Water quality simulation and prediction method

Abstract

The invention discloses a water quality simulation and prediction method, which comprises the following steps: collecting data of a target detection area, preprocessing the data, and constructing a training data set; the training data set is used to train the Transform deep learning model; predicting a key hydrological parameter sequence of a future time period by using the trained Transform deep learning model; and simulating the water quality of the target water area by using an EFDC model based on the key hydrological parameter sequence and the water quality measured data. The reliability of water quality simulation is improved, and more scientific and accurate decision support is provided for water environment management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water environment simulation, in particular to a water quality simulation and prediction method. BACKGROUND

[0002] With the development of water environment management towards refinement and precision, higher requirements are put forward for water quality monitoring and prediction of small and medium scale river and lake water bodies such as urban rivers, tributaries and small lakes. As the key collection and diffusion area of regional pollution, the dynamic change of the water quality of such water bodies not only reflects the health status of the regional water environment, but also directly relates to the regional water ecological safety and the quality of the living environment. At present, for the water quality simulation and prediction of small and medium scale river and lake water bodies, the traditional method mainly relies on the water dynamic-water quality numerical model based on physical process, such as MIKE series, EFDC (Environmental Fluid Dynamics Code), Delft3D, etc. These models accurately depict the spatio-temporal evolution process of pollutants in water bodies by solving hydrodynamic equations and pollutant transport and transformation equations, and have the advantages of good physical consistency and process explainability, which can clearly reveal the migration and transformation mechanism of pollutants in water bodies, and provide important theoretical basis and technical support for water environment management and research. However, the reliability of the physical numerical model in practical application depends on the spatio-temporal integrity and accuracy of the input hydrological driving data, including water level, flow, flow rate, etc. In the practical application scene of small and medium scale river and lake area, due to unreasonable monitoring site layout, monitoring cost limitation and other factors, there are problems such as sparse hydrological monitoring sites, serious data missing, lagging update and even no monitoring station, which makes it difficult to build boundary conditions required by the model or the quality is low, which seriously affects the simulation accuracy and reliability of the model. Especially under the disturbance of rainstorm, sudden pollution or gate pump scheduling, the traditional hydrological data interpolation method shows obvious limitations in dealing with nonlinear and non-stationary hydrological processes of small rivers, and it is difficult to accurately capture the dramatic changes of hydrological processes, so that the water quality simulation result is distorted, and accurate and effective decision support for water environment management cannot be provided. SUMMARY

[0003] In order to solve the problem of poor applicability and weak physical consistency in the application scene of small and medium scale river under the condition of sparse monitoring sites and serious data missing, the present application provides a water quality simulation and prediction method, which improves the reliability of water quality simulation and provides more scientific and accurate decision support for water environment management.

[0004] In order to achieve the above purpose, the present application is realized by the following technical scheme:

[0005] The present application is a water quality simulation and prediction method, comprising:

[0006] Collecting data of the target detection area and preprocessing the data to construct a training data set;

[0007] training the Transformer deep learning model by using the training data set;

[0008] predicting the sequence of the key hydrological parameters in the future period by using the trained Transformer deep learning model;

[0009] simulating the water quality of the target water area by using the EFDC model based on the sequence of the key hydrological parameters and the water quality measured data.

[0010] The further improvement of the present application is that the preprocessing comprises data cleaning, outlier elimination, spatio-temporal alignment and normalization operation on the collected data of the target detection area, wherein the data of the target detection area comprises topographic data, hydrological data, meteorological data, water quality measured data, point source pollution data, non-point source pollution data and tributary inflow data.

[0011] The further improvement of the present application is that the Transformer deep learning model comprises an encoder and a decoder, wherein the encoder is composed of multiple self-attention layers and a feedforward neural network, and the decoder generates a sequence through a masked self-attention mechanism, wherein the calculation expression of the multi-head self-attention layer in each self-attention layer is:

[0012] ;

[0013] wherein: 、 are the first , the second attention head, 、 、 is the projection matrix of the first attention head, is the output projection matrix, is a vector, is the connection attention head, is a function.

[0014] The further improvement of the present application is that the water quality of the target water area is simulated by using the EFDC model based on the sequence of the key hydrological parameters and the water quality measured data, comprising:

[0015] performing data coupling processing on the sequence of the key hydrological parameters, and inputting the obtained result as a dynamic boundary condition into the EFDC model;

[0016] performing parameter calibration of the EFDC model in combination with the water quality measured data, solving the control equation of the target water area, and simulating the spatio-temporal distribution and evolution process of the pollutants in the target water area.

[0017] A further improvement of the present invention is that the data coupling processing includes:

[0018] The key hydrological parameter sequences are transformed by spatial interpolation and time step to match the grid boundary and calculation step of the EFDC model;

[0019] The key hydrological parameter sequences were converted into NetCDF format that matches the EFDC model.

[0020] A further improvement of the present invention is that the governing equations for the target water area include the continuity equation, the momentum equation, and the pollutant transport equation;

[0021] Continuity equation:

[0022] ;

[0023] in, Indicates the total water depth. for Flow velocity in the direction, For external sources entering and exiting the system, For a specific moment;

[0024] Momentum equation:

[0025] ;

[0026] in, Indicates the total water depth. For absolute values, It is the acceleration due to gravity. This refers to the water level shift that deviates from the equilibrium water level. This is the bottom drag coefficient. The horizontal eddy viscosity coefficient, for Wind stress in the direction;

[0027] Pollutant transport equations:

[0028] ;

[0029] in, Concentration is a state variable in water quality. for Turbulent diffusion coefficient in the direction, Let be the degradation rate constant of the water quality state variable. These are the source and sink terms for water quality state variables.

[0030] A further improvement of this invention lies in: constructing the Transformer deep learning model and the EFDC model into a closed-loop optimization system, establishing a bidirectional iterative coupling and optimization mechanism, including:

[0031] The error feedback phase includes comparing the water quality simulation results output by the EFDC model with the measured water quality data of the target water area and calculating the simulation error of the water quality state variables.

[0032] The reverse correction phase includes propagating the simulation error back to the decoder part of the Transformer deep learning model and adjusting the decoder parameter weights through the gradient descent algorithm.

[0033] The iterative optimization phase includes repeated water quality simulation, error feedback, and reverse correction steps for the target water area until the Nash efficiency coefficient of the water quality simulation results reaches the preset standard or the number of iterations reaches the preset upper limit, and the optimal prediction result is output.

[0034] The beneficial effects of this invention are as follows: This invention utilizes the sequence modeling capabilities of the Transformer deep learning model to complete and predict key hydrological driving data, and couples this data as reliable boundary conditions to the mature, open-source EFDC model. Ultimately, water quality simulation and predictive analysis are achieved within the framework of the mechanistic model, improving the reliability of water quality simulation and providing more scientific and accurate decision support for water environment management. This invention enables the mature water quality mechanistic model, namely the EFDC model, which was originally limited by data completeness, to be applied in small- and medium-scale rivers and lakes with insufficient water quality monitoring sections and no continuous flow observation. The feasibility of the EFDC model has increased from less than 40% to 100%, while reducing the basic data collection cost required for constructing water quality modules in specific watersheds by more than 70%. This invention proposes a bidirectional iterative coupling optimization mechanism, achieving deep integration of the Transformer deep learning model and the EFDC model through closed-loop feedback correction, rather than a simple combination. This provides a solution for water quality simulation in small- and medium-scale rivers and lakes, and is more practically significant for management. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method in an embodiment of the present invention; Figure 2 This is the water level verification result in the embodiment of the present invention; Figure 3 This is the water quality verification result in the embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0037] like Figure 1As shown, the water quality simulation and prediction method in this embodiment includes:

[0038] Step 1, Data Collection and Preprocessing: Collect data from the target detection area, preprocess the data, and construct a training dataset;

[0039] Step 2, drive data generation: train the Transformer deep learning model using the training dataset, and use the trained Transformer deep learning model to predict key hydrological parameter sequences for future periods;

[0040] Step 3, Model Coupling: Couple the data of the interaction between the Transformer deep learning model and the EFDC model to achieve coupling between the Transformer deep learning model and the EFDC model;

[0041] Step 4, Water quality simulation: Simulate the water quality of the target water area using the EFDC model.

[0042] Step 5, Result Evaluation and Optimization: Construct the Transformer deep learning model and the EFDC model into a closed-loop optimization system, establish a two-way iterative coupling and optimization mechanism, and optimize the water quality prediction results.

[0043] The data collected in Step 1 for the target detection area includes at least topographic data, hydrological data, meteorological data, measured water quality data, point source pollution data, and tributary inflow information. The measured water quality data includes, but is not limited to, permanganate index, ammonia nitrogen, total phosphorus, and total nitrogen. Preprocessing involves data quality control and fusion operations, including data cleaning, outlier removal, spatiotemporal alignment, and normalization of meteorological and hydrological data to improve data integrity and consistency.

[0044] In step 2, based on the Transformer deep learning model, the historical hydrological and meteorological sequences obtained in step 1 are trained to predict key hydrological parameter sequences for future periods, which serve as the dynamic input to the EFDC model. Key hydrological parameters include at least the water level and flow rate upstream and downstream of the target detection area. The Transformer deep learning model is specifically a sequence modeling architecture based on a self-attention mechanism, capable of effectively capturing global dependencies in long-term sequences and suitable for prediction tasks involving non-stationary, multivariate hydrological and meteorological data. The Transformer deep learning model structure includes an encoder and a decoder. The encoder consists of multiple self-attention layers and a feedforward neural network, while the decoder generates sequences through a masked self-attention mechanism.

[0045] The core calculation formula for the multi-head self-attention layer in the Transformer deep learning model is as follows:

[0046] ;

[0047] Each attention head is calculated as follows:

[0048] ;

[0049] in: , The first , No. One point of attention, , , For the first The projection matrix of each attention head. To output the projection matrix, For vectors, To connect attention, It is a function.

[0050] After the self-attention layer outputs from the encoder of the Transformer deep learning model, residual connections and layer normalization are performed sequentially to enhance the model's training stability and expressive power. The feedforward neural network adopts a two-layer fully connected structure.

[0051] Step 3 involves using the key hydrological parameter sequences output by the Transformer deep learning model as time-series boundary conditions to drive the EFDC model. This replaces the boundary condition settings in traditional models that rely on historical averages or simple extrapolations, enabling the EFDC model to respond to more accurate and predictive hydrological dynamics. The coupling process mainly includes:

[0052] Spatiotemporal scale matching: The key hydrological parameter sequence output by the Transformer deep learning model is matched to the grid boundary and computational step size required by the hydrodynamic module of the EFDC model through spatial interpolation and time step conversion.

[0053] Data format standardization: The data exchange format is uniformly defined as NetCDF to enable data transfer between two heterogeneous models.

[0054] In step 4, the governing equations for applying the EFDC model to small- and medium-scale rivers and lakes mainly include the continuity equation, momentum equation, and pollutant transport equation.

[0055] Continuity equation:

[0056] ;

[0057] in, Indicates the total water depth. for Flow velocity in the direction, For external sources entering and exiting the system, For a moment.

[0058] Momentum equation:

[0059] ;

[0060] in, Indicates the total water depth. For absolute values, It is the acceleration due to gravity. This refers to the water level shift that deviates from the equilibrium water level. This is the bottom drag coefficient. The horizontal eddy viscosity coefficient, for Wind stress in the direction;

[0061] Pollutant transport equations:

[0062] ;

[0063] in, Concentration is a state variable in water quality. for Turbulent diffusion coefficient in the direction, Let be the degradation rate constant of the water quality state variable. These are the source and sink terms for water quality state variables.

[0064] Before the EFDC model can make water quality predictions, it is necessary to establish a robust and reasonable hydrodynamic module based on the collected topographic data, tributary inflow conditions and key hydrological parameter sequences transmitted from the Transformer deep learning model, so as to ensure the physical consistency of the simulation process.

[0065] The hydrodynamic module of the EFDC model relies on the continuity equation and momentum equation to solve for water level, flow velocity, etc. After the solution is completed, the data is transmitted to the water quality module of the EFDC model, and then the water quality is solved based on the pollutant transport equation to obtain the water quality simulation results.

[0066] The EFDC model is calibrated using measured water quality data to ensure that the simulation process is physically reasonable and numerically stable.

[0067] In step 5, a weighted error backpropagation algorithm is used to compare the water quality simulation results output by the EFDC model with the measured water quality data, generating an error signal. This error signal is then backpropagated to the decoder part of the Transformer deep learning model. A gradient descent algorithm is used to adjust the parameter weights of the Transformer deep learning model decoder. The adjusted Transformer deep learning model is then used to re-predict the key hydrological parameter sequence until preset conditions are met. Finally, the parameter configuration with the best water quality simulation effect during the iteration process is selected, and the final water quality simulation and prediction results are output. The specific process is as follows:

[0068] Step 1) Three statistical indicators are selected for error signals: root mean square error (RMSE), mean absolute error (MAE), and Nash efficiency coefficient (NSE). The water quality simulation results are compared with the measured water quality data to ensure that the simulation error is controlled within an acceptable range. If the water quality simulation results do not meet the preset accuracy requirements, they are fed back to the Transformer deep learning model training stage for local or global optimization.

[0069] Step 2) Calculate the gradient of the simulation error of each water quality state variable on the output of the Transformer deep learning model;

[0070] Step 3) Adjust the parameter weights of the Transformer deep learning model decoder using the gradient descent algorithm. The adjustment formula is as follows:

[0071] ;

[0072] in, For model parameters, The adjusted model parameters, For learning rate, For gradient, For the first Loss function for each water quality state variable, For the first Loss function for each water quality state variable The weights of water quality state variables include, but are not limited to, permanganate index, ammonia nitrogen, total phosphorus, and total nitrogen.

[0073] Step 4) Re-predict the key hydrological parameter sequence using the adjusted Transformer deep learning model: Repeat steps 1) to 3) until any of the following conditions are met: the average Nash efficiency coefficient (NSE) of the water quality simulation reaches 0.85 or higher, the number of iterations reaches the preset upper limit, and the rate of change of the mean absolute error (MAE) of the water quality simulation between two adjacent iterations is less than 1%. Then, select the set of parameter configurations that have the best water quality simulation effect during the iteration process and output the final water quality simulation and prediction results.

[0074] The following describes the specific implementation examples in detail, using specific scenarios as examples.

[0075] This study focuses on a key water quality monitoring section (hereinafter referred to as "the section") on a critical inflow channel in the northwest of a major lake. The aim is to address the problem of distorted boundary conditions and difficulty in water quality prediction caused by the sparse hydrological monitoring stations and severe data gaps during the flood season in this small-to-medium scale inflow channel. The coupling method described in this invention is used to select total phosphorus (TP) as the indicator for prediction.

[0076] Step 1, Data Acquisition and Preprocessing

[0077] Study Scope Definition: The study area is defined as the main channel upstream of a certain cross section, with a total length of 21.5 km, including the confluence of two important tributaries.

[0078] Data collection:

[0079] Topographic data: Riverbed elevation data from relevant research findings, 0.5m downstream and 1.9m upstream, were used to calculate the river slope for topographic interpolation in the EFDC model.

[0080] Hydrological data: Annual runoff data for one of the major tributaries was collected. Due to the lack of continuous daily flow observations, this embodiment uses this annual runoff as a long-term average reference and relies heavily on the dynamic flow series generated in subsequent steps.

[0081] Meteorological data: Daily rainfall, temperature, wind speed, and other data for the study area from 2018 to 2023 were collected to drive the Transformer deep learning model.

[0082] Water quality measurement data: Monthly routine monitoring data of a certain section in 2023 and daily data of automatic water quality stations were used, with a focus on total phosphorus.

[0083] Pollution source data: Information on the discharge outlets of 97 key enterprises and regional sewage treatment plants in the provincial pollution source monitoring system was integrated; combined with land use data and the results of a pre-flood on-site source tracing survey, industrial, agricultural, urban and rural domestic pollution sources were identified.

[0084] Data preprocessing: All time series data are cleaned, outliers are removed, and spatiotemporal resolution is standardized. Discrete monthly monitoring data are spatiotemporally aligned with continuous daily meteorological and automatic station data to form a unified standardized dataset.

[0085] Step 2, Driving Data Generation (Application of Transformer Deep Learning Model)

[0086] Model Objective: To construct a Transformer prediction model for the key hydrological parameter sequences required for the upper boundary of the hydrodynamic module. Traditional methods can only provide annual average flow rates and cannot reflect the drastic flow changes caused by torrential rains during the flood season.

[0087] Model Construction: A Transformer deep learning model based on a self-attention mechanism was built. Meteorological and hydrological data from 2018 to 2021 in the standardized dataset obtained in step 1 were used for model training, and validation and optimization were performed using a validation set (meteorological and hydrological data from 2022). Finally, meteorological data from the 2023 flood season (May to September) were used as input to predict the upstream flow and water level sequence for the corresponding period.

[0088] Model structure: A 6-layer encoder-decoder structure is adopted, with 8 multi-head attention heads and 512 hidden layer dimensions.

[0089] Model accuracy: By adjusting the model hyperparameters, the Nash efficiency coefficient (NSE) of the model on the validation set is made to reach above 0.85.

[0090] Dynamic boundary generation: The trained Transformer deep learning model was applied to the 2023 flood season (May-September) to generate dynamic flow sequences with high spatiotemporal resolution. This sequence characterizes the real hydrological processes under the flood season's torrential rainfall events and outperforms the constant flow assumption.

[0091] Steps 3 and 4: Model Coupling and Water Quality Simulation (Application of EFDC Model)

[0092] EFDC model construction:

[0093] Grid generation: The main river channel and important tributaries are generalized in one dimension and divided into 304 computational grids.

[0094] Boundary condition settings:

[0095] Upper boundary: The dynamic flow sequence generated by the Transformer deep learning model in step 2 is used as the boundary of the main tributary inflow.

[0096] Lower boundary: The measured water level sequence of the only lake in the river channel is used as the open boundary.

[0097] Initial conditions: Given an initial water depth of 1.0m.

[0098] Pollution source generalization: Based on the investigation and analysis results, the complex pollution sources are generalized into 6 main input points, including wastewater treatment plant effluent in the northern part of the region, upstream agricultural non-point sources, industrial point sources, and urban and rural domestic pollution, and are assigned corresponding TP concentrations and flow rates.

[0099] Parameter calibration and validation of the EFDC model:

[0100] Data from the dry season of 2023 were used to calibrate the parameters of the EFDC model. The channel roughness was calibrated to 0.025 m, and the TP degradation coefficient for the water quality module was set to 0.04 d. -1 .

[0101] Hydrodynamic condition verification:

[0102] The EFDC model was validated using data from the 2023 normal water period:

[0103] The hydrodynamic module verification showed NSE=0.87, RMSE=0.059m, and MRE=1.42%, indicating that the water level simulation accuracy meets the requirements and can be used for subsequent water quality prediction. The water level simulation results are as follows: Figure 2 As shown.

[0104] Step 5: Iterative optimization of water quality prediction results

[0105] The initial EFDC simulation yielded a TP concentration NSE of 0.09, an RMSE of 0.018 mg / L, and an MRE of 9.47%, indicating that the EFDC model has room for optimization.

[0106] The TP simulation error is backpropagated to the Transformer deep learning model, and the gradient descent algorithm is used to adjust the parameters in the decoder that are directly related to the generation of the traffic sequence. These parameters mainly include the weight matrix of the feedforward neural network of the mask self-attention layer and the output projection matrix of the last layer of the decoder.

[0107] After five iterations of optimization, the simulation accuracy of each iteration is shown in Table 1: Table 1: Simulation accuracy after 5 iterations

[0108]

[0109] After the aforementioned bidirectional iterative coupling optimization, the simulation accuracy of TP concentration was significantly improved. Specifically, NSE increased from 0.09 to 0.61, RMSE decreased from 0.018 mg / L to 0.011 mg / L, and MRE decreased from 9.47% to 4.80%. Although some high-value fluctuations still exist, the simulated mean is highly close to the measured data, indicating that the model has good reliability and can be used for subsequent continuous water quality prediction and analysis in this region. Water quality validation results are as follows... Figure 3 As shown.

[0110] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0111] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A water quality simulation and prediction method, characterized in that, include: Collect data from the target detection area, preprocess the data, and construct a training dataset; The Transformer deep learning model is trained using the training dataset; Predict key hydrological parameter sequences for future periods using a trained Transformer deep learning model; Based on key hydrological parameter sequences and measured water quality data, the EFDC model was used to simulate the water quality of the target water area.

2. The water quality simulation and prediction method according to claim 1, characterized in that, Preprocessing includes: data cleaning, outlier removal, spatiotemporal alignment and normalization of the collected data from the target detection area. The data from the target detection area includes topographic data, hydrological data, meteorological data, measured water quality data, point source pollution data, non-point source pollution data and tributary inflow data.

3. The water quality simulation and prediction method according to claim 1, characterized in that, The Transformer deep learning model includes an encoder and a decoder. The encoder consists of multiple self-attention layers and a feedforward neural network. The decoder generates sequences through a masked self-attention mechanism. The computational expression for the multi-head self-attention layer in each self-attention layer is as follows: ； ； in: , The first , No. One point of attention, For the first The projection matrix of each attention head. To output the projection matrix, For vectors, To connect attention, It is a function.

4. The water quality simulation and prediction method according to claim 1, characterized in that, Based on key hydrological parameter sequences and measured water quality data, the EFDC model is used to simulate the water quality of the target water area, including: The key hydrological parameter sequences were coupled data, and the results were used as dynamic boundary conditions input into the EFDC model. The parameters of the EFDC model were calibrated by combining measured water quality data, the governing equations of the target water area were solved, and the spatiotemporal distribution and evolution of pollutants in the target water area were simulated.

5. The water quality simulation and prediction method according to claim 4, characterized in that, Data coupling processing includes: The key hydrological parameter sequences are transformed by spatial interpolation and time step to match the grid boundary and calculation step of the EFDC model; The key hydrological parameter sequences were converted into NetCDF format that matches the EFDC model.

6. The water quality simulation and prediction method according to claim 4, characterized in that, The governing equations for the target water area include the continuity equation, the momentum equation, and the pollutant transport equation. Continuity equation: ； in, Indicates the total water depth. for Flow velocity in the direction, For external sources entering and exiting the system, For a specific moment; Momentum equation: ； in, Indicates the total water depth. For absolute values, It is the acceleration due to gravity. This refers to the water level shift that deviates from the equilibrium water level. This is the bottom drag coefficient. The horizontal eddy viscosity coefficient, for Wind stress in the direction; Pollutant transport equations: ； in, Concentration is a state variable in water quality. for Turbulent diffusion coefficient in the direction, Let be the degradation rate constant of the water quality state variable. These are the source and sink terms for water quality state variables.

7. The water quality simulation and prediction method according to claim 1, characterized in that, The Transformer deep learning model and the EFDC model are constructed into a closed-loop optimization system, establishing a bidirectional iterative coupling and optimization mechanism, including: The error feedback phase includes comparing the water quality simulation results output by the EFDC model with the measured water quality data of the target water area and calculating the simulation error of the water quality state variables. The reverse correction phase includes propagating the simulation error back to the decoder part of the Transformer deep learning model and adjusting the decoder parameter weights through the gradient descent algorithm. The iterative optimization phase includes repeated water quality simulation, error feedback, and reverse correction steps for the target water area until the Nash efficiency coefficient of the water quality simulation results reaches the preset standard or the number of iterations reaches the preset upper limit, and the optimal prediction result is output.

Images