hydrodynamics and water quality that integrates artificial intelligence. Background Technology

[0002] While existing hydrodynamic water quality numerical models based on governing equations possess strong physical interpretability, their high computational cost makes them insufficient to meet the timeliness requirements of daily water environment management and emergency response to sudden events. Purely data-driven rapid prediction methods, though computationally efficient, still suffer from insufficient physical consistency and engineering reliability in their prediction results. Large-scale open water transfer systems such as the South-to-North Water Diversion Project, the Datong-Minjiang Water Diversion Project, and the National Modern Water Network Project place even higher demands on the rapid and accurate simulation of water quality spatiotemporal patterns, necessitating a rapid simulation method that integrates the advantages of artificial intelligence and numerical simulation. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for rapid simulation and emergency control of hydrodynamics and water quality that integrates artificial intelligence, employing the following technical solution, including the following steps:

[0004] S1. Construct a training dataset based on historical hydrological sequences;

[0005] S2. Construct a hydrodynamic proxy model based on Physical Information Neural Networks (PINNs): Construct a fully connected feedforward neural network, taking historical boundary time series data as input and calculating the water level value of the grid cell as output; construct a loss function that includes a data-driven term and a global water balance regularization term; train the network using an optimizer to obtain the hydrodynamic proxy model;

[0006] S3. Implement grid-scale mass conservation constraints and flow flux correction: Construct a mass conservation loss function focused on each computational grid cell and refine the training of the neural network; based on the idea of ​​the pressure Poisson equation, perform post-processing correction on the predicted flow flux field to ensure that it strictly satisfies the continuity equation of each grid cell.

[0007] S4. Perform simultaneous and rapid prediction of multiple parameters of hydrodynamic field: Adjust the neural network structure, take historical water level sequence and current flow data as input, and output the center water level value of all grid cells and the flow flux value of all grid surfaces simultaneously. Use multi-task joint loss function and layer-by-layer pre-training strategy to train and fine-tune the model.

[0008] S5. Based on the hydrodynamic surrogate model, perform rapid solution and overall performance evaluation of water quality transport: Based on the flow flux field and water level field predicted in S4, solve the two-dimensional convection-diffusion equation; use the finite volume method explicit scheme for discretization, and linearly interpolate the hydrodynamic field between substeps; adopt a modular decoupling coupling strategy for the hydrodynamic surrogate model and the water quality solver; establish a rapid hydrodynamic and water quality simulation model to conduct high-fidelity rapid simulation of the spatiotemporal evolution of the water level field, flow flux field, and water quality element concentration field; conduct a horizontal comparison of the simulation results of the rapid simulation model and the TUWMM high-fidelity numerical model, including performance indicators in three dimensions: water level prediction accuracy, water quality prediction accuracy, and computation time, to achieve an overall performance evaluation of the model;

[0009] S6. Implement differentiated emergency control measures for water diversion periods and non-water diversion periods: Utilize the spatiotemporal evolution characteristics of pollutants predicted in S5 to formulate strategies for shutting down gate pumps and physical interception during water diversion periods, and formulate strategies for joint scheduling to increase downstream flow or shutting down outlet gates for retention and dilution during non-water diversion periods. Utilize the rapid simulation capabilities of hydrodynamics and water quality to conduct dynamic quantitative evaluation and iterative optimization of the effects of different emergency control strategies.

[0010] Preferably, the data augmentation based on traffic time pattern recognition in step S1 specifically includes the following steps:

[0011] S11. Collect historical hydrological data and extract time series data;

[0012] S12. Construct sliding window input-output sample pairs;

[0013] S13. Data set partitioning and normalization.

[0014] Preferably, step S2, the step of constructing a hydrodynamic proxy model based on a physical information neural network, specifically includes:

[0015] Constructing a hydrodynamic proxy model framework: A fully connected feedforward neural network is constructed, with N × C nodes in the input layer, where N = 8 is the number of input time steps, and C is the feature dimension of each time step. The hidden layer depth is 4 layers, with 128 neurons per layer. The activation function is the hyperbolic tangent function. ,in: Let be the value of the hyperbolic tangent function, dimensionless, and its range is . , For input values, It is a natural constant;

[0016] The input data time interval is Δt = 21600 s, and the total time span of the input sequence is N × Δt = 48 hours;

[0017] Construct a grid-scale mass conservation loss function;

[0018] Efficient physical loss calculation based on OpenFOAM;

[0019] Perform post-processing corrections on the flow flux field;

[0020] The steps in step S3, which implement grid-scale mass conservation constraints and flux correction, specifically include:

[0021] Constructing a grid-scale mass conservation loss function: Focusing physical constraints on each computational grid cell, the mass conservation equation for each grid cell i is discretized as follows:

[0022] Where t is the current time, The water level at the center point of the grid. =5400 s is the discrete time interval of the water level time gradient. For the grid projection area, Let j be the flow flux through the j-th boundary of the grid. The total number of grid boundaries;

[0023] Define the mass-conserving residual for each grid cell: ;

[0024] The mass conservation loss function at the global grid scale is: ,in, =2680 is the total number of grid cells;

[0025] The improved composite loss function is: ,in: The value of the improved composite loss function, in m². This is a data-driven term, specifically the mean square error between the predicted and actual values, expressed in m². The coefficients of the physical regularization term at the grid scale are dimensionless. , This represents the mass conservation loss function value at the global grid scale, in meters. 2 / s 2 ;

[0026] Efficient physical loss calculation based on OpenFOAM: The predicted water level and flow flux output by the neural network are denormalized to restore the water level at the grid center point. and the flow flux through the j-th boundary of the grid The divergence operator is invoked to automatically sum and normalize the flux around each grid cell based on the Gaussian divergence theorem, yielding the net flux field. This net flux field is then added to the water level time gradient field to obtain the mass-conserved residual field, which is then used to calculate... ;

[0027] Post-processing correction of flux field: Using the idea of ​​the pressure Poisson equation, the original flux field φ predicted by the neural network is regarded as a predicted value that does not strictly satisfy mass conservation, and a correction potential function is constructed. Poisson's equation: ,in: To correct the potential function Laplace operator, unit: s -1 , The corrected potential function, unit: m 2 / s, For the original flux field divergence, unit: s -1 , The original flux field predicted by the neural network, in meters. 3 / s;

[0028] The linear system is solved using the conjugate gradient method combined with incomplete Cholesky preprocessing; the flux correction for each grid surface is calculated. ,in: For grid surface Flux correction, unit: m 3 / s, For grid surface Area, unit: m² For grid surface One side of the grid center Potential function value, unit: m 2 / s, For grid surface The center of the grid on the other side Potential function value, unit: m 2 / s, For grid center and The distance between them, in meters;

[0029] Corrected flux field The mass conservation of each grid cell is satisfied, where: For the corrected flux field on the grid surface The value above satisfies the law of conservation of mass, with units of m³ / s. The original flux field predicted by the neural network on the grid surface Values ​​above, unit: m 3 / s, For grid surface Flux correction, unit: m 3 / s.

[0030] Preferably, the step of performing simultaneous and rapid prediction of multiple parameters of the hydrodynamic field in step S4 specifically includes:

[0031] S41. Network Input / Output Structure Adjustment: The input layer is expanded to simultaneously contain water level sequences of N = 8 historical time steps and current flow data. The input tensor shape is [8, 14], where 14 corresponds to twice the feature dimension of the 7 boundaries; the output layer synchronously outputs N... cell = Center water level values ​​of 2680 grid cells and N face = Flux values ​​for 4988 grid surfaces, with a total of 7668 output nodes;

[0032] S42. Design of Multi-Task Joint Loss Function: The data-driven term is expanded to a weighted sum of water level error and flow rate error. ,in: This is the flow rate normalization factor, in m³ / s. For the expanded data-driven item, unit: m². The total number of grid cells, dimensionless. , For grid indexing, , For the first Water level prediction values ​​for each grid cell, in meters. For the first Actual water level values ​​for each grid cell, unit: meters. The weighting coefficient for the flow error term. , This represents the total number of grid faces. , For grid face indexing, , For the first Predicted flow flux values ​​for each grid surface, in m³. 3 / s, For the first The actual flow flux value of each grid surface, in m³. 3 / s;

[0033] The joint loss function is: ,in: The value represents the joint loss function, in m². For the expanded data-driven item, unit: m 2 , For the physical regularization coefficients at the grid scale, , This represents the mass conservation loss function value at the global grid scale, in meters. 2 / s2 ;

[0034] S43. Layer-by-layer pre-training and model fine-tuning: First, train the model to convergence using water level prediction as a single task, and fix the network parameters of the water level output branch; then add a flow prediction branch, and use the hidden layer features of the water level output branch as the shared input of the flow prediction branch for joint fine-tuning; the learning rate in the fine-tuning stage is set to 0.1 times the initial learning rate, and the training is carried out for 200 epochs; after the model outputs, post-processing correction is performed on the flow flux field.

[0035] Preferably, the step in S5, which involves rapidly solving for water transport and evaluating overall performance based on a hydrodynamic proxy model, specifically includes:

[0036] The water transport process is described using a two-dimensional convection-diffusion equation:

[0037] ,in: Water depth, unit: meters. Pollutant concentration, unit: mg / L For time variables, the unit is seconds. For concentration flux Regarding time The partial derivatives, in mg / (L·s), This is the advection term, representing the transport flux of pollutants along the water flow, measured in mg / (L·s). This is a velocity vector, with units of m / s. This is the diffusion term, representing the diffusion flux of pollutants driven by the concentration gradient, in mg / (L·s). The diffusion coefficient, in m² / s, ranges from 0.01 to 1.0 m² / s. Source and sink terms, unit: mg / (L·s); ,in: Source and sink terms, unit: mg / (L·s) Attenuation coefficient, unit: s -1 , Pollutant concentration, unit: mg / L;

[0038] The solution is obtained by explicit discretization using the finite volume method, with a time step of [missing information]. = 1800 s satisfies the CFL stability condition; a modular decoupling coupling strategy is adopted to decouple hydrodynamic prediction from water quality solution. Within one hydrodynamic time step Δt = 21600 s, the water quality solver performs 12 sub-step iterations, and the hydrodynamic field is linearly interpolated between sub-steps.

[0039] Establish a rapid simulation model for hydrodynamics and water quality, and conduct rapid simulation of the spatiotemporal evolution of high-fidelity water level field, flow flux field and water quality element concentration field.

[0040] A horizontal comparison was conducted between the simulation results of the rapid simulation model and the TUWMM high-fidelity numerical model, including performance indicators in three dimensions: water level prediction accuracy (measured by root mean square error RMSE), water quality prediction accuracy (measured by mean relative error MRE), and computation time, to evaluate the overall performance of the model.

[0041] Preferably, the step of implementing differentiated emergency regulation for water diversion periods and non-water diversion periods in S6 specifically includes:

[0042] S61. Comparative analysis of pollution diffusion characteristics between water diversion period and non-water diversion period: Set up two comparative working conditions, water diversion period and non-water diversion period, run the rapid water quality simulation method, output the concentration field at different times after the accident, and track the center location and diffusion range of the pollution plume.

[0043] S62. Differentiated emergency control strategies: For the water diversion period, develop emergency strategies including closing the secondary dam gates, deploying physical barriers and adsorption barriers downstream of the accident point, and carrying out in-situ treatment; for the non-water diversion period, develop emergency strategies such as joint scheduling to increase the discharge flow to quickly push the pollution plume out of the lake area, or closing the outlet gates to retain the pollution plume in the middle of the lake for dilution and degradation.

[0044] S63. Strategy Effectiveness Evaluation and Dynamic Optimization: Quantify the control measures into model boundary conditions or parameter modifications, and rerun the rapid water quality simulation method; calculate the peak concentration reduction rate. Pollution duration reduction rate Scope of influence reduction rate If the strategy does not achieve the expected results, adjust the location and density of the enclosure, adjust the opening and closing sequence of the gate, and optimize the amount and location of the adsorption material. Each iteration simulation takes 5 to 10 seconds, and the optimal combination of control strategies is output.

[0045] Preferably, the differentiated emergency control measures implemented in S6 for water diversion periods and non-water diversion periods also include an adaptive enclosure layout optimization method based on pollutant concentration gradients, specifically including:

[0046] Based on the pollutant concentration field predicted by S5, the concentration gradient along the mainstream direction of each grid cell is calculated: ,in, The concentration gradient vector at position (x,y) (unit: ) and These are the partial derivatives of the concentration in the x and y directions, respectively. and These are the unit vectors in the x and y directions, respectively;

[0047] Calculate the magnitude of the concentration gradient: ;

[0048] The concentration gradient magnitude exceeds the threshold θ grad = 0.05 The grid cells were identified as pollution front areas;

[0049] Within the pollution front area, the center point of the grid unit is used as the candidate deployment point. The k-means clustering algorithm is used to cluster the candidate points, where k = 3 to 8. The cluster center is used as the actual deployment location of the enclosure.

[0050] Enclosure density Positively correlated with the magnitude of the local concentration gradient:

[0051] ,in, For position Density of enclosure layout at the location = 0.5 lanes / km is the minimum deployment density. = 3.0 lanes / km is the maximum deployment density. For position Concentration gradient modulus at that location, This represents the maximum value of the concentration gradient modulus within the entire pollution front area; based on the adaptive deployment results, a containment deployment scheme is generated, and the strategy effectiveness is evaluated.

[0052] Preferably, the rapid solution for water quality transport based on the hydrodynamic proxy model in S5 further includes a method for accelerating the solution for pollutant transport based on an adaptive time-step splitting operator, specifically including:

[0053] The two-dimensional convection-diffusion equation is split into two successive stages: the convection substep and the diffusion substep.

[0054] Convection step: The convection terms are discretized using a second-order TVD (Total Variation Diminishing) scheme, with a critical time step of [missing information]. Satisfying CFL conditions:

[0055] Where Δx is the grid scale (unit: m). = 10 -6 To prevent division by zero for small constants;

[0056] Diffusion substep: The diffusion term is discrete using an implicit scheme, with a critical time step of [missing information]. satisfy: ;

[0057] The adaptive global time step takes the minimum of the two values: Where β = 0.8 is the safety factor;

[0058] An adaptive time step is used within each hydrodynamic time step Δt = 21600 s. The process involves sub-loop propagation, with linear interpolation of the hydrodynamic field between adjacent sub-steps; when the average magnitude of the pollutant concentration gradient... mg·L -1 ·m -1 When the time step is reached, it will automatically increase to 1.5 times the current value, with a maximum not exceeding Δt. adaptive_max = 900s.

[0059] Preferably, step S3, which implements grid-scale mass conservation constraints and flux correction, further includes a dynamic adjustment method for physical regularization weights based on multi-scale residual adaptive adjustment, specifically including:

[0060] The grid cells are divided into three scale levels based on their area: small-scale grid. <5000m 2 5000 m mesoscale grid 2 ≤ < 20000 m 2 Large-scale grid A i ≥ 20000 m 2 ; Calculate the average mass conservation residuals for each scale level, where: For the first Area of ​​each grid cell:

[0061] ,in, For the first The average quality-conserving residual within the class scale level, where s is the scale level index, s ∈ {small, medium, large}. Let be the number of grid cells of type s. For grid cell indexing, For those belonging to the first Summing all grid cells at the class-scale level, For the first The mass conservation residual of each grid cell For the first Mass conservation residual of each grid cell The absolute value;

[0062] Define the scale-adaptive weighting coefficients: ,in, For the first Adaptive weighting coefficients for class scale levels are used to amplify or reduce the contribution of that scale level to the loss function. To maximize the function, the weight coefficients must be at least 0.1 to prevent them from being too small and causing the scale level to be ignored. To minimize the function, the weight coefficients are set to no higher than 3.0 to prevent instability during training due to excessively large values. For the first Average quality-conserving residuals within the class-scale level The global average residual;

[0063] Construct a multi-scale weighted mass conservation loss function: ,in, The multi-scale weighted mass conservation loss function value is used to replace the original Participate in model training, This represents the total number of grid cells, with a value of 2680. To sum over all grid cells, For grid Belonging to the scale level The corresponding adaptive weight coefficients, For the first Mass conservation residual of each grid cell The square of , s(i) is the scale level to which grid i belongs; the physical regularization coefficient λ mass During training, adjustments are made dynamically according to the following rules: the validation set is computed every 10 epochs. If the loss does not decrease after three consecutive assessments, then λ will be... mass Multiply by 1.2; if Then λ mass Multiply by 0.8; λ mass The adjustment range is limited to the interval [0.01, 0.5].

[0064] To address the aforementioned technical problems, this invention also provides a rapid simulation and emergency control device for hydrodynamics and water quality integrating artificial intelligence, employing the following technical solution, including:

[0065] The data module is used to construct training datasets based on historical hydrological sequences;

[0066] The module is used to construct a hydrodynamic proxy model based on a physical information neural network: a fully connected feedforward neural network is constructed, which takes historical boundary time series data as input and calculates the water level value of the grid cell as output; a loss function containing data-driven terms and global water balance regularization terms is constructed; and an optimizer is used to train the network to obtain the hydrodynamic proxy model.

[0067] The correction module is used to implement grid-scale mass conservation constraints and flow flux correction: it constructs a mass conservation loss function focused on each computational grid cell and performs refined training on the neural network; based on the idea of ​​the pressure Poisson equation, it performs post-processing correction on the predicted flow flux field to ensure that it strictly satisfies the continuity equation of each grid cell.

[0068] The prediction module is used to perform synchronous and rapid prediction of multiple parameters of the hydrodynamic field: the neural network structure is adjusted, and the historical water level sequence and the current flow data are used as inputs to synchronously output the center water level value of all grid cells and the flow flux value of all grid surfaces. The model is trained and fine-tuned by using a multi-task joint loss function and a layer-by-layer pre-training strategy.

[0069] The solver module is used for rapid solution and overall performance evaluation of water transport based on a hydrodynamic surrogate model: it solves the two-dimensional convection-diffusion equations based on the flow flux field and water level field; it uses the finite volume method explicit scheme for discretization, and linear interpolation of the hydrodynamic field between substeps; the hydrodynamic surrogate model and the water quality solver adopt a modular decoupling coupling strategy; it establishes a rapid hydrodynamic and water quality simulation model to conduct high-fidelity rapid simulation of the spatiotemporal evolution of the water level field, flow flux field, and water quality element concentration field; it conducts a horizontal comparison of the simulation results of the rapid simulation model and the TUWMM high-fidelity numerical model, including performance indicators in three dimensions: water level prediction accuracy, water quality prediction accuracy, and computation time, to achieve an overall performance evaluation of the model;

[0070] The control module is used to implement differentiated emergency control for water diversion periods and non-water diversion periods: it uses the predicted spatiotemporal evolution characteristics of pollutants to formulate strategies for shutting down gate pumps and physical interception during water diversion periods, and formulates strategies for joint scheduling to increase the downstream flow or shutting down the outlet gate for retention and dilution during non-water diversion periods, and uses rapid simulation capabilities to dynamically evaluate and iteratively optimize the effectiveness of the strategies.

[0071] Compared with the prior art, the present invention has the following main advantages:

[0072] (1) It takes into account both physical reliability and computational efficiency: By combining physical information neural networks with high-precision numerical simulation, the data and physical mechanism dual-driven proxy model is constructed. While ensuring that the prediction accuracy is comparable to that of traditional numerical models, the computation speed is improved by 92.2% to 97.6%, achieving a leap in computational efficiency.

[0073] (2) Constructing a continuous, high-quality training dataset: Using a sliding window to generate input-output sample pairs provides a supervisory signal for the neural network model to learn the dynamic evolution of water level. It is a bridge connecting the original observation data and the physical information neural network model, and determines the basis of the model's predictive ability.

[0074] (3) It has strict physical consistency guarantee: By focusing the mass conservation constraint on the grid scale and using the flow flux post-processing correction algorithm based on the pressure Poisson equation, it ensures that each computational grid cell strictly satisfies the continuity equation, providing a high-fidelity hydrodynamic background field for water transport simulation.

[0075] (4) A modular decoupling and coupling strategy was adopted: By taking into account the essential differences in the propagation speed of hydrodynamic and water quality parameters, different prediction methods were adopted (surrogate model prediction of hydrodynamic field + analytical solution of convection-diffusion equation), which avoided the problem that a single neural network could not maintain physical self-consistency in long-term prediction.

[0076] (5) It has differentiated emergency control capabilities: It can quickly deduce the spatiotemporal evolution characteristics of pollution plumes under different working conditions, and formulate differentiated emergency control strategies for water diversion periods and non-water diversion periods. At the same time, it supports rapid iterative evaluation of the effectiveness of the strategies, providing scientific support and efficient tools for emergency decision-making in sudden water pollution events. Attached Figure Description

[0077] To more clearly illustrate the solutions in this invention, the accompanying drawings used in the description of the embodiments of this invention will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0078] Figure 1 This is a flowchart of an embodiment of the method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to the present invention;

[0079] Figure 2 For 2025, a certain East Railway Station COD Mn A diagram showing the comparison of predicted concentrations;

[0080] Figure 3 This is a schematic diagram showing the evolution of the proportion of polluted area over time under different operating conditions after adopting this method;

[0081] Figure 4 This is a schematic diagram of a structure of an embodiment of the rapid simulation and emergency control device for hydrodynamics and water quality integrating artificial intelligence of the present invention;

[0082] Figure 5 This is a schematic diagram of the structure of an embodiment of the computer device of the present invention. Detailed Implementation

[0083] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the specification is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects and not to describe a particular order.

[0084] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0085] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0086] It should be noted that the rapid simulation and emergency control method for hydrodynamics and water quality integrating artificial intelligence provided in the embodiments of the present invention is generally executed by a server / terminal device. Correspondingly, the rapid simulation and emergency control device for hydrodynamics and water quality integrating artificial intelligence is generally installed in the server / terminal device.

[0087] It should be understood that the number of terminal devices, networks, and servers is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be used.

[0088] Example 1

[0089] Please refer to Figure 1 The flowchart illustrates an embodiment of the rapid simulation and emergency control method for hydrodynamic and water quality integrating artificial intelligence according to the present invention. The rapid simulation and emergency control method for hydrodynamic and water quality integrating artificial intelligence includes the following steps:

[0090] S1. Construct a training dataset based on historical hydrological sequences.

[0091] In specific implementation, step S1 further includes the following steps:

[0092] S11. Collect historical hydrological data and extract time series data.

[0093] Historical hourly hydrological data from multiple boundary locations (including the secondary dam and Hanzhuang sluice gate) were extracted from the long-term hydrological monitoring system of the target water area (such as the lower lakes of Nansi Lake). Each record includes a timestamp, water level, and flow information. The data spans at least 3 years, with a time resolution of 1 hour (3600 s). The extracted raw data are arranged in chronological order to form a continuous multivariate time series.

[0094] Raw monitoring records are used directly without event segmentation, filtering, or pattern clustering. Missing data are imputed using linear interpolation, and outliers (such as negative flow rates or abrupt changes exceeding three times the standard deviation) are replaced with the mean of the preceding and following times. All boundary water level and flow data maintain their original time alignment.

[0095] The purpose of step S11 is to provide a real, continuous, and complete sequence of hydrological processes, preserve the dynamic coupling characteristics of water level and flow under the combined effects of artificial regulation (such as gate scheduling during water diversion periods) and natural inflow, and provide high-confidence training samples for subsequent supervised learning.

[0096] S12. Construct sliding window input-output sample pairs.

[0097] A sliding window method is used to convert continuous time series into sample pairs required for supervised learning. The window length is set. Time interval (6 hours), therefore the total duration covered by each input window is .

[0098] Assume there is a total Each boundary (such as the secondary dam, Hanzhuang sluice gate, etc.) is at time... Record water level and traffic Then the first Input tensor of each sample for:

[0099] .

[0100] The feature vector at each time step Defined as:

[0101] .

[0102] in: : No. The input tensor for each sample has the shape of... , dimensionless (but the internal element units are m or m³ / s). : No. The current time corresponding to each sample, in seconds; : Time steps, dimensionless; Time interval; :time eigenvectors; : No. A boundary at time Water level, unit: m; : No. A boundary at time Flow rate, unit: m³ / s; Total number of boundary values, dimensionless.

[0103] Output target For the current moment The center water level value of all calculated grid cells:

[0104] .

[0105] in: : No. The output target vector for each sample, with length Unit: m; : No. Each grid cell at time... The central water level value, in meters; Total number of grid cells, dimensionless.

[0106] Slide window in steps Moving along the time axis generates a large number of overlapping samples, significantly expanding the size of the training set.

[0107] Step S12 transforms the continuous hydrological time series into an input-output pair that can be directly used for supervised learning. The input consists of the water level and flow rate sequences of each boundary over the past 48 hours, and the output is the current global water level field. This structure enables the neural network to learn the nonlinear mapping relationship between historical boundary conditions and the current spatial distribution of water levels.

[0108] S13. Data set partitioning and normalization.

[0109] The generated sample set is divided into a training set (70%), a validation set (15%), and a test set (15%) in chronological order, ensuring that the test set is later than the training set in time, so as to simulate the generalization ability to future events in actual prediction scenarios.

[0110] Z-score normalization is performed on each feature dimension (each boundary water level, each boundary flow rate) and the output water level field: .

[0111] in: : The dimensionless value after normalization; Original values ​​(water level in m, flow rate in m³ / s). The mean of this feature on the training set (in units of...) same); The standard deviation of this feature on the training set (in units of...) same).

[0112] The normalization of the output water level field also uses the above formula, where For grid water levels (unit: m). All normalized parameters ( It is computed only from the training set and applied to the validation and test sets.

[0113] The purpose of step S13 is to eliminate the differences in dimensions and magnitudes between different physical quantities (water level and flow rate) and different boundaries, making neural network training more stable and efficient; at the same time, it evaluates the model's ability to extrapolate future unknown working conditions by dividing the time sequence.

[0114] The purpose of step S1 is to provide a data foundation and monitoring signals for step S2.

[0115] S2. Construct a hydrodynamic proxy model based on a physical information neural network.

[0116] In specific implementation, step S2 further includes the following steps:

[0117] S21. Construct a hydrodynamic proxy model framework.

[0118] A fully connected feedforward neural network is constructed with N × C nodes in the input layer, where N = 8 is the number of input time steps, and C is the feature dimension at each time step. The hidden layer has a depth of 4 layers, with 128 neurons per layer. The hyperbolic tangent activation function is used. Among them, among them: Let be the value of the hyperbolic tangent function, dimensionless, and its range is . , Input value, dimensionless. It is a natural constant, approximately equal to 2.71828, and is dimensionless.

[0119] The number of output layer nodes equals the total number of computational grid cells, and the output is the water level value at the center point of each grid. The input data time interval is Δt = 21600 s, and the total time span of the input sequence is N × Δt = 48 hours.

[0120] The purpose of step S21 is to determine the topology, time parameters, and data preprocessing methods of the neural network, laying the architectural foundation for the model's learning ability, convergence speed, and generalization performance.

[0121] S22. Construct the grid-scale mass conservation loss function.

[0122] The total loss function is: .in: The total loss function value is dimensionless. This is a data-driven term, specifically the mean square error between the predicted and actual water levels. Unit: m² The coefficient of the physical regularity term is dimensionless. In this invention... , For reference flow rate, unit: m³ / s, historical average flow rate is used. m³ / s, This is the physical regularization term, which is the square of the overall water balance deviation, in units of: , ,in, for Total water storage of the lake at any given time, in m³. The time interval is expressed in seconds (s). For the first A boundary at Flow rate at any given time, in m³ / s. Dimensionless physical regularization term, unit and Keep it consistent, both are m².

[0123] The global water balance in discrete time is expressed as: .in: for Total water storage of the lake at any given time, in m³. for Total water storage of the lake at any given time, in m³. The time interval is expressed in seconds (s). Boundary index, dimensionless. For the first A boundary at Flow rate at any given time, in m³ / s.

[0124] The purpose of step S22 is to define the objective function for model optimization, which is composed of a weighted average of data-driven terms and physical regularization terms, so that the model can fit the data and comply with the basic physical law of water conservation during the training process.

[0125] S23. Based on OpenFOAM, perform efficient physical loss calculation.

[0126] The predicted water level and flow flux output from the neural network are denormalized to restore the water level at the grid center point. and the flow flux through the j-th boundary of the grid The divergence operator is invoked to automatically sum and normalize the flux around each grid cell based on the Gaussian divergence theorem, yielding the net flux field. This net flux field is then added to the water level time gradient field to obtain the mass-conserved residual field, which is then used to calculate... ;

[0127] S24. Perform post-processing corrections for the flow flux field.

[0128] Using the idea of ​​the pressure Poisson equation, the original flux field φ predicted by the neural network is regarded as a predicted value that does not strictly satisfy mass conservation, and a corrected potential function is constructed. Poisson's equation: ,in: To correct the potential function Laplace operator, unit: s -1 , The corrected potential function, unit: m 2 / s, For the original flux field divergence, unit: s -1 , The original flux field predicted by the neural network, in meters. 3 / s;

[0129] The linear system is solved using the conjugate gradient method combined with incomplete Cholesky preprocessing; the flux correction for each grid surface is calculated. ,in: For grid surface Flux correction, unit: m 3 / s, For grid surface Area, unit: m² For grid surface One side of the grid center Potential function value, unit: m 2 / s, For grid surface The center of the grid on the other side Potential function value, unit: m 2 / s, For grid center and The distance between them, in meters;

[0130] Corrected flux field The mass conservation of each grid cell is satisfied, where: For the corrected flux field on the grid surface The value above satisfies the law of conservation of mass, with units of m³ / s. The original flux field predicted by the neural network on the grid surface Values ​​above, unit: m 3 / s, For grid surface Flux correction, unit: m 3 / s.

[0131] The purpose of step S2 is to establish a rapid mapping relationship from boundary time series data to the water level field, serving as the basis for a surrogate model to replace the traditional high-precision numerical model. By embedding physical laws (mass conservation) into the loss function in the form of regularization terms, it ensures that the model can maintain basic physical consistency under limited samples.

[0132] S3. Implement grid-scale mass conservation constraints and flow flux correction.

[0133] In specific implementation, step S3 further includes the following steps:

[0134] S31. Construct the grid-scale mass conservation loss function.

[0135] Based on the continuity equation, the mass conservation law within each computational grid cell is discretized as follows:

[0136] .in: for Time Grid Water level at the center point, unit: m. for Time Grid Water level at the center point, unit: m. The discrete time interval of the water level time gradient, in seconds, is given in this invention. s, For grid Projected area, unit: m² For grid The total number of boundary values, dimensionless. Boundary index, dimensionless. , To pass through the grid No. The flow flux at the boundary, in m³ / s.

[0137] Define the mass-conserving residual for each grid cell: .in: For grid The mass conservation residual, in m / s. for Time Grid Water level at the center point, unit: m. for Time Grid Water level at the center point, unit: m. The discrete time interval of the water level time gradient, in seconds. For grid Projected area, unit: m² For grid The total number of boundary values, dimensionless. Boundary index, dimensionless. To pass through the grid No. The flow flux at the boundary, in m³ / s.

[0138] The mass conservation loss function at the global grid scale is: .in: This represents the mass conservation loss function value at the global grid scale, in meters. 2 / s 2 , The total number of grid cells is dimensionless. In this embodiment... The total number of grid cells. For grid indexing, dimensionless. , For grid The mass conservation residual, in m / s.

[0139] The improved composite loss function is: .in: The value of the improved composite loss function, in m². This is a data-driven term, specifically the mean square error between the predicted and actual values, expressed in m². The coefficients of the physical regularization term at the grid scale are dimensionless. In this invention... , This represents the mass conservation loss function value at the global grid scale, in meters. 2 / s 2 .

[0140] The purpose of step S31 is to define the mass conservation residual for each grid cell based on the discretized form of the continuity equation, construct the mass conservation loss function at the global grid scale, and provide stronger physical constraints for the central region of the computational domain far from the boundary constraints.

[0141] S32. Based on OpenFOAM, perform efficient physical loss calculation.

[0142] The predicted water level and flow flux output from the neural network are denormalized to restore the water level value with physical meaning. and flow flux value The divergence calculation operator is invoked, and based on the Gaussian divergence theorem, the summation of flux around each grid cell and area normalization are automatically performed. .in: For grid cells Volume, unit: m³ velocity vector divergence, unit: , This is a velocity vector, with units of m / s. For volumetric elements, the unit is m³. For grid cells Surface area, unit: m² For closed surfaces The integral symbol, The unit outward normal vector is dimensionless. The area is represented by a micro-element, in m².

[0143] Adding the net flux field to the water level time gradient field yields the mass-conserved residual field, which is then used to calculate... .

[0144] The purpose of step S32 is to utilize the tensor field computation capabilities of the OpenFOAM platform to avoid inefficient explicit loop traversal and significantly improve the computational efficiency and numerical accuracy of the physical regularization loss term during backpropagation.

[0145] S33. Perform post-processing corrections for the flow flux field.

[0146] Constructing the modified potential function Poisson's equation: .in: To correct the potential function The Laplace operator, unit: , This is the corrected potential function, in m² / s. For the original flux field divergence, unit: , The original flux field predicted by the neural network, in m³ / s.

[0147] The linear system is solved using the conjugate gradient method combined with incomplete Cholesky preprocessing.

[0148] Calculate the flux correction for each grid surface: Among them, among them: For grid surface Flux correction, in m³ / s. For grid surface Area, unit: m² For grid surface One side of the grid center The potential function value, in m² / s. For grid surface The center of the grid on the other side The potential function value, in m² / s. For grid center and The distance between them, in meters.

[0149] Corrected flux field .in: For the corrected flux field on the grid surface The value above strictly satisfies the law of conservation of mass, with units of m³ / s. The original flux field predicted by the neural network on the grid surface The value, in m³ / s, For grid surface Flux correction, in m³ / s.

[0150] In specific implementation, step S3 implements grid-scale mass conservation constraints and flow flux correction, and also includes a dynamic adjustment method for physical regularization weights based on multi-scale residual adaptation, specifically including:

[0151] The grid cells are divided into three scale levels based on their area: small-scale grid. <5000m 2 5000 m mesoscale grid 2 ≤ < 20000 m 2 Large-scale grids ≥ 20000 m 2 ; Calculate the average mass conservation residuals for each scale level, where: For the first Area of ​​each grid cell:

[0152] ,in, For the first The average quality-conserving residual within the class scale level, where s is the scale level index, s ∈ {small, medium, large}. Let be the number of grid cells of type s. For grid cell indexing, For those belonging to the first Summing all grid cells at the class-scale level, For the first The mass conservation residual of each grid cell For the first Mass conservation residual of each grid cell The absolute value;

[0153] Define the scale-adaptive weighting coefficients: ,in, For the first Adaptive weighting coefficients for class scale levels are used to amplify or reduce the contribution of that scale level to the loss function. To maximize the function, the weight coefficients must be at least 0.1 to prevent them from being too small and causing the scale level to be ignored. To minimize the function, the weight coefficients are set to no higher than 3.0 to prevent instability during training due to excessively large values. For the first Average quality-conserving residuals within the class-scale level The global average residual;

[0154] Construct a multi-scale weighted mass conservation loss function: ,in, The multi-scale weighted mass conservation loss function value is used to replace the original Participate in model training, This represents the total number of grid cells, with a value of 2680. To sum over all grid cells, For grid Belonging to the scale level The corresponding adaptive weight coefficients, For the first Mass conservation residual of each grid cell The square of , s(i) is the scale level to which grid i belongs; the physical regularization coefficient λ mass During training, adjustments are made dynamically according to the following rules: the validation set is computed every 10 epochs. If the loss does not decrease after three consecutive assessments, then λ will be... mass Multiply by 1.2; if Then λ mass Multiply by 0.8; λ mass The adjustment range is limited to the interval [0.01, 0.5].

[0155] The purpose of step S3 is to overcome the deficiency of insufficient constraint force of global water balance constraint on local areas, refine the physical constraint to each computational grid cell, and introduce a flow flux post-processing correction algorithm to strictly ensure the mass conservation consistency between water level and flow field.

[0156] The purpose of step S33 is to use the idea of ​​the pressure Poisson equation to correct the original flux field predicted by the neural network, so that the corrected flux field strictly satisfies the continuity equation of each grid cell, providing a physically self-consistent flow field input for subsequent water quality simulation.

[0157] S4. Perform simultaneous and rapid prediction of multiple parameters of the hydrodynamic field.

[0158] In practice, step S4 further includes the following steps:

[0159] S41. Adjust the network input / output structure.

[0160] The input layer is expanded to simultaneously contain water level sequences from N = 8 historical time steps and current flow rate data. The input tensor shape is [8, 14], where 14 corresponds to twice the feature dimension (water level and flow rate) of the 7 boundaries. The output layer synchronously outputs N... cell = 2680 grid cells' center water level values ​​and N_face = 4988 grid faces' flow flux values, with a total output node count of 7668.

[0161] The purpose of step S41 is to: expand the input layer to simultaneously include historical water level sequences and current flow data, thereby enhancing the model's ability to perceive flow information; and expand the output layer to simultaneously output the water level field and the flow flux field.

[0162] S42. Design a joint loss function for multiple tasks.

[0163] The data-driven term is expanded to a weighted sum of water level error and flow rate error:

[0164] .

[0165] in: This is the flow rate normalization factor, in m³ / s. For the expanded data-driven item, unit: m². The total number of grid cells, dimensionless. , For grid indexing, dimensionless. , For the first Water level prediction values ​​for each grid cell, in meters. For the first Actual water level values ​​for each grid cell, unit: meters. The weighting coefficient for the flow error term is dimensionless. , The total number of grid faces is dimensionless. , For grid surface indexing, dimensionless. , For the first Predicted flow flux values ​​for each grid surface, in m³ / s. For the first The actual flow flux of each grid surface, in m³ / s.

[0166] The joint loss function is: .in: The value represents the joint loss function, in m². For the expanded data-driven item, unit: m². The coefficients of the physical regularization term at the grid scale are dimensionless. , This represents the mass conservation loss function value at the global grid scale, in meters. 2 / s 2 .

[0167] The purpose of step S42 is to expand the data-driven term into a weighted sum of water level error and flow rate error, which together with the grid-scale mass conservation loss constitutes a joint loss function, ensuring that both prediction tasks can be fully optimized during joint training.

[0168] S43. Perform layer-by-layer pre-training and model fine-tuning.

[0169] First, the model is trained to convergence using water level prediction as a single task, with the network parameters of the water level output branch fixed. Then, a flow prediction branch is added, and the hidden layer features of the water level output branch are used as the shared input of the flow prediction branch for joint fine-tuning. The learning rate during the fine-tuning phase is set to 0.1 times the initial learning rate, i.e., 0.0001, and training is performed for 200 epochs. After the model outputs, post-processing correction S33 is applied to the flow flux field.

[0170] The purpose of step S43 is to adopt a phased training strategy of first water level and then flow rate, to first establish a good feature representation using the water level prediction task, and then transfer it to the flow rate prediction task, thereby reducing the difficulty of multi-task joint training.

[0171] The purpose of step S4 is to extend the function of the surrogate model from single water level prediction to simultaneous prediction of water level field and flow flux field, so as to achieve complete and rapid output of hydrodynamic conditions.

[0172] S5. Based on the hydrodynamic proxy model, perform rapid solution for water quality transport and overall performance evaluation.

[0173] In practice, step S5 further includes the following steps:

[0174] S51. Design of a solver for the convection-diffusion equation.

[0175] The water transport process is described using a two-dimensional convection-diffusion equation: .in: Water depth, unit: meters. Pollutant concentration, unit: mg / L For time variables, the unit is seconds. For concentration flux Regarding time The partial derivatives, in mg / (L·s), This is the advection term, representing the transport flux of pollutants along the water flow, measured in mg / (L·s). This is a velocity vector, with units of m / s. This is the diffusion term, representing the diffusion flux of pollutants driven by the concentration gradient, in mg / (L·s). The diffusion coefficient, in m² / s, ranges from 0.01 to 1.0 m² / s. Source and sink terms, unit: mg / (L·s).

[0176] The solution is obtained by explicit discretization using the finite volume method, with a time step of [missing information]. = 1800 s satisfies the CFL stability condition: .in: The time step for water quality analysis is expressed in seconds. s, The grid scale is in meters (m), and its value ranges from 200 to 500 m. The velocity is expressed in m / s. The diffusion coefficient is expressed in m² / s. For reference time step, unit: seconds.

[0177] The purpose of step S51 is to establish a mathematical model describing the transport and diffusion process of pollutants, link the hydrodynamic field with the water quality concentration field, and realize the connection between the physical processes of hydrodynamics and water quality.

[0178] S52, Execute the hydrodynamic-water quality module coupling strategy.

[0179] Within one hydrodynamic time step Δt = 21600 s, the water quality solver performs 12 sub-step iterations (sub-step time step size 1800 s), and the hydrodynamic field is linearly interpolated between sub-steps.

[0180] The purpose of step S52 is to adopt a modular decoupling and coupling strategy, whereby the hydrodynamic field is predicted and stored once by the proxy model, and the water quality solver directly reads and drives the water quality simulation, so that the hydrodynamic field can drive the water quality simulation of multiple scenarios after being predicted once, which greatly improves the computational efficiency.

[0181] S53. Construct a rapid simulation model of hydrodynamics and water quality and evaluate its overall performance.

[0182] The specific assessment content includes:

[0183] (1) Water level prediction accuracy assessment: The root mean square error (RMSE) is used to measure the deviation between the water level predicted by the surrogate model and the water level simulated by the high-fidelity numerical model.

[0184] .

[0185] in, This represents the predicted water level (in meters) for the i-th grid and the n-th time step. This corresponds to the actual water level value (unit: m).

[0186] (2) Water quality prediction accuracy assessment: The accuracy of water quality concentration prediction is measured using the mean relative error (MRE). .in: The average relative error is expressed in %. The total number of grid cells, dimensionless. The number of time steps is dimensionless. For grid indexing, dimensionless. , For time step indexing, dimensionless. , For the first The grid, the first Predicted concentration values ​​at each time step, in mg / L. For the first The grid, the first The actual concentration values ​​at each time step, in mg / L. It is a small constant, with units of mg / L, used to prevent division by zero.

[0187] (3) Evaluation of computation time: Comparing the computation time of the fast simulation model and the TUWMM high-fidelity numerical model under the same simulation task, the improvement rate of computation speed is:

[0188] .in, The computation time (in seconds) for the TUWMM high-fidelity numerical model. The computation time (in seconds) for quickly simulating the model.

[0189] Through the horizontal comparison across the above three dimensions, a comprehensive evaluation of the overall performance of the model is achieved. In the application test of the lower lakes of the South Four Lakes, the method of this invention improved the calculation speed by 92.2% to 97.6% while maintaining comparable accuracy to the numerical model, verifying the technical superiority of this invention.

[0190] The purpose of step S53 is to establish a complete hydrodynamic and water quality rapid simulation model, and to conduct rapid simulation of the spatiotemporal evolution of the water level field, flow flux field, and water quality element concentration field with high fidelity; and to quantitatively evaluate the model performance from multiple dimensions by comparing it with the simulation results of the TUWMM high fidelity numerical model, thereby verifying the technical advantages of the present invention.

[0191] In specific implementation, the rapid solution for water quality transport based on the hydrodynamic proxy model in S5 also includes a method for accelerating the solution for pollutant transport based on an adaptive time-step splitting operator, specifically including:

[0192] The two-dimensional convection-diffusion equation is split into two successive stages: the convection substep and the diffusion substep.

[0193] Convection step: The convection terms are discretized using a second-order TVD (Total Variation Diminishing) scheme, with a critical time step of [missing information]. Satisfying CFL conditions:

[0194] Where Δx is the grid scale (unit: m). = 10 -6 To prevent division by zero for small constants;

[0195] Diffusion substep: The diffusion term is discrete using an implicit scheme, with a critical time step of [missing information]. satisfy: ;

[0196] The adaptive global time step takes the minimum of the two values: Where β = 0.8 is the safety factor;

[0197] An adaptive time step is used within each hydrodynamic time step Δt = 21600 s. The process involves sub-loop propagation, with linear interpolation of the hydrodynamic field between adjacent sub-steps; when the average magnitude of the pollutant concentration gradient... mg·L -1 ·m -1 When the time step is reached, it will automatically increase to 1.5 times the current value, with a maximum not exceeding Δt. adaptive_max = 900s.

[0198] Step S5 serves to: utilize the hydrodynamic field rapidly predicted in S4 to drive the solution of the convection-diffusion equations, achieving near real-time prediction of the spatiotemporal evolution of water quality parameters; simultaneously, establish a rapid hydrodynamic and water quality simulation model, and evaluate the overall performance of the model from three dimensions: water level prediction accuracy, water quality prediction accuracy, and computation time, through horizontal comparison with a high-fidelity numerical model. A modular decoupling strategy is adopted to decouple hydrodynamic prediction from water quality solution.

[0199] S6. Implement differentiated emergency regulation for water diversion periods and non-water diversion periods.

[0200] In specific implementation, step S6 further includes the following steps:

[0201] S61. Conduct a comparative analysis of pollution diffusion characteristics between the water diversion period and the non-water diversion period.

[0202] Two comparative operating conditions were set up: one during the water diversion period and the other during the non-water diversion period. A sudden pollution accident was assumed to occur on the water diversion main line 2 km upstream of the secondary dam, with instantaneous COD emissions. Mn 500 kg. The S5 rapid water quality simulation method was used to output COD_Mn concentration fields at 1 h, 6 h, 12 h, 24 h, and 48 h after the accident. During the water diversion period, the pollution plume migrated downstream towards Hanzhuang Sluice Gate along the main water diversion line, with an average migration velocity of approximately 0.08 m / s; during the non-water diversion period, the pollution plume diffused outwards in an approximately circular pattern under the influence of local circulation and turbulent diffusion.

[0203] The purpose of step S61 is to reveal the control effect of different flow field structures on the transport and diffusion of contaminant plumes, and to provide a scientific basis for the formulation of differentiated strategies.

[0204] S62. Develop differentiated emergency response and control strategies.

[0205] The following emergency strategies were formulated for the water diversion period: the secondary dam gates were closed to cut off the hydraulic channel for the pollution plume to move downstream; two physical barriers were set up 0.5 km and 1.0 km downstream of the accident point, and activated carbon adsorption barriers were set up at key sections of the main channel; high-concentration sewage was pumped to shore treatment facilities or treated in situ by adding chemical oxidants within the enclosure and control area.

[0206] The following emergency strategies are formulated for non-water diversion periods: If the downstream river channel has the capacity to receive pollution, the discharge from the upper lake and the outlet gates of the lower lake will be coordinated to increase the flow velocity in the lake area from 0.01-0.03 m / s to 0.10-0.15 m / s, thereby quickly pushing the pollution plume out of the lake area; If the downstream is a sensitive water area, the outlet gates will be closed to retain the pollution plume in the middle of the lake, and the lake's environmental capacity and turbulent diffusion will be used to dilute and degrade it.

[0207] The purpose of step S62 is to match the control measures with the hydrological conditions and improve the effectiveness of the emergency response.

[0208] S63. Conduct strategy effectiveness evaluation and dynamic optimization.

[0209] Quantify the control measures as model boundary conditions or parameter modifications, and rerun the S5 rapid water quality simulation method. Calculate the following quantitative indicators:

[0210] Peak concentration reduction rate: .in, Peak concentration reduction rate, in %. Peak concentration (mg / L) without any control measures. Peak concentration after control measures were implemented (unit: mg / L).

[0211] Pollution duration reduction rate: ,in, The percentage reduction in pollution duration is expressed as a percentage. The duration of pollution without any control measures (in seconds). The duration of pollution after control measures were implemented (in seconds).

[0212] Reduction rate of impact area: .in, The affected area reduction rate is expressed in %. The pollution impact area (unit: m²) is the area without any control measures. The extent of pollution impact after control measures are implemented (unit: m²).

[0213] If the strategy does not achieve the expected results, adjust the location and density of the enclosure, adjust the opening and closing sequence of the gates, and optimize the amount and location of the adsorbent material. Each iteration of the simulation takes 5 to 10 seconds, and the optimal combination of control strategies can be quickly output.

[0214] The purpose of step S63 is to quantitatively evaluate the effectiveness of the strategy, support rapid iterative optimization, and output the optimal combination of control strategies.

[0215] In specific implementation, step S6 also includes step S64, an adaptive enclosure layout optimization method based on pollutant concentration gradients, specifically including:

[0216] Based on the pollutant concentration field predicted in step S5, the concentration gradient along the mainstream direction of each grid cell is calculated: ,in, The concentration gradient vector at position (x,y) (unit: ), and These are the partial derivatives of the concentration in the x and y directions, respectively. and These are the unit vectors in the x and y directions, respectively;

[0217] Calculate the magnitude of the concentration gradient: ;

[0218] The concentration gradient magnitude exceeds the threshold θ grad = 0.05 The grid cells were identified as pollution front areas;

[0219] Within the pollution front area, the center point of the grid unit is used as the candidate deployment point. The k-means clustering algorithm is used to cluster the candidate points, where k = 3 to 8. The cluster center is used as the actual deployment location of the enclosure.

[0220] Enclosure density Positively correlated with the magnitude of the local concentration gradient:

[0221] ,in, For position Density of enclosure layout at the location = 0.5 lanes / km is the minimum deployment density. = 3.0 lanes / km is the maximum deployment density. For position Concentration gradient modulus at that location, This represents the maximum value of the concentration gradient modulus within the entire pollution front area; based on the adaptive deployment results, a containment deployment scheme is generated, and the strategy effectiveness is evaluated.

[0222] The purpose of step S6 is to formulate differentiated emergency control strategies for different hydrological conditions based on the spatiotemporal evolution characteristics of pollutants predicted by the rapid simulation method, and to support the rapid iterative evaluation and optimization of the strategy effects.

[0223] Regarding the hyperparameter settings of the hydrodynamic proxy model architecture, apart from the input layer, output layer, and improved loss function adjusted for the new task requirements, the optimal hyperparameter configurations for hidden layer depth, number of neurons per layer, and training settings were determined through trial and error. For the selection strategy of time parameters, a parameter optimization strategy with a clear physical mechanism was followed, employing the optimal combination of Δt = 21600 s and N = 8, thus ensuring that the model can fully capture hydrodynamic characteristics in the time dimension.

[0224] Figure 2 For 2025, a certain East Railway Station COD Mn A diagram showing the comparison of predicted concentrations. (Example) Figure 2 As shown, the COD predicted by the modelMn The concentration curves and measured scatter plots showed good agreement, with a mean relative error (MRE) of only 14.70%. In terms of temporal evolution, the model accurately reflects the abrupt changes in water quality that occur during the transition between the water diversion period and the non-water diversion period: during the water diversion period from January to May, the lake's water quality remained at a low level (approximately 4 mg / L) due to the continuous replenishment and dilution effect of high-quality Yangtze River water from the South-to-North Water Diversion Project; however, during the non-water diversion period and flood season after June, with the cessation of water diversion and the influx of upstream tributary runoff, the predicted concentration rapidly increased to the 6–8 mg / L range.

[0225] Figure 3 This diagram illustrates the evolution of the proportion of contaminated area over time under different operating conditions after adopting this method. Figure 3 As shown, the dynamic evolution of the proportion of polluted area over time under two sudden operating conditions is intuitively demonstrated.

[0226] This invention can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0227] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0228] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0229] Example 2

[0230] Further reference Figure 4 As a response to the above Figure 1 The present invention provides an embodiment of a device for rapid simulation and emergency control of hydrodynamic water quality that integrates artificial intelligence, based on the method shown. This embodiment of the device is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0231] like Figure 4 As shown, the rapid simulation and emergency control device 70 for hydrodynamics and water quality integrating artificial intelligence described in this embodiment includes: a data module 71, a construction module 72, a correction module 73, a prediction module 74, a solution module 75, and a control module 76. Wherein:

[0232] Data module 71 is used to construct a training dataset based on historical hydrological sequences;

[0233] Module 72 is used to construct a hydrodynamic proxy model based on a physical information neural network: a fully connected feedforward neural network is constructed, which takes historical boundary time series data as input and calculates the water level value of the grid cell as output; a loss function containing data-driven terms and global water balance regularization terms is constructed; and an optimizer is used to train the network to obtain the hydrodynamic proxy model.

[0234] The correction module 73 is used to implement grid-scale mass conservation constraints and flow flux correction: it constructs a mass conservation loss function focused on each computational grid cell and performs refined training on the neural network; based on the idea of ​​the pressure Poisson equation, it performs post-processing correction on the predicted flow flux field to ensure that it strictly satisfies the continuity equation of each grid cell.

[0235] Prediction module 74 is used to perform multi-parameter synchronous and rapid prediction of hydrodynamic field: adjust the neural network structure, take historical water level sequence and current flow data as input, synchronously output the center water level value of all grid cells and the flow flux value of all grid surfaces, and use multi-task joint loss function and layer-by-layer pre-training strategy to train and fine-tune the model.

[0236] The solver module 75 is used for rapid solution and overall performance evaluation of water quality transport based on a hydrodynamic surrogate model: it solves the two-dimensional convection-diffusion equations based on the flow flux field and water level field; it uses the finite volume method explicit scheme for discretization, and the hydrodynamic field is linearly interpolated between substeps; the hydrodynamic surrogate model and the water quality solver adopt a modular decoupling coupling strategy; it establishes a rapid hydrodynamic and water quality simulation model to carry out high-fidelity rapid simulation of the spatiotemporal evolution of the water level field, flow flux field, and water quality element concentration field; it conducts a horizontal comparison of the simulation results of the rapid simulation model and the TUWMM high-fidelity numerical model, including performance indicators in three dimensions: water level prediction accuracy, water quality prediction accuracy, and computation time, to achieve an overall performance evaluation of the model;

[0237] The control module 76 is used to implement differentiated emergency control for water transfer periods and non-water transfer periods: using the spatiotemporal evolution characteristics of pollutants predicted by S5, it formulates strategies for shutting down gate pumps and physical interception during water transfer periods, and formulates strategies for joint scheduling to increase the downstream flow or shutting down the outlet gate for retention and dilution during non-water transfer periods, and uses rapid simulation capabilities to dynamically evaluate and iteratively optimize the effectiveness of the strategies.

[0238] Example 3

[0239] To address the aforementioned technical problems, embodiments of the present invention also provide a computer device. Please refer to [link / reference needed]. Figure 5 , Figure 5 This is a basic structural block diagram of the computer device in this embodiment.

[0240] The aforementioned computer device 8 includes a memory 81, a processor 82, and a network interface 83 that are interconnected via a system bus. It should be noted that only the computer device 8 with components 81, 82, and 83 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described herein is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0241] The aforementioned computer devices can be desktop computers, laptops, handheld computers, and cloud servers, among other computing devices. These devices can facilitate human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0242] The aforementioned memory 81 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the aforementioned memory 81 may be an internal storage unit of the aforementioned computer device 8, such as the hard disk or memory of the computer device 8. In other embodiments, the aforementioned memory 81 may also be an external storage device of the aforementioned computer device 8, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 8. Of course, the aforementioned memory 81 may also include both the internal storage unit and its external storage device of the aforementioned computer device 8. In this embodiment, the aforementioned memory 81 is typically used to store the operating system and various application software installed on the aforementioned computer device 8, such as computer-readable instructions for rapid simulation and emergency control methods of hydrodynamic water quality integrating artificial intelligence. In addition, the aforementioned memory 81 can also be used to temporarily store various types of data that have been output or will be output.

[0243] In some embodiments, the processor 82 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 82 is typically used to control the overall operation of the computer device 8. In this embodiment, the processor 82 is used to execute computer-readable instructions stored in the memory 81 or to process data, such as executing the computer-readable instructions of the aforementioned method for rapid simulation and emergency control of hydrodynamic and water quality incorporating artificial intelligence.

[0244] The network interface 83 may include a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the computer device 8 and other electronic devices.

[0245] The beneficial effects of implementing the above embodiments are as follows:

[0246] (1) It takes into account both physical reliability and computational efficiency: By combining physical information neural networks with high-precision numerical simulation, the data and physical mechanism dual-driven proxy model is constructed. While ensuring that the prediction accuracy is comparable to that of traditional numerical models, the computation speed is improved by 92.2% to 97.6%, achieving a leap in computational efficiency.

[0247] (2) Constructing a continuous, high-quality training dataset: Using a sliding window to generate input-output sample pairs provides a supervisory signal for the neural network model to learn the dynamic evolution of water level. It is a bridge connecting the original observation data and the physical information neural network model, and determines the basis of the model's predictive ability.

[0248] (3) It has strict physical consistency guarantee: By focusing the mass conservation constraint on the grid scale and using the flow flux post-processing correction algorithm based on the pressure Poisson equation, it ensures that each computational grid cell strictly satisfies the continuity equation, providing a high-fidelity hydrodynamic background field for water transport simulation.

[0249] (4) A modular decoupling and coupling strategy was adopted: By taking into account the essential differences in the propagation speed of hydrodynamic and water quality parameters, different prediction methods were adopted (surrogate model prediction of hydrodynamic field + analytical solution of convection-diffusion equation), which avoided the problem that a single neural network could not maintain physical self-consistency in long-term prediction.

[0250] (5) It has differentiated emergency control capabilities: It can quickly deduce the spatiotemporal evolution characteristics of pollution plumes under different working conditions, and formulate differentiated emergency control strategies for water diversion periods and non-water diversion periods. At the same time, it supports rapid iterative evaluation of the effectiveness of the strategies, providing scientific support and efficient tools for emergency decision-making in sudden water pollution events.

[0251] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0252] Obviously, the embodiments described above are merely some embodiments of the present invention, not all embodiments. The accompanying drawings show preferred embodiments of the present invention, but do not limit the patent scope of the present invention. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the patent protection scope of this invention.

Claims

1. A method for rapid simulation and emergency control of hydrodynamic water quality integrating artificial intelligence, characterized in that, Includes the following steps: S1. Construct a training dataset based on historical hydrological sequences; S2. Construct a hydrodynamic proxy model based on a physical information neural network: Construct a fully connected feedforward neural network, taking historical boundary time series data as input and calculating the water level value of the grid cell as output; construct a loss function that includes a data-driven term and a global water balance regularization term; train the network using an optimizer to obtain the hydrodynamic proxy model; S3. Implement grid-scale mass conservation constraints and flux correction: Construct a mass conservation loss function focused on each computational grid cell to refine the training of the neural network; Based on the idea of ​​the pressure Poisson equation, the predicted flow flux field is post-processed and corrected to ensure that it strictly satisfies the continuity equation of each grid cell. S4. Perform simultaneous and rapid prediction of multiple parameters of hydrodynamic field: Adjust the neural network structure, take historical water level sequence and current flow data as input, and output the center water level value of all grid cells and the flow flux value of all grid surfaces simultaneously. Use multi-task joint loss function and layer-by-layer pre-training strategy to train and fine-tune the model. S5. Based on the hydrodynamic proxy model, perform rapid solution for water transport and evaluate the overall performance of the model: Based on the hydrodynamic field, solve the two-dimensional convection-diffusion equation; The finite volume method is used for explicit discretization, with linear interpolation of the hydrodynamic field between substeps. The hydrodynamic surrogate model and the water quality solver adopt a modular decoupling coupling strategy. A rapid hydrodynamic and water quality simulation model with both physical interpretability and computational efficiency is established to conduct high-fidelity simulations of the spatiotemporal evolution of water level, flow flux, and water quality element concentration fields. A horizontal comparison is conducted between the simulation results of the rapid simulation model and the TUWMM high-fidelity numerical model, including performance indicators in three dimensions: water level prediction accuracy, water quality prediction accuracy, and computation time, to evaluate the overall performance of the model. S6. Implement differentiated emergency control measures for sudden pollution events during and outside the water transfer period: Utilize step S5 to conduct emergency scenario simulations of water quality elements under different operating conditions and predict the spatiotemporal evolution characteristics of different pollutants. For the water diversion period, control strategies such as shutting down gate pumps, physical interception, and adsorption barriers are formulated. For the non-water diversion period, control strategies such as joint scheduling of upstream and downstream gates to increase the discharge flow or closing the outlet gate to retain and dilute the water are formulated. The effects of different emergency control strategies are dynamically and quantitatively evaluated and iteratively optimized by utilizing the rapid simulation capability of hydrodynamics and water quality.

2. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, Step S1, which involves constructing a training dataset based on historical hydrological sequences, specifically includes the following steps: S11. Collect historical hydrological data and extract time series data; S12. Construct sliding window input-output sample pairs; S13. Data set partitioning and normalization.

3. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, Step S2, the step of constructing a hydrodynamic proxy model based on a physical information neural network, specifically includes: Constructing a hydrodynamic proxy model framework: A fully connected feedforward neural network is constructed, with N × C nodes in the input layer, where N = 8 is the number of input time steps, and C is the feature dimension of each time step. The hidden layer depth is 4 layers, with 128 neurons per layer. The activation function is the hyperbolic tangent function. ,in: Let be the value of the hyperbolic tangent function, dimensionless, and its range is . , For input values, It is a natural constant; The input data time interval is Δt = 21600 s, and the total time span of the input sequence is N × Δt = 48 hours; Construct a grid-scale mass conservation loss function; Efficient physical loss calculation based on OpenFOAM; Perform post-processing corrections on the flow flux field; The steps in step S3, which implement grid-scale mass conservation constraints and flux correction, specifically include: Constructing a grid-scale mass conservation loss function: Focusing physical constraints on each computational grid cell, the mass conservation equation for each grid cell i is discretized as follows: Where t is the current time, The water level at the center point of the grid. = 5400 s is the discrete time interval of the water level time gradient. For the grid projection area, Let j be the flow flux through the j-th boundary of the grid. The total number of grid boundaries; Define the mass-conserving residual for each grid cell: ; The mass conservation loss function at the global grid scale is: ,in, = 2680 is the total number of grid cells; The improved composite loss function is: ,in: The value of the improved composite loss function, in m². This is a data-driven term, specifically the mean square error between the predicted and actual values, expressed in m². The coefficients of the physical regularization term at the grid scale are dimensionless. , This represents the mass conservation loss function value at the global grid scale, in meters. 2 / s 2 ; Efficient physical loss calculation based on OpenFOAM: The predicted water level and flow flux output by the neural network are denormalized to restore the water level at the grid center point. and the flow flux through the j-th boundary of the grid The divergence operator is invoked to automatically sum and normalize the flux around each grid cell based on the Gaussian divergence theorem, yielding the net flux field. This net flux field is then added to the water level time gradient field to obtain the mass-conserved residual field, which is then used to calculate... ; Post-processing correction of flux field: Using the idea of ​​the pressure Poisson equation, the original flux field φ predicted by the neural network is regarded as a predicted value that does not strictly satisfy mass conservation, and a correction potential function is constructed. Poisson's equation: ,in: To correct the potential function Laplace operator, unit: s -1 , The corrected potential function, unit: m 2 / s, For the original flux field divergence, unit: s -1 , The original flux field predicted by the neural network, in meters. 3 / s; The linear system is solved using the conjugate gradient method combined with incomplete Cholesky preprocessing; the flux correction for each grid surface is calculated. ,in: For grid surface Flux correction, unit: m 3 / s, For grid surface Area, unit: m² For grid surface One side of the grid center Potential function value, unit: m 2 / s, For grid surface The center of the grid on the other side Potential function value, unit: m 2 / s, For grid center and The distance between them, in meters; Corrected flux field ,in: For the corrected flux field on the grid surface The value above satisfies the law of conservation of mass, with units of m³ / s. The original flux field predicted by the neural network on the grid surface Values ​​above, unit: m 3 / s, For grid surface Flux correction, unit: m 3 / s.

4. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, The step of performing simultaneous and rapid prediction of multiple parameters of the hydrodynamic field in step S4 specifically includes: S41. Network Input / Output Structure Adjustment: The input layer is expanded to simultaneously contain water level sequences of N = 8 historical time steps and current flow data. The input tensor shape is [8, 14], where 14 corresponds to twice the feature dimension of the 7 boundaries; the output layer synchronously outputs N... cell = Center water level values ​​of 2680 grid cells and N face = Flux values ​​for 4988 grid surfaces, with a total of 7668 output nodes; S42. Design of Multi-Task Joint Loss Function: The data-driven term is expanded to a weighted sum of water level error and flow rate error. ,in: This is the flow rate normalization factor, in m³ / s. For the expanded data-driven item, unit: m². The total number of grid cells, dimensionless. , For grid indexing, , For the first Water level prediction values ​​for each grid cell, in meters. For the first Actual water level values ​​for each grid cell, unit: meters. The weighting coefficient for the flow error term. , This represents the total number of grid faces. , For grid face indexing, , For the first Predicted flow flux values ​​for each grid surface, in m³. 3 / s, For the first The actual flow flux value of each grid surface, in m³. 3 / s; The joint loss function is: ,in: The value represents the joint loss function, in m². For the expanded data-driven item, unit: m 2 , For the physical regularization coefficients at the grid scale, , This represents the mass conservation loss function value at the global grid scale, in meters. 2 / s 2 ; S43. Layer-by-layer pre-training and model fine-tuning: First, train the model to convergence using water level prediction as a single task, and fix the network parameters of the water level output branch; then add a flow prediction branch, and use the hidden layer features of the water level output branch as the shared input of the flow prediction branch for joint fine-tuning; the learning rate in the fine-tuning stage is set to 0.1 times the initial learning rate, and the training is carried out for 200 epochs; after the model outputs, post-processing correction is performed on the flow flux field.

5. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, The steps in S5 for rapidly solving water transport and evaluating the overall performance of the model based on the hydrodynamic proxy model specifically include: The water transport process is described using a two-dimensional convection-diffusion equation: ,in: Water depth, unit: meters. Pollutant concentration, unit: mg / L For time variables, the unit is seconds. For concentration flux Regarding time The partial derivatives, in mg / (L·s), This is the advection term, representing the transport flux of pollutants along the water flow, measured in mg / (L·s). This is a velocity vector, with units of m / s. This is the diffusion term, representing the diffusion flux of pollutants driven by the concentration gradient, in mg / (L·s). The diffusion coefficient, in m² / s, ranges from 0.01 to 1.0 m² / s. Source and sink terms, unit: mg / (L·s); ,in: Source and sink terms, unit: mg / (L·s) Attenuation coefficient, unit: s -1 , Pollutant concentration, unit: mg / L; The solution is obtained by explicit discretization using the finite volume method, with a time step of [missing information]. = 1800 s satisfies the CFL stability condition; a modular decoupling coupling strategy is adopted to decouple hydrodynamic prediction from water quality solution. Within one hydrodynamic time step Δt = 21600 s, the water quality solver performs 12 sub-step iterations, and the hydrodynamic field is linearly interpolated between sub-steps. Establish a rapid simulation model for hydrodynamics and water quality, and conduct rapid simulation of the spatiotemporal evolution of high-fidelity water level field, flow flux field and water quality element concentration field. A horizontal comparison was conducted between the simulation results of the rapid simulation model and the TUWMM high-fidelity numerical model, including performance indicators in three dimensions: water level prediction accuracy (measured by root mean square error RMSE), water quality prediction accuracy (measured by mean relative error MRE), and computation time, to evaluate the overall performance of the model.

6. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, The specific steps in S6 for implementing differentiated emergency regulation for water transfer periods and non-water transfer periods include: S61. Comparative analysis of pollution diffusion characteristics between water diversion period and non-water diversion period: Set up two comparative working conditions, water diversion period and non-water diversion period, run the rapid water quality simulation method, output the concentration field at different times after the accident, and track the center location and diffusion range of the pollution plume. S62. Differentiated emergency control strategies: For the water diversion period, develop emergency strategies including closing the secondary dam gates, deploying physical barriers and adsorption barriers downstream of the accident point, and carrying out in-situ treatment; for the non-water diversion period, develop emergency strategies such as joint scheduling to increase the discharge flow to quickly push the pollution plume out of the lake area, or closing the outlet gates to retain the pollution plume in the middle of the lake for dilution and degradation. S63. Strategy Effectiveness Evaluation and Dynamic Optimization: Quantify the control measures into model boundary conditions or parameter modifications, and rerun the rapid water quality simulation method; calculate the peak concentration reduction rate. Pollution duration reduction rate Scope of influence reduction rate If the strategy does not achieve the expected results, adjust the location and density of the enclosure, adjust the opening and closing sequence of the gate, and optimize the amount and location of the adsorption material. Each iteration simulation takes 5 to 10 seconds, and the optimal combination of control strategies is output.

7. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, The S6 implementation of differentiated emergency control for water diversion periods and non-water diversion periods also includes an adaptive enclosure layout optimization method based on pollutant concentration gradients, specifically including: Based on the pollutant concentration field predicted by S5, the concentration gradient along the mainstream direction of each grid cell is calculated: ,in, The concentration gradient vector at position (x,y) (unit: ), and These are the partial derivatives of the concentration in the x and y directions, respectively. and These are the unit vectors in the x and y directions, respectively; Calculate the magnitude of the concentration gradient: ; The concentration gradient magnitude exceeds the threshold θ grad = 0.05 The grid cells were identified as pollution front areas; Within the pollution front area, the center point of the grid unit is used as the candidate deployment point. The k-means clustering algorithm is used to cluster the candidate points, where k = 3 to 8. The cluster center is used as the actual deployment location of the enclosure. Enclosure density Positively correlated with the magnitude of the local concentration gradient: ,in, For position Density of enclosure layout at the location = 0.5 lanes / km is the minimum deployment density. = 3.0 lanes / km is the maximum deployment density. For position Concentration gradient modulus at that location, This represents the maximum value of the concentration gradient modulus within the entire pollution front area; based on the adaptive deployment results, a containment deployment scheme is generated, and the strategy effectiveness is evaluated.

8. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, The rapid solution for water quality transport based on the hydrodynamic surrogate model in S5 also includes a method for accelerating the solution for pollutant transport based on an adaptive time-step splitting operator, specifically including: The two-dimensional convection-diffusion equation is split into two successive stages: the convection substep and the diffusion substep. Convection step: The convection terms are discretized using a second-order TVD (Total Variation Diminishing) scheme, with a critical time step of [missing information]. Satisfying CFL conditions: Where Δx is the grid scale (unit: m). = 10 -6 To prevent division by zero for small constants; Diffusion substep: The diffusion term is discrete using an implicit scheme, with a critical time step of [missing information]. satisfy: ; The adaptive global time step takes the minimum of the two values: Where β=0.8 is the safety factor; An adaptive time step is used within each hydrodynamic time step Δt = 21600 s. The process involves sub-loop propagation, with linear interpolation of the hydrodynamic field between adjacent sub-steps; when the average magnitude of the pollutant concentration gradient... mg·L -1 ·m -1 When the time step is reached, it will automatically increase to 1.5 times the current value, with a maximum not exceeding Δt. adaptive_max = 900 s.

9. The method for rapid simulation and emergency control of hydrodynamics and water quality integrating artificial intelligence according to claim 1, characterized in that, Step S3 implements grid-scale mass conservation constraints and flow flux correction, and also includes a dynamic adjustment method for physical regularization weights based on multi-scale residual adaptive adjustment, specifically including: The grid cells are divided into three scale levels based on their area: small-scale grid. <5000m 2 5000 m mesoscale grid 2 ≤ < 20000 m 2 Large-scale grids ≥ 20000 m 2 ; Calculate the average mass conservation residuals for each scale level, where: For the first Area of ​​each grid cell: ,in, For the first The average quality-conserving residual within the class scale level, where s is the scale level index, s ∈ {small, medium, large}. Let be the number of grid cells of type s. For grid cell indexing, For those belonging to the first Summing all grid cells at the class-scale level, For the first The mass conservation residual of each grid cell For the first Mass conservation residual of each grid cell The absolute value; Define the scale-adaptive weighting coefficients: ,in, For the first Adaptive weighting coefficients for class scale levels are used to amplify or reduce the contribution of that scale level to the loss function. To maximize the function, the weight coefficients must be at least 0.1 to prevent them from being too small and causing the scale level to be ignored. To minimize the function, the weight coefficients are set to no higher than 3.0 to prevent instability during training due to excessively large values. For the first Average quality-conserving residuals within the class-scale level The global average residual; Construct a multi-scale weighted mass conservation loss function: ,in, The multi-scale weighted mass conservation loss function value is used to replace the original Participate in model training, This represents the total number of grid cells, with a value of 2680. To sum over all grid cells, For grid Belonging to the scale level The corresponding adaptive weight coefficients, For the first Mass conservation residual of each grid cell The square of , s(i) is the scale level to which grid i belongs; the physical regularization coefficient λ mass During training, adjustments are made dynamically according to the following rules: the validation set is computed every 10 epochs. If the loss does not decrease after three consecutive assessments, then λ will be... mass Multiply by 1.2; if Then λ mass Multiply by 0.8; λ mass The adjustment range is limited to the interval [0.01, 0.5].

10. A device for rapid simulation and emergency control of hydrodynamic water quality integrating artificial intelligence, characterized in that, include: The data module is used to construct training datasets based on historical hydrological sequences; The module is used to construct a hydrodynamic proxy model based on a physical information neural network: a fully connected feedforward neural network is constructed, which takes historical boundary time series data as input and calculates the water level value of the grid cell as output; a loss function containing data-driven terms and global water balance regularization terms is constructed; and an optimizer is used to train the network to obtain the hydrodynamic proxy model. The correction module is used to implement grid-scale mass conservation constraints and flux correction: it constructs a mass conservation loss function focused on each computational grid cell and performs refined training on the neural network; Based on the idea of ​​the pressure Poisson equation, the predicted flow flux field is post-processed and corrected to ensure that it strictly satisfies the continuity equation of each grid cell. The prediction module is used to perform synchronous and rapid prediction of multiple parameters of the hydrodynamic field: the neural network structure is adjusted, and the historical water level sequence and the current flow data are used as inputs to synchronously output the center water level value of all grid cells and the flow flux value of all grid surfaces. The model is trained and fine-tuned by using a multi-task joint loss function and a layer-by-layer pre-training strategy. The solver module is used for rapid solution of water transport and overall performance evaluation based on the hydrodynamic proxy model: solving the two-dimensional convection-diffusion equation based on the flow flux field and the water level field; The finite volume method is used for explicit discretization, and the hydrodynamic field is linearly interpolated between substeps. The hydrodynamic surrogate model and the water quality solver adopt a modular decoupling coupling strategy. A rapid hydrodynamic and water quality simulation model is established to conduct high-fidelity simulations of the spatiotemporal evolution of water level field, flow flux field, and water quality element concentration field. A horizontal comparison of the simulation results of the rapid simulation model and the TUWMM high-fidelity numerical model is carried out, including performance indicators in three dimensions: water level prediction accuracy, water quality prediction accuracy, and computation time, to evaluate the overall performance of the model. The control module is used to implement differentiated emergency control for water diversion periods and non-water diversion periods: using the spatiotemporal evolution characteristics of pollutants predicted by S5, it formulates strategies for shutting down gate pumps and physical interception during water diversion periods, and formulates strategies for joint scheduling to increase the downstream flow or shutting down the outlet gate for retention and dilution during non-water diversion periods, and uses rapid simulation capabilities to dynamically evaluate and iteratively optimize the effectiveness of the strategies.

Images