Water source chlorophyll concentration prediction model design method based on machine learning

Abstract

The invention discloses a water source chlorophyll a concentration prediction model design method based on machine learning. The method comprises the following steps: acquiring chlorophyll a concentration data in a to-be-predicted region for a continuous period of time; carrying out data preprocessing on the chlorophyll a concentration data, and filtering high-frequency noise by adopting wavelet transform preprocessing; constructing a concentration prediction model, carrying out data preprocessing on chlorophyll a concentration data, and filtering high-frequency noise by adopting wavelet transform preprocessing; constructing different concentration prediction models, and inputting the processed chlorophyll a concentration data and physicochemical parameters into the prediction models to obtain a chlorophyll a concentration data prediction result; and comparing prediction results of different prediction models, and determining the prediction model. According to the prediction model design method, the WT-GRU model is adopted to preprocess the data through wavelet transform, the wavelet transform effectively extracts key time scale characteristics through signal decomposition, and the accuracy of chlorophyll a concentration prediction is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of chlorophyll concentration prediction, and particularly relates to a plain river network water source chlorophyll a concentration prediction model design method based on machine learning. BACKGROUND

[0002] With the rapid economic growth and steady population increase, the water bodies such as rivers and lakes are increasingly affected by human activities. The excessive discharge of nitrogen, phosphorus and other nutrients has become the main factor for the eutrophication of many water bodies, which not only leads to frequent occurrence of water bloom, but also seriously affects water quality, destroys water ecological balance and further increases human health risks. Chlorophyll a in water bodies reflects the growth status of phytoplankton and is an important indicator for evaluating water nutrient status, which is closely related to water ecological environment. The suddenness of algal bloom brings great challenges to water management and governance, and therefore, accurately predicting the dynamic changes of chlorophyll a concentration in water bodies and predicting the trend of water quality change in advance are effective means for quickly responding to potential water bloom risks, and the prediction of water bloom in drinking water source closely related to human activities is of great significance.

[0003] Many studies on chlorophyll-a prediction have been carried out by domestic and foreign scholars. Mechanistic methods, such as water quality analysis simulation program (WASP), soil and water assessment tool (SWAT), water environment and water ecological model (AQUATOX), etc., predict the change of chlorophyll-a by constructing a mathematical model based on physical and chemical processes, but are limited by data acquisition and model parameter determination in practical application, resulting in poor prediction applicability in different water bodies. Traditional non-mechanism methods, such as regression analysis, time series analysis, etc., have less demand for long time series data and are not sensitive to noise; remote sensing analysis is affected by factors such as climate and terrain, and the prediction accuracy is not stable enough. In recent years, artificial intelligence, machine learning and big data technologies have iterated and developed, and data mining and prediction analysis methods based on deep learning neural networks provide a feasible way for effective protection of water resources. Liang et al. proved that using long short-term memory (LSTM) network can well simulate the prediction ability of environmental fluid dynamics water quality model by overcoming the limitations of traditional mechanism model such as data limitation, large amount of calculation and future boundary conditions; Liao et al. used machine learning model to predict the chlorophyll-a concentration of Miyun Reservoir, and the results showed that the random forest (RF) model had higher prediction accuracy, more stable results, less overfitting and stronger prediction ability when there was data missing or abnormal; Huang et al. used machine learning method to predict chlorophyll-a in Poyang Lake, and the results showed that the support vector machine (SVR) model had higher precision and accuracy than the other two models; Ren Shushun et al. used three time series models to predict algal blooms in Jiulong River reservoir area and carried out comparative analysis, and the results showed that the LSTM model was better than the other two models in short-term prediction of total chlorophyll-a concentration. However, the performance of a single machine learning model depends largely on the quality and integrity of the input data, such as data bias, incompleteness or errors, which may affect the prediction accuracy of the model. Chlorophyll-a online monitoring data is non-steady-state data, which is affected by observation error, system error and other factors, and the test results contain random noise, so the monitoring is not accurate. SUMMARY

[0004] In view of the technical problems of the prior art, the present application provides a water source chlorophyll-a concentration prediction model design method based on machine learning, which improves the accuracy of chlorophyll-a concentration prediction.

[0005] To solve the above technical problems, the technical solution provided by the present application is:

[0006] A water source chlorophyll-a concentration prediction model design method based on machine learning, comprising:

[0007] Obtaining chlorophyll-a concentration data in a to-be-predicted area for a continuous period of time;

[0008] The chlorophyll a concentration data is preprocessed, and wavelet transform is used for preprocessing to filter high-frequency noise.

[0009] Different concentration prediction models are constructed, the processed chlorophyll a concentration data and physicochemical parameters are input into the prediction models, and the chlorophyll a concentration data prediction results are obtained.

[0010] The prediction results of different prediction models are compared to determine the prediction model.

[0011] As a further improvement of the above technical solution:

[0012] In the above technical solution, preferably, constructing different concentration prediction models at least includes constructing a long short-term memory network model, constructing a gated recurrent unit model, constructing a convolutional neural network model, and constructing a multilayer perceptron model.

[0013] In the above technical solution, preferably, the long short-term memory network model includes an input gate, a forget gate and an output gate, the input gate determines the new information added to the unit, the forget gate determines the old information to be retained or forgotten in the unit state, and the output gate determines the output information of the current unit state.

[0014] The long short-term memory network model is:

[0015] f t = σ(W f [h t-1 ,x t ]+b f )

[0016] i t = σ(W i [h t-1 ,x t ]+b i )

[0017]

[0018]

[0019] o t = σ(W o [h t-1 ,x t ]+b o )

[0020] h t = o t tanh(C t )

[0021] Wherein, f t represents the forget gate; i t represents the input gate; o tdenotes an output gate; W denotes a weight matrix corresponding to connecting two layers; b denotes a corresponding bias matrix; C t denotes a candidate vector of a storage unit at t time; denotes an update vector of a storage unit at t time; h t denotes an output vector of a model at t time; h t-1 denotes an output vector of a model at t-1 time; sigma denotes a sigmoid activation function; tanh denotes a hyperbolic tangent function, W f denotes a weight matrix of a forgetting gate, W i denotes a weight matrix of an input gate, W c denotes a weight matrix of a storage unit, W o denotes a weight matrix of an output gate, x t denotes an input at the current t time, b f denotes a bias matrix of a forgetting gate, b i denotes a bias matrix of an input gate, b c denotes a bias matrix of a storage unit, b o denotes a bias matrix of an output gate.

[0022] In the technical scheme, preferably, the gated recurrent unit model comprises a reset gate and an update gate, the reset gate controls the combination of memory information and new input information, and the update gate controls the input degree of previous state information in the current state.

[0023] In the technical scheme, preferably, the wavelet transform preprocessing decomposes the chlorophyll a concentration data by using a db4 wavelet function in a wavelet family for three layers, filters out high-frequency noise, and retains low-frequency data reflecting long-term trends after processing.

[0024] In the technical scheme, preferably, the prediction method further comprises model evaluation, and the marine chlorophyll concentration prediction model is evaluated in performance according to an average absolute error index, a root mean square error index and an average absolute percentage error index.

[0025] In the technical scheme, preferably, according to the evaluation indexes of the prediction results of chlorophyll a of different models after wavelet transform, the gated recurrent unit model after wavelet transform is selected as the chlorophyll a concentration prediction model in water.

[0026] The machine learning-based water source chlorophyll a concentration prediction model design method provided by the application has the following advantages compared with the prior art:

[0027] The machine learning-based water source chlorophyll a concentration prediction model design method of the present application adopts a WT-GRU model (wavelet-transformed gated recurrent unit model) to use wavelet transformation to pre-process data, significantly improving the accuracy of chlorophyll a concentration prediction. The wavelet transformation effectively extracts key time scale features by decomposing signals. The gating unit of the GRU model can adaptively adjust the information flow, enhancing the ability to capture long-term dependencies. The GRU reduces the number of parameters by simplifying the gating structure, thereby reducing the risk of overfitting and improving the generalization of the model. The WT-GRU has a significant advantage in predicting chlorophyll a concentration. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is the regression prediction fitting of the four machine learning models of the present application.

[0029] Figure 2 is the comparison of the predicted values and actual values of the four machine learning models of the present application. DETAILED DESCRIPTION

[0030] The specific embodiments of the present application are described in detail below. It should be understood that the specific embodiments described herein are only for illustration and explanation of the present application, and are not intended to limit the present application.

[0031] The machine learning-based water source chlorophyll a concentration prediction model design method of the present application comprises the following steps:

[0032] Step S1, selecting data: water quality monitoring data comes from the water station in a county-level water source, and the water quality monitoring indicators include 11 indicators such as chlorophyll a, water temperature, pH value, dissolved oxygen, turbidity, ammonia nitrogen, permanganate index, total nitrogen, and total phosphorus. The monitoring frequency is once every 4 hours. The data selected are daily monitoring results for 3 years.

[0033] Step S2, identifying the main factors affecting the concentration of chlorophyll a.

[0034] The relationship between chlorophyll a and its influencing indicators is nonlinear, and it is unnecessary and impractical to establish a model containing all influencing indicators. Therefore, the main factors affecting the concentration of chlorophyll a are identified by Pearson correlation analysis.

[0035] Step S3, data pre-processing.

[0036] In order to optimize the original data and improve the accuracy of the prediction model, wavelet transformation (WT) preprocessing is implemented. The selected monitoring data is decomposed into three layers by using db4 wavelet function in Daubechies wavelet family, and high-frequency noise is filtered out. The processed data retains the low-frequency components reflecting long-term trends.

[0037] Step S4: construct a prediction model.

[0038] S4-1, build a long short-term memory network (LSTM) model.

[0039] The key feature of the LSTM model is its cell structure (gating mechanism), which consists of three main parts: the input gate, which determines what new information is added to the cell state, the forget gate, which determines what old information is retained or forgotten in the cell state, and the output gate, which determines the output information of the current cell state.

[0040] f t =σ(W f [h t-1 ,x t ]+b f ) (1)

[0041] i t =σ(W i [h t-1 ,x t ]+b i ) (2)

[0042]

[0043] o t =σ(W o [h t-1 ,x t ]+b o ) (5)

[0044] h t =o t tanh(C t ) (6)

[0045] Where: f t represents the forget gate; i t represents the input gate; o t represents the output gate; W represents the weight matrix corresponding to the connection between the two layers; b represents the corresponding bias matrix; C t Represents the candidate vector of the storage unit at time t; represents the update vector of the storage unit at time t; h t represents the output vector of the model at time t; h t-1 represents the output vector of the model at time t-1; σ represents the sigmoid activation function; tanh represents the hyperbolic tangent function, W f Represents the weight matrix of the forget gate, W i Represents the weight matrix of the input gate, W c Represents the weight matrix of the storage unit, W oweight matrix of the output gate, x t input at the current time t, b f bias matrix of the forget gate, b i bias matrix of the input gate, b c bias matrix of the memory cell, b o bias matrix of the output gate.

[0046] S4-2, build a gated recurrent unit (GRU) model.

[0047] The GRU model is considered an efficient improvement form of the LSTM model, with a relatively simple structure and fewer internal parameters, which can improve computational efficiency. The GRU model mainly includes reset gate and update gate, which control the flow of information and the update of state. The reset gate controls the combination of memory information and new input information, and the update gate controls the input degree of previous state information in the current state.

[0048] S4-3, build a convolutional neural network (CNN) model.

[0049] The CNN model usually contains convolutional layers, activation layers, pooling layers, fully connected layers and other multi-layer networks. It extracts the features of input data through core convolutional operations, thereby improving the performance and generalization ability of the model, and is a classic deep learning model.

[0050] S4-4, build a multi-layer perceptron (MLP) model.

[0051] The MLP model is a feedforward neural network composed of three or more layers: input layer, one or more hidden layers, and output layer. Each layer is composed of multiple neurons or nodes, and the neurons between different layers are connected through weights. The MLP model can learn and simulate complex nonlinear relationships, so it is widely used in classification, regression and feature learning tasks.

[0052] Step S5, model evaluation.

[0053] The accuracy of the prediction results of different prediction models is compared by using three indicators: mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The specific indicators are shown in equations (7) to (9):

[0054]

[0055] Where: MAE represents the mean absolute error, RMSE represents the root mean square error, and MAPE represents the mean absolute percentage error, predicted value, y t actual value, n represents the number of samples.

[0056] Step S6, determine the prediction model.

[0057] Tensorflow's Keras platform was used in this prediction model confirmation. Four machine learning models were built and corrected to predict the change of chlorophyll a through the following steps:

[0058] (1) Data preprocessing, normalize the original data to ensure consistency of different monitoring indicators value range;

[0059] (2) Feature selection, based on the correlation between chlorophyll a and other monitoring indicators, select pH, water temperature, turbidity and chlorophyll a as input features;

[0060] (3) Data set division, divide the monitoring data set of 3 years of relevant indicators into training set (first 80%) and test set (last 20%), provide the basis for model training and evaluation;

[0061] (4) Model establishment, build CNN, MLP, LSTM, GRU four kinds of models, each model is designed for the characteristics of time series data;

[0062] (5) Model training, use Adam optimizer, set learning rate to 0.001, determine time window to 32 through random grid search, sample batch size to 16, hidden layer design to two, containing 32 and 16 neurons respectively, 50 times of iteration training;

[0063] (6) Model correction, use early stopping method to avoid overfitting, optimize hidden layer neuron number and network structure, improve model generalization ability;

[0064] (7) Performance evaluation, predict on the test set, restore the original value range through inverse normalization, and use MSE, R 2 and other indicators to evaluate the performance of the model.

[0065] The prediction evaluation results of chlorophyll a of the four models before and after wavelet transform are shown in Table 1. The prediction performance of the four models after wavelet transform is improved, among which the prediction accuracy of GRU is improved most significantly, followed by LSTM and CNN, and the improvement effect of MLP is the worst. The MAE, RMSE and MAPE evaluation indexes of LSTM are improved by 57.6%, 61.3% and 9.7 percentage points respectively compared with those before wavelet transform, the MAE, RMSE and MAPE evaluation indexes of GRU are improved by 75.4%, 78.9% and 9.8 percentage points respectively compared with those before wavelet transform, the MAE, RMSE and MAPE evaluation indexes of CNN are improved by 38.9%, 45.3% and 7 percentage points respectively compared with those before wavelet transform, and the MAE, RMSE and MAPE evaluation indexes of MLP are improved by 10.8%, 11.4% and 6.8 percentage points respectively compared with those before wavelet transform. This improvement is mainly due to the effective removal of high-frequency noise in the data by wavelet transform preprocessing, thereby improving the model's ability to capture the trend of water quality changes.

[0066] Among the four models after wavelet transform, the WT-GRU model shows the best performance in the evaluation indexes of the prediction results of chlorophyll a, with MAE, RMSE and MAPE values of 0.71 μg / L, 0.91 μg / L and 3.3% respectively, which are 48.9%, 48.0% and 3.8 percentage points lower than those of the WT-LSTM model, indicating that it has good advantages in capturing the complex patterns of time series data. The WT-LSTM model shows the second best prediction performance in the three prediction result evaluation indexes, with MAE, RMSE and MAPE values of 1.39 μg / L, 0.91 μg / L and 3.3% respectively, which are 49.5%, 47.9% and 6.2 percentage points lower than those of the WT-CNN model, showing good prediction ability. The prediction performance of the WT-CNN model is in the middle in this validation, with MAE, RMSE and MAPE values of 2.75 μg / L, 3.36 μg / L and 13.3% respectively, which may be because its convolutional layer structure is not sensitive to the time sequence of water quality information and cannot complete the prediction task well. The prediction performance of the WT-MLP model is the worst in this validation, with MAE, RMSE and MAPE values of 2.75 μg / L, 3.36 μg / L and 13.3% respectively. The MLP model does not have the memory function of the cycle structure, and it is difficult to handle non-stationarity, and its feature extraction ability is limited. The non-stationary change of chlorophyll a concentration with time leads to its poor performance in prediction results.

[0067] Table 1 Prediction evaluation results of chlorophyll a of four models

[0068] Model MAE / (pg / L) RMSE / (pg / L) MAPE / (%) LSTM 3.28 4.52 16.8 GRU 2.89 4.31 13.1 CNN 4.50 6.14 20.3 MLP 4.74 6.43 23.3 WT-LSTM 1.39 1.75 7.1 WT-GRU 0.71 0.91 3.3 WT-CNN 2.75 3.36 13.3 WT-MLP 4.23 5.70 16.5

[0069] The comparison of the predicted values and actual values of the four machine learning models in the test set is as followsFigure 1 (scatter plot) and Figure 2 (line plot) shown by Figure 1 It can be seen that the points in the scatter plot of WT-GRU are very closely distributed around the diagonal line, showing the smallest prediction error, the points in the scatter plot of WT-LSTM are relatively concentrated near the diagonal line, but slightly more dispersed than WT-GRU, the points in the scatter plot of WT-CNN show a certain degree of dispersion, indicating that the deviation between the predicted value and the actual value is slightly larger, and the points in the scatter plot of WT-MLP show a larger dispersion, which may mean that the model has a larger uncertainty or deviation in prediction. The correlation coefficient (R 2 ) values of WT-GRU, WT-LSTM, WT-CNN and WT-MLP are 0.9955, 0.9738, 0.9481 and 0.9115 respectively, and the fitting results of the predicted value and the actual value are WT-GRU>WT-LSTM>WT-CNN>WT-MLP. This result shows that the WT-GRU model can effectively handle the long-term dependence problem in time series data and adapt to the nonlinear and complex patterns in the data, which is consistent with the fact that the WT-GRU model performs best in the three evaluation indicators of predicting chlorophyll a.

[0070] The WT-LSTM model of the application can effectively capture the long-term dependence relationship in time series data by introducing the gating mechanism, which makes it also perform well in handling data with time continuity. The WT-CNN model can capture local features through convolutional layers, but when processing time series data, it may not be able to fully utilize the global information and long-term dependence of the data. In contrast, the WT-MLP model has the smallest R 2 value and the predicted value is relatively loose, which means that there is still a large error in the prediction result of the model. MLP lacks the ability to capture memory and sequence dependence in time series data and cannot effectively handle non-stationary data. When dealing with time series problems such as changes in chlorophyll a concentration in water sources, the limitations of the MLP model may lead to its inability to accurately predict future concentration changes, thereby affecting the timely response and disposal of environmental problems such as algal blooms in water sources.

[0071] Figure 2 ​It can be seen that the WT-GRU prediction value is the best, followed by the WT-LSTM, and the WT-CNN and WT-MLP are relatively poor. Especially in the high value area and data mutation, WT-CNN and WT-MLP are difficult to accurately capture the rapid changes of chlorophyll a concentration, and the overall prediction value is relatively flat, which cannot reflect the actual trend; WT-LSTM shows a certain limitation in that the prediction value is lower than the actual value at some moments. In contrast, the prediction result of the WT-GRU model is obviously better than the other three models, which may be because the unique gating mechanism of the GRU model can better learn the long-term and short-term patterns in the data, and when dealing with data with complex time dependence, it can provide more accurate and adaptive prediction results. Especially in the case of data mutation, the GRU model can quickly adapt to the new pattern without prediction lag or deviation.

[0072] The SARIMA model shows high accuracy in predicting the concentration of chlorophyll a in a certain lake water body, but there is still a large gap between the predicted value and the measured value in some specific months, showing certain deficiencies, while the advantage of the WT-GRU model lies in its ability to handle the nonlinear characteristics of time series data, which is particularly important when dealing with complex water environment systems. The support vector regression model combined with the random forest recursive feature elimination method significantly improves the prediction accuracy of chlorophyll a in water bodies, with a correlation coefficient of 0.94, but lower than that of the WT-GRU model, indicating that the WT-GRU model performs better in generalization ability and adaptability due to its deep learning characteristics. Based on the NARX neural network, the correlation coefficient of the short-term prediction of chlorophyll a concentration in Thousand Island Lake remains between 0.953 and 0.986, and the mean square error is between 1.387 and 4.483 μg / L. In contrast, the WT-GRU model shows lower RMSE and higher correlation coefficient, indicating that the WT-GRU model may be more sensitive and accurate in capturing the dynamic changes of time series data. When using the LSTM model to predict the changes of chlorophyll a concentration in the Three Gorges Reservoir, the RMSE is 0.049-0.221 μg / L. In contrast, the RMSE of the WT-GRU model is slightly higher, but the overall trend prediction is very accurate. The WT-GRU model of the present invention uses wavelet transform to preprocess data, significantly improving the accuracy of chlorophyll a concentration prediction. Wavelet transform effectively extracts key time scale features by decomposing signals, which is crucial for time series analysis. The gating unit of the GRU model can adaptively adjust the information flow, enhancing the ability to capture long-term dependencies. Compared with LSTM, GRU simplifies the gating structure, reduces the number of parameters, and thus reduces the risk of overfitting, improving the generalization of the model. These improvements make WT-GRU have a significant advantage in predicting chlorophyll a concentration. Therefore, due to its deep learning architecture, the WT-GRU model can better handle complex nonlinear relationships and long-term dependencies, and it is a good choice to use the WT-GRU model to predict algae in plain river network water sources and improve early warning of water bloom.

[0073] The above implementation cases are only preferred embodiments of the present application, and do not limit the present application in any form. Although the present application has been disclosed as above with preferred embodiments, it is not intended to limit the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the present application, according to the technical essence of the present application, shall fall within the protection scope of the technical solution of the present application.

Claims

1. A method for designing a chlorophyll a concentration prediction model for a water source based on machine learning, characterized in that: include: Obtain chlorophyll a concentration data for a continuous period of time in the area to be predicted; The chlorophyll a concentration data were preprocessed by using wavelet transform to filter out high-frequency noise; Construct different concentration prediction models, input the processed chlorophyll a concentration data and physical and chemical parameters into the prediction model, and obtain the chlorophyll a concentration data prediction results; Compare the prediction results of different prediction models and determine the prediction model.

2. The method for designing a water source chlorophyll a concentration prediction model based on machine learning according to claim 1, characterized in that: Constructing different concentration prediction models at least includes constructing a long short-term memory network model, constructing a gated recurrent unit model, constructing a convolutional neural network model, and constructing a multi-layer perceptron model.

3. The method for designing a water source chlorophyll a concentration prediction model based on machine learning according to claim 1, characterized in that: The long short-term memory network model includes an input gate, a forget gate, and an output gate. The input gate determines the new information added to the unit, the forget gate determines the old information to be retained or forgotten in the unit state, and the output gate determines the output information of the current unit state. The long short-term memory network model is: f t =σ(W f [h t-1 ,x t ]+b f ) i t =σ(W i [h t-1 ,x t ]+b i ) the t =σ(W o [h t-1 ,x t ]+b o ) h t =o t fishy(C) t ) Among them, f t represents the forget gate; i t represents the input gate; o t represents the output gate; W represents the weight matrix corresponding to the connection between the two layers; b represents the corresponding bias matrix; C t Represents the candidate vector of the storage unit at time t; represents the update vector of the storage unit at time t; h t represents the output vector of the model at time t; h t-1 represents the output vector of the model at time t-1; σ represents the sigmoid activation function; tanh represents the hyperbolic tangent function, W f Represents the weight matrix of the forget gate, W i Represents the weight matrix of the input gate, W c Represents the weight matrix of the storage unit, W o represents the weight matrix of the output gate, x t represents the input at the current time t, b f represents the bias matrix of the forget gate, b i represents the bias matrix of the input gate, b c represents the bias matrix of the storage unit, b o Represents the bias matrix of the output gate.

4. The method for designing a water source chlorophyll a concentration prediction model based on machine learning according to claim 2, wherein: The improved prediction model adopts a gated recurrent unit model, which includes a reset gate and an update gate. The reset gate controls the combination of memory information and new input information, and the update gate controls the input degree of previous state information in the current state.

5. The method for designing a water source chlorophyll a concentration prediction model based on machine learning according to claim 2, characterized in that: The wavelet transform preprocessing uses the db4 wavelet function in the wavelet family to perform three-layer decomposition on the chlorophyll a concentration data, filter out high-frequency noise, and retain low-frequency data reflecting long-term change trends in the processed data.

6. The method for designing a water source chlorophyll a concentration prediction model based on machine learning according to claim 5, characterized in that: The prediction method also includes model evaluation, which evaluates the performance of the ocean chlorophyll concentration prediction model based on the mean absolute error index, the root mean square error index and the mean absolute percentage error index.

7. The method for designing a water source chlorophyll a concentration prediction model based on machine learning according to claim 5, characterized in that: According to the evaluation indexes of the prediction results of chlorophyll a by different models after wavelet transform, the chlorophyll a concentration prediction model in water body of gated cyclic unit model after wavelet transform was selected.

Images