Coupled model based method for predicting river hydrodynamics and pollutant transport

Abstract

The application discloses a kind of river hydrodynamics and pollutant migration collaborative prediction method based on coupling model, step 1, the MIKE21 hydrodynamic field of river channel is constructed;Step 2, construct pollutant migration model;Step 3: construct LSTM-MIKE 21 coupling model;Step 4, real-time prediction is carried out to the whole basin pollutant concentration field by coupling model.The application significantly improves the accuracy and reliability of pollutant migration simulation by bidirectional coupling of physical mechanism model MIKE 21 and data-driven model LSTM21, realizes the collaborative optimization solution of hydrodynamic field and pollutant concentration field.LSTM21 residual correction mechanism effectively compensates the complex process not covered by the pure physical model, reduces the pollutant peak concentration prediction error from more than 20% of the traditional method to less than 8%, and the compliance of mass conservation and non-negative concentration is more than 99.5%.

Description

Technical Field

[0001] This invention belongs to the field of water environment simulation and ecological governance technology, specifically involving a method for synergistic prediction of river hydrodynamics and pollutant migration based on a coupled model. Background Technology

[0002] With the increasing demands for refined water environment management, accurate simulation of river hydrodynamic processes and pollutant migration and diffusion is crucial for water quality early warning, pollution prevention and control, and the formulation of ecological restoration strategies. Currently, this field mainly relies on two types of mainstream methods: pure mechanistic models (such as MIKE 21) and pure data-driven machine learning models (such as LSTM), but both have significant limitations. Pure mechanistic models strictly rely on precise physical governing equations such as the Navier-Stokes equations and convection-diffusion equations. Their limitations include: low computational efficiency (often on the order of hours), high dependence on manual calibration of a large number of parameters such as the Manning coefficient and diffusion coefficient, and the influence of parameter uncertainty, often resulting in a mass conservation error rate as high as 5%-10%, affecting the reliability of the results. Pure data-driven machine learning models, on the other hand, directly utilize historical monitoring data for training, avoiding the explicit expression of physical mechanisms. The model output is prone to violating physical constraints (such as generating negative concentrations and mass non-conservation), leading to distorted results and a lack of interpretability. When monitoring data is scarce, the generalization ability drops sharply, limiting its practicality. In summary, the core dilemma of existing technologies lies in the fragmented application of mechanistic models and purely data-driven machine learning models. This results in an inability to simultaneously achieve efficient computing power, strict compliance with physical laws, and strong adaptability to sudden pollution scenarios, which severely restricts the precise management of river water environment and the effectiveness of emergency response, becoming a technological bottleneck that urgently needs to be overcome. Summary of the Invention

[0003] To achieve these objectives and other advantages according to the present invention, a method for synergistic prediction of river hydrodynamics and pollutant migration based on a coupled model is provided, comprising the following steps:

[0004] Step 1: Construct the MIKE21 hydrodynamic field of the river channel;

[0005] Step 2: Construct a pollutant migration model based on the MIKE21 hydrodynamic field;

[0006] Step 3: Couple the physical mechanism model MIKE 21 bidirectionally with LSTM to construct the LSTM-MIKE 21 coupled model;

[0007] Step 4: Real-time prediction of pollutant concentration field across the entire river basin using the LSTM-MIKE 21 coupled model.

[0008] Preferably, step 1 includes:

[0009] Step 1.1: Mesh Analysis and Terrain Import

[0010] Data on river velocity, flow rate, water depth, and elevation were collected. The irregular triangular mesh of the MIKE21 software was used to divide the data into grids. Based on satellite elevation data, underwater topographic data obtained from satellite data and hydrological stations were mapped to grid nodes using a weighted inverse distance interpolation method to generate a three-dimensional topographic model of the river.

[0011] The weighted inverse distance interpolation method is as follows:

[0012]

[0013] in, Zi represents the elevation of the node to be interpolated, and d represents the elevation of the node to be interpolated. k p = 2, where p is the distance between adjacent elevation points; N is the number of monitoring points.

[0014] Step 1.2: Constructing the hydrodynamic field

[0015] A two-dimensional hydrodynamic model is established based on the Navier-Stokes equations, coupled with the following environmental dynamic factors:

[0016]

[0017] Where h: water depth (m), v = (u, v): velocity vector (m / s), u is the longitudinal velocity, v is the transverse velocity, and v t : Real-time flow velocity; η: Water level (m); g: Acceleration due to gravity, g = 9.81 m / s² 2 Dry and wet treatment: water depth ≤ 0.01m is defined as dry cell mesh.

[0018] Preferably, step 2 includes:

[0019] Based on the pollutant kinetic model established in step 1, daily monitoring data from the monitoring station were imported into the MIKE 21 software, and the following pollutant migration model was established according to the mathematical relationships between various water quality indicators:

[0020]

[0021] Where C: pollutant concentration (mg / L), Δ: grid scale, D: diffusion tensor, unit m 2 / s, including turbulent diffusion and molecular diffusion, S 反应 =-K d hC: Primary degradation term, K d Degradation rate constant, in d -1 S 外源 Point source / area source emissions, in g / s; diffusion coefficient D: using the Smagorinsky subgrid model, Cs = 0.2;

[0022] D = (C s Δ) 2 |S|,

[0023] The simulated pollutant concentration C can be obtained based on the watershed pollutant migration model. MIKE .

[0024] Preferably, in step 3, an LSTM neural network is used to predict the δC value of the simulated river section to establish a residual model; where δC = C 实测 -C MIKE C 实测 These are the measured values ​​of pollutant concentrations.

[0025] The steps for constructing a residual model include:

[0026] Interpolate the MIKE 21 grid output to the monitoring point location X. i Make position X i Spatiotemporal matching with measured data was performed, and the following parameters were collected every 15 minutes: flow velocity v, flow rate Q, water level h, and pollutant concentration C. 实测 ;

[0027] The time series is constructed with a time step Δt = 15 min, and the time series data {Xt} = [h(X i ,t),v(X i ,t), C measured (X) i ,t),Q(t),C MIKE (X i ,t)(i=1,…,N,N is the number of monitoring points);

[0028] A three-layer stacked LSTM neural network structure is adopted, with 128, 64, and 32 hidden units in each LSTM layer, respectively. A fully connected layer is then added to output the residual prediction values ​​for N monitoring points. In this three-layer LSTM structure, the output of each layer serves as the input to the next. Let the hidden state of the first LSTM layer be h. t (1) The second layer is h t (2) The third layer is h t (3) The specific transmission is as follows:

[0029]

[0030] First LSTM layer: Input: X t Output: h t (1) Dimension 128, second-layer LSTM: Input: h t (1)Output: h t (2) Dimension 64, third-layer LSTM: Input: h t (2) Output: h t (3) Dimension 32, C t It is the real-time concentration, C t-1 The concentration of the previous LSTM cell is calculated for each LSTM cell at time step t as follows:

[0031] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i )

[0032] Forgotten Gate:f t =σ(W f ·[h t-1 ,x t ]+b f )

[0033] Output gate: o t =σ(W o ·[h t-1 ,x t ]+b o )

[0034] Hidden layer output: h t =o t ☉tanh(C t )

[0035] Among them, X t It is the input vector at the current time step, with a dimension of 5, i.e., [h, v, C]. 实测 Q, C MIKE ]; h t-1 It is the hidden state of the previous time step; C t-1 It represents the memory cell state of the previous time step; σ is the sigmoid activation function; ⊙ represents element-wise multiplication; W and b are the weight matrix and bias vector, respectively;

[0036] The output h of the third LSTM layer at time step t t (3) The residual prediction values ​​of N monitoring points are obtained through a fully connected layer;

[0037]

[0038] Among them, W o This is the weight matrix of the output layer, with dimensions 32×N, b oIt is a bias vector with dimension N, therefore... It is an N-dimensional vector, where each element corresponds to the residual prediction of a monitoring point.

[0039] Preferably, the location of the monitoring point is X. i The predicted residual is Output the predicted residual field at the grid point, for any grid point X. i Its residual prediction value is:

[0040]

[0041] Wherein, weight λ i The following system of equations was solved to obtain the following:

[0042]

[0043] γ is the semivariogram, and μ is the Lagrange multiplier.

[0044] Preferably, the pollutant migration equation generated by MIKE is modified using the pollutant whole-basin residual field obtained by the LSTM neural network. An LSTM residual correction term is introduced, and the modified MIKE prediction model is defined as an LSTM-MIKE 21 coupled model.

[0045]

[0046] Right now:

[0047]

[0048] Where: α is the adaptive weight, which is calculated as follows:

[0049]

[0050] β is the basic weight, ε is the standard deviation of the residual in the current period, and γ is the hydrological abrupt change factor. When the rate of change of flow velocity exceeds the threshold, γ = 0.5, otherwise it is 1.0.

[0051] when When this happens, the correction weight α at that point is automatically reset to zero to avoid overcorrection.

[0052] Preferably, the time step and spatial grid of the exported simulation results are used to adjust the diffusion coefficient and degradation rate parameters using historical measured concentration data to ensure that the simulated value Cpred is consistent with the actual trend of the measured value C. After calibration and debugging, the relative error between the simulated value and the measured value is controlled within 20%.

[0053] Preferably, in step 4, the input includes: real-time hydrological data and pollutant concentration C at the monitoring point. 实测 Monitoring time t;

[0054] Perform LSTM-MIKE 21 coupled model computation:

[0055] First, hydrodynamic field calculations are performed using MIKE 21. Based on real-time collected flow rate data Q(t), wind speed W(t), and boundary water level data, the hydrodynamic fields h and v are solved. Simultaneously, real-time water quality data from monitoring points are input to drive an LSTM neural network to predict the residual δC of the physical model. LSTM The pollutant migration equation generated by MIKE is modified using the pollutant whole-basin residual field obtained by LSTM. An LSTM residual correction term is introduced, and the modified MIKE prediction model is used to obtain the LSTM-MIKE 21 coupled model and calculate C. pred (x, t), ultimately outputting the pollutant concentration field C for the entire watershed. pred (x, t).

[0056] The present invention has at least the following beneficial effects:

[0057] This invention significantly improves the accuracy and reliability of pollutant migration simulation by bidirectionally coupling a physical mechanism model (MIKE 21) with a data-driven model (LSTM), achieving synergistic optimization of the hydrodynamic field and the pollutant concentration field. Specifically, the LSTM residual correction mechanism effectively compensates for complex processes not covered by the pure physical model, reducing the pollutant concentration prediction error from >20% in traditional methods to <8%, while achieving a consistency of over 99.5% for mass conservation and non-negative concentrations.

[0058] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0059] Figure 1 System construction flowchart;

[0060] Figure 2 Target river location in a specific embodiment;

[0061] Figure 3 River COD migration distribution in the examples;

[0062] Figure 4 Comparative analysis of COD migration distribution in rivers. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0064] Example

[0065] Reference Figure 1-3 This embodiment provides a method and system for the synergistic prediction of river hydrodynamics and pollutant migration based on the LSTM-MIKE 21 coupled model. The specific process is as follows: Figure 1 As shown.

[0066] This embodiment uses the Zhangcun River as the research object, selecting the western starting point of the simulated section at the Qingyin Expressway, extending eastward to the Binhai Avenue, with a total river length of approximately 4000 meters and a study width varying from 150 to 480 meters, covering a total area of ​​approximately 821,300 square meters (excluding the north bank). The geographical coordinates are between 120°24′33″ and 120°43′ east longitude and 36°03′10″ and 36°20′23″ north latitude. The target river is a typical northern seasonal river, with precipitation concentrated from June to September, and a rapid rise and fall in flow during the flood season. During the dry season, the river lacks a certain ecological base flow. The river flows from east to west; the lowest elevation is 8.21 meters, and the highest elevation is 123.3 meters; the elevation difference between the upstream and downstream sections is approximately 110 meters; the riverbed gradient is 0.67%.

[0067] Step 1: Building the MIKE 21 model:

[0068] 1. Mesh Analysis and Terrain Import

[0069] The river channel was gridded using a triangular network (TIN) derived from MIKE 21 software, with a total number of grids ≥ 20,000. The shoreline boundary grids were refined to a resolution ≤ 5m. Based on satellite elevation data (50m × 50m resolution), a weighted inverse distance interpolation method was used. Underwater topographic data is mapped to grid nodes to generate a 3D topographic model of the river channel; where Z i Let d be the elevation of the node to be interpolated. k Given the distance to nearby elevation points, p = 2 (secondary weight), a three-dimensional terrain model is generated.

[0070] 2. Construction of the hydrodynamic field:

[0071] A two-dimensional hydrodynamic model is established based on the Navier-Stokes equations, coupled with the following environmental dynamic factors:

[0072]

[0073] h: water depth (m), v = (u, v): flow velocity vector (m / s), η: water level (m), g: gravitational acceleration (9.81 m / s²) 2 ), Dry and wet treatment: water depth ≤ 0.01m is defined as dry cell mesh.

[0074] The river section was divided into 21,374 triangular grids, with the original data being satellite data at a resolution of 50m×50m.

[0075] Step 2: Construction of pollutant migration equations:

[0076] Based on the pollutant kinetic model established in step 1, daily monitoring data from the Donghan Station of Zhangcun River in Qingdao were imported into the MIKE 21 software. The inflow boundary conditions were selected as inflow from the upstream tributary and water replenishment from the water purification plant. According to rainfall and evaporation data from the Qingdao meteorological station where the river section is located, the average annual surface evaporation is 1472.6 mm, and the average annual rainfall is 905.48 mm. After fully balancing the model's computational stability and efficiency, and through multiple adjustments and trials, the model's computational time step was determined to be 50 seconds. The following COD variable equations were defined based on the mathematical relationships between various water quality indicators:

[0077]

[0078] C: Pollutant concentration (mg / L)

[0079] Δ: Grid scale

[0080] D: Diffusion tensor (m 2 / s), including turbulent diffusion and molecular diffusion

[0081] S 反应 =-K d hC: Primary degradation term (K) d： Degradation rate constant, d -1 )

[0082] S 外源 Point source / area source emissions (g / s)

[0083] Diffusion coefficient D: Using the Smagorinsky subgrid model, Cs = 0.2.

[0084] D t =(C s Δ) 2 |S|,

[0085] Step 3: Construct the LSTM-MIKE 21 coupled model

[0086] In the simulated section, 13 sampling points were selected: C1, C2 South, C2 North, C3 South, C3 North, C4 South, C4 North, C5 South, C5 North, C6 South, C6 North, C7 South, and C7 North (e.g., ...). Figure 2 As shown in the figure, the measured values ​​of flow velocity, flow rate, water level, and COD are C. 实测 Determine the residual δC = C between the measured and simulated values. 实测 -C MIKE We used LSTM to predict the δC value of the simulated river section and built a model.

[0087] 3.1 LSTM Prediction Network Design

[0088] Data collection:

[0089] Interpolate the MIKE 21 grid output to the monitoring point location (X). i The data was spatiotemporally matched with measured data, and the following parameters were collected every 15 minutes: flow velocity v (m / s), flow rate Q (m³ / s). 3 / s), water level h (m), COD concentration C 实测 (mg / L).

[0090] Input data: Construct a time series with a time step of Δt = 15 min, and the time series data {Xt} = [h(X i ,t),v(X i ,t),C 实测 (X i ,t),Q(t),C MIKE (X i ,t)(i=1,…,N monitoring points).

[0091] Data processing:

[0092] This embodiment employs a three-layer stacked LSTM structure, with 128, 64, and 32 hidden units per layer, respectively. A fully connected layer is then added to output the residual prediction values ​​for N monitoring points. In this three-layer LSTM structure, the output of each layer serves as the input to the next. Let the hidden state of the first LSTM layer be h. t (1) The second layer is h t (2) The third layer is h t (3) The specific transmission is as follows:

[0093]

[0094] First LSTM layer: Input: X t Output: h t (1) (Dimension 128);

[0095] Second LSTM layer: Input: h t (1) Output: h t (2) (Dimension 64);

[0096] Third-layer LSTM: Input: h t (2) Output: h t (3) (Dimension 32);

[0097] The calculation details for each LSTM unit at time step t are as follows:

[0098] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i )

[0099] Forgotten Gate:f t =σ(W f ·[h t-1 ,x t ]+b f )

[0100] Output gate: o t =σ(W o ·[h t-1 ,x t ]+b o )

[0101] Hidden layer output: h t =o t ☉tanh(C t )

[0102] X t It is the input vector at the current time step (dimension 5, i.e., [h, v, C)). 实测 Q, C MIKE ])h t-1 It is the hidden state of the previous time step; C t-1 σ is the memory cell state of the previous time step; σ is the sigmoid activation function; ⊙ represents element-wise multiplication; W and b are the weight matrix and bias vector, respectively. The output ht(3) of the third LSTM at time step t is passed through a fully connected layer to obtain the residual prediction values ​​of N monitoring points.

[0103]

[0104] Among them, W o This is the weight matrix of the output layer, with dimensions 32×N, b o It is a bias vector with dimension N. Therefore, It is an N-dimensional vector, where each element corresponds to the residual prediction of a monitoring point.

[0105] Since the residual field generation LSTM outputs residual predictions for N monitoring points, the Kriging interpolation method is used to extend it to the entire simulation region (i.e., grid points). Let the monitoring point location be X. i The predicted residuals for i = 1, ..., N are

[0106] Output the predicted residual field at the grid point, for any grid point X. i Its residual prediction value is:

[0107]

[0108] Wherein, weight λ i The following system of equations was solved to obtain the following:

[0109]

[0110] Where γ is the semivariogram and μ is the Lagrange multiplier.

[0111] This embodiment uses the pollutant whole-basin residual field obtained by LSTM to correct the pollutant migration equation generated by MIKE, and introduces an LSTM residual correction term. The corrected MIKE prediction model is defined as an LSTM-MIKE 21 coupled model.

[0112]

[0113] Right now:

[0114]

[0115] Where: α is the adaptive weight, which is calculated as follows:

[0116]

[0117] β is the basic weight (set to 0.8), ε is the standard deviation of the residual in the current period, and γ is the hydrological abrupt change factor. When the rate of change of flow velocity exceeds the threshold, γ = 0.5, otherwise it is 1.0.

[0118] At the same time, a threshold condition is set: when When this happens, the correction weight α at that point is automatically reset to zero to avoid overcorrection.

[0119] The time step and spatial grid of the exported simulation results are used to adjust parameters such as diffusion coefficient and degradation rate using historical measured concentration data to ensure that the simulated value C is accurate. pred Compared with the measured value C 实际 The trend is consistent. After continuous calibration and debugging, the relative error between the simulated value and the measured value is controlled within 20%, indicating that the model has a good simulation effect on pollutant concentration.

[0120] The final output is as follows Figure 3 As shown, after comparing with the actual values ​​at the measured points, it was found that the error of the predicted values ​​was less than 20%, indicating a good fit.

[0121] Comparative Example

[0122] This comparative example only uses a MIKE 21 to simulate the target river in Example 1, and the results are as follows: Figure 4 As shown, the specific steps include:

[0123] MIKE 21 model building:

[0124] 1. Mesh Analysis and Terrain Import

[0125] The river channel was gridded using a triangular network (TIN) derived from MIKE 21 software, with a total number of grids ≥ 20,000. The shoreline boundary grids were refined to a resolution ≤ 5m. Based on satellite elevation data (50m × 50m resolution), a weighted inverse distance interpolation method was used. Underwater topographic data is mapped to grid nodes to generate a 3D topographic model of the river channel; where Z i Let d be the elevation of the node to be interpolated. k Given the distance to nearby elevation points, p = 2 (secondary weight), a three-dimensional terrain model is generated.

[0126] 2. Construction of the hydrodynamic field:

[0127] A two-dimensional hydrodynamic model is established based on the Navier-Stokes equations, coupled with the following environmental dynamic factors:

[0128]

[0129] h: water depth (m), v = (u, v): flow velocity vector (m / s)

[0130] η: water level (m), g: acceleration due to gravity (9.81 m / s²) 2 )

[0131] Dry / wet treatment: Water depth ≤ 0.01m is defined as dry cell mesh.

[0132] The river section was divided into 21,374 triangular grids, with the original data being satellite data at a resolution of 50m×50m.

[0133] Construction of pollutant migration equations:

[0134] Based on the pollutant kinetic model established in step 1, daily monitoring data from the Donghan Station of Zhangcun River in Qingdao were imported into the MIKE 21 software. The inflow boundary conditions were selected as inflow from the upstream tributary and water replenishment from the water purification plant. According to rainfall and evaporation data from the Qingdao meteorological station where the river section is located, the average annual surface evaporation is 1472.6 mm, and the average annual rainfall is 905.48 mm. After fully balancing the model's computational stability and efficiency, and through multiple adjustments and trials, the model's computational time step was determined to be 50 seconds. The following COD variable equations were defined based on the mathematical relationships between various water quality indicators:

[0135]

[0136] C: Pollutant concentration (mg / L)

[0137] Δ: Grid scale

[0138] D: Diffusion tensor (m 2 / s), including turbulent diffusion and molecular diffusion

[0139] S 反应 =-K d hC: Primary degradation term (K) d： Degradation rate constant, d -1 )

[0140] S 外源 Point source / area source emissions (g / s)

[0141] The output results are as follows Figure 4 As shown, the simulation results deviate significantly from the measured values, with a large error range, while the simulation results in the examples all have errors of less than 20%.

[0142] As described above, this invention significantly improves the accuracy and reliability of pollutant migration simulation by bidirectionally coupling the physical mechanism model (MIKE 21) and the data-driven model (LSTM), achieving synergistic optimization of the hydrodynamic field and the pollutant concentration field. Specifically, the LSTM residual correction mechanism effectively compensates for complex processes not covered by the pure physical model, reducing the pollutant concentration prediction error from >20% in traditional methods to <8%, while achieving a consistency of over 99.5% for mass conservation and non-negative concentrations.

[0143] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for coupled prediction of river hydrodynamics and pollutant transport based on a coupling model, characterized in that, Comprising the following steps: Step 1, constructing the MIKE21 hydrodynamic field of the river channel; comprising: Step 1.1, mesh analysis and terrain import Collect the flow velocity, flow rate, water depth, elevation data of the river channel, and use the irregular triangle net of the MIKE21 software to divide the grid of the flow velocity, flow rate, water depth, elevation data of the river channel; based on satellite elevation data, the underwater topographic data obtained through satellite data and hydrological station is mapped to the grid node through weighted inverse distance interpolation method, and a three-dimensional terrain model of the river channel is generated; The weighted inverse distance interpolation method is: wherein, Zi is the elevation of the node to be interpolated, d k is the distance between adjacent elevation points, p = 2; N is the number of monitoring points; Step 1.2, constructing the hydrodynamic field Based on the Navier-Stokes equation, a two-dimensional hydrodynamic model is established, which couples the following environmental dynamic factors: where h: water depth (m), v = (u, v): velocity vector (m / s), u is the longitudinal velocity, v is the transverse velocity, v t : real-time flow rate; η: water level (m), g: gravitational acceleration, g = 9.81 m / s 2 , dry-wet treatment: water depth ≤ 0.01 m is defined as a dry unit grid; Step 2, according to the MIKE21 hydrodynamic field, constructing the pollutant migration model; comprising: Based on step 1, a pollutant dynamics model is established, and daily monitoring data from monitoring stations is imported into MIKE 21 software, and the following pollutant migration model is established according to the mathematical relationship between each water quality index: Where C: pollutant concentration (mg / L), Δ: grid scale, D: diffusion tensor, unit m 2 / s, containing turbulent diffusion and molecular diffusion, S 反应 = -K d hC: first-order degradation term, K d : degradation rate constant, unit d -1 , S 外源 : point source / surface source emission, unit g / s; diffusion coefficient D: Smagorinsky sub-grid model is adopted, Cs=0.2; According to the pollutant migration model of the river basin, a simulated value C of the pollutant concentration can be obtained MIKE ; Step 3, coupling the physical mechanism model MIKE 21 and LSTM bidirectionally to construct the LSTM-MIKE 21 coupling model; Step 4, real-time prediction of pollutant concentration field in the whole river basin by LSTM-MIKE 21 coupling model; Wherein, in step 3, the LSTM neural network is used to predict the river simulation section delta C value, and a residual model is established; wherein delta C = C 实测 -C MIKE , C 实测 is the measured value of the pollutant concentration; The construction steps of the residual model include: Interpolate MIKE 21 grid output to monitoring point locations X i Match locations X i with spatio-temporal measured data, collect the following parameters every 15 minutes: flow velocity v, flow rate Q, water level h, pollutant concentration C 实测 ; The time sequence is constructed with a time step Δt = 15 min, and the time sequence data {Xt} = [h(X i , t), v(X i , t), C measured(X i , t), Q(t), C MIKE (X i , t) (i = 1, …, N, N is the number of monitoring points); Adopt three layers of stacked LSTM neural network structure, the number of hidden layer units of each layer of LSTM is 128, 64, 32 respectively, and finally connect a full connection layer to output the residual prediction value of N monitoring points, three layers of LSTM, the output of each layer is taken as the input of the next layer, let the hidden state of the first layer of LSTM be h t (1) , the second layer be h t (2) , and the third layer be h t (3) ; The output h t (3) The residual prediction value of the N monitoring points is obtained through a fully connected layer. where W o is the weight matrix of the output layer with dimension 32 x N, b o is the bias vector with dimension N, and thus, is an N-dimensional vector, each element of which corresponds to the residual prediction of a monitoring point; Let the monitoring point position be X i The prediction residual of X is Output the prediction grid point residual field, for any grid point X i The residual prediction value of X is: where the weight λ i By solving the following system of equations: γ is the semi-variogram function, and μ is the Lagrange multiplier; The pollutant whole basin residual field obtained by using the LSTM neural network is used to correct the pollutant migration equation generated by MIKE, and the LSTM residual correction term is introduced, and the corrected MIKE prediction model is defined as the LSTM-MIKE 21 coupling model: That is: Where: α is the adaptive weight, and its calculation method is: β is the basic weight, ε is the standard deviation of the current period residual, and γ is the hydrological mutation factor, which is 0.5 when the flow rate change rate exceeds the threshold, otherwise it is 1.0; When the correction weight a of the point position is automatically zeroed to avoid over-correction.

2. The coupled model based river hydrodynamics and pollutant transport co- prediction method of claim 1, wherein, The time step and spatial grid of the exported simulation results are used to adjust the diffusion coefficient and degradation rate parameters with historical measured concentration data, so as to ensure that the simulation value Cpred and the measured value Cactual trend is consistent, and after calibration and debugging, the relative error between the simulation value and the measured value is controlled within 20%.

3. The coupled model based river hydrodynamics and pollutant transport co- prediction method of claim 2, wherein, In step 4, input: real-time hydrological data, pollutant concentration C at monitoring point, monitoring time t 实测 , Carrying out LSTM-MIKE 21 coupling model operation: Firstly, the hydrodynamic field is calculated by MIKE 21, based on the real-time collected flow data Q(t), wind speed W(t) and boundary water level data, to solve the hydrodynamic field h, v; the real-time water quality data of the monitoring points are accessed synchronously to drive the LSTM neural network to predict the physical model residual δC LSTM ; the pollutant residual field of the whole basin obtained by the LSTM is used to modify the pollutant migration equation generated by MIKE, and the LSTM residual correction term is introduced; the modified MIKE prediction model is used to obtain the LSTM-MIKE 21 coupling model and calculate C pred (x, t), and finally the pollutant concentration field C pred (x, t) of the whole basin is output.

Images