Multi-index sewage water quality prediction method based on parallel CNN-GRU network

Abstract

The present application relates to a kind of multi-index sewage water quality prediction method based on parallel CNN-GRU network, comprising the following steps: after data preprocessing to historical inflow and effluent data, construct inflow and effluent two-dimensional matrix, form data set and carry out division, wherein inflow and effluent two-dimensional matrix are respectively composed of the value of multiple inflow and effluent input index;Build multi-network combination model, which includes two parallel CNN-GRU networks and DNN networks;The model is trained to obtain a trained model;Using the trained model to predict the historical water quality data to be predicted, obtain the multi-index sewage water quality prediction result, including the value of effluent chemical oxygen demand, effluent PH value, effluent ammonia nitrogen, effluent total phosphorus and effluent total nitrogen at future time point.The present application uses the prediction method of multi-index input, multi-index output, can make full use of the biochemical relationship between other water quality indicators and prediction indicators, thereby improving the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the fields of wastewater treatment and artificial intelligence, and in particular to a multi-indicator wastewater quality prediction method based on a parallel CNN-GRU network. This method uses different neural network combinations to predict multiple indicators of future effluent using historical data of effluent and influent. Background Technology

[0002] Wastewater treatment plays a vital role in water resource utilization, especially in urban areas. With urban development and rapid population growth, the volume of urban wastewater has increased dramatically, placing immense pressure on urban wastewater treatment. The most widely used wastewater treatment process in urban wastewater treatment plants is the CASS (Continuous Adaptive Fluid Process), but because CASS involves intermittent treatment, its process adjustment is more complex. Effluent quality prediction algorithms can provide data support and optimization basis for production processes, making them a crucial element in process adjustment.

[0003] Most current wastewater quality prediction methods adopt a single-indicator input to predict a single indicator. This method ignores the complexity of wastewater composition and the interaction between various components, resulting in low accuracy of water quality prediction. Summary of the Invention

[0004] To address the problem that traditional single-index prediction methods suffer from low accuracy due to neglecting the complexity of wastewater composition and the interactions between its components, this invention provides a multi-index wastewater quality prediction method based on a parallel CNN-GRU network. This method uses historical influent and effluent data from actual wastewater treatment plant production to train a multi-network combined model, and then uses the trained model to predict future effluent data, thus achieving prediction of multiple indicators in future effluent quality.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A multi-index wastewater quality prediction method based on a parallel CNN-GRU network includes the following steps:

[0007] Step 1: Obtain historical influent and effluent data from the wastewater treatment plant, and perform data preprocessing on the historical influent and effluent data, including outlier detection, missing value filling, and data normalization.

[0008] Step 2: Construct two-dimensional matrices for inflow and outflow based on the processed historical inflow and outflow data to form a dataset, and divide the dataset into a training set, a test set, and a validation set.

[0009] In the two-dimensional water inflow matrix, each element in the same row represents the value of multiple water inflow input indicators at the same historical time point, and the historical time points of each row are different. The multiple water inflow input indicators are air temperature, precipitation, water inflow chemical oxygen demand, water inflow pH value, water inflow ammonia nitrogen content, water inflow flow rate, aeration time, average aeration dissolved oxygen content, and average aeration dissolved oxygen content variance.

[0010] In the effluent two-dimensional matrix, the historical time points of each row correspond one-to-one with the historical time points of each row in the influent two-dimensional matrix, and each element in the same row is the value of multiple effluent input indicators at the same historical time point. The multiple effluent input indicators are effluent chemical oxygen demand, effluent pH value, effluent ammonia nitrogen, effluent total phosphorus and effluent total nitrogen.

[0011] Step 3: Construct a multi-network combined model, which includes two parallel CNN-GRU networks and a DNN network connected to the CNN-GRU network. The CNN network is used to obtain the weights in the data and input them into the GRU network. The GRU network is used to obtain the water quality time series and input it into the DNN network. The DNN network performs effluent water quality prediction and outputs multi-index wastewater water quality prediction data for future time points. The multi-index wastewater water quality prediction data includes the values ​​of effluent chemical oxygen demand, effluent pH value, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen.

[0012] Step 4: Input the inflow and outflow 2D matrices from the training set into the two parallel CNN-GRU networks of the multi-network ensemble model for training, and find the optimal weights and biases of the network. After each training iteration, input the inflow and outflow 2D matrices from the test set into the two parallel CNN-GRU networks of the multi-network ensemble model to find the optimal hyperparameters, loss function, optimizer, and activation function. The hyperparameters include the kernel size, the number of hidden layers in each neural network, the learning rate, and the number of training epochs. The loss function is the absolute mean error, the optimizer is adaptive matrix estimation, and the activation function is the tanh function. After multiple rounds of training and testing, the trained model is validated using a validation set. The evaluation metrics used in the validation process are goodness of fit, mean squared error, and mean percentage error.

[0013] Step 5: Use the model trained in Step 4 to predict the historical water quality data to be predicted, and obtain the multi-index wastewater quality prediction results.

[0014] The main beneficial effects of this invention are as follows:

[0015] (1) The multi-index wastewater quality prediction method proposed in this invention is a prediction method with multiple index inputs and multiple index outputs. Compared with the traditional single index prediction method, this invention makes full use of the biochemical relationship between other water quality indicators and prediction indicators, thereby improving the prediction accuracy.

[0016] (2) The multi-network combination model in this invention adopts two CNN-GRU networks for parallel operation. The CNN-GRU network combines the ability of convolutional neural networks to efficiently extract data features and the ability of recurrent neural networks to process temporal relationships. By using the parallel operation of the two CNN-GRU networks, the effluent water quality can be predicted from two dimensions: the biochemical relationship between the effluent and the influent, and the temporal relationship within the time series of the effluent water quality itself. Furthermore, the model proposed in this invention has a simple structure, is efficient to implement, has a small error rate, and has achieved relatively ideal results in the prediction of wastewater effluent.

[0017] (3) By constructing a combined neural network, this invention can construct a biochemical model between water inlet and outlet, and a temporal relationship within the water. At the same time, it adopts multi-index prediction on input and output indicators, enabling it to be applied in complex water environments. Attached Figure Description

[0018] Figure 1 This is a flowchart of a multi-index wastewater quality prediction method based on a parallel CNN-GRU network according to one embodiment of the present invention;

[0019] Figure 2 This is a time-series diagram showing the relationship between the input and output of the multi-network ensemble model in this invention;

[0020] Figure 3 This is a diagram showing the specific index relationship between the input and output of the multi-network combination model in this invention;

[0021] Figure 4 This is a schematic diagram of the structure of the multi-network combination model in this invention;

[0022] Figure 5 A schematic diagram of the CNN network structure used in multi-network combination models;

[0023] Figure 6 A schematic diagram of the GRU network structure used in the multi-network combination model. Detailed Implementation

[0024] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0025] In one embodiment, such as Figure 1As shown, this invention provides a multi-index wastewater quality prediction method based on a parallel CNN-GRU network, which specifically includes the following steps:

[0026] Step 1: Obtain historical influent and effluent data from the wastewater treatment plant, and perform data preprocessing on the historical influent and effluent data, including outlier detection, missing value filling, and data normalization.

[0027] In this step, historical data extracted from the wastewater treatment plant is preprocessed. The data preprocessing process specifically includes:

[0028] Outlier detection: First, query the collected actual production data of the wastewater treatment plant for outliers (excessive values, negative values) and delete them, leaving the corresponding gaps for the deleted outliers;

[0029] Imputation of missing values: Then, linear interpolation is used to fill in the missing values. The formula for linear interpolation is:

[0030] x k =x w +(x r -x w (kw) / (rw)

[0031] In the formula: k is the sequence number of the missing data, w is the sequence number of the nearest known data preceding the missing data, r is the sequence number of the nearest known data following the missing data, and x k x is the value to be filled in. w For x k The most recent known data; x r For x k The most recent known data;

[0032] Data normalization: To accelerate gradient descent, the water quality data is normalized using a formula, employing max-min normalization to scale all data to the range [0,1]. The normalization formula is:

[0033] x′=(xx min ) / (x max -x min )

[0034] In the formula, x is the data to be normalized; x′ is the normalized data; x max x min These represent the maximum and minimum values ​​of the input data, respectively.

[0035] Step 2: Construct two-dimensional influent and effluent matrices based on the processed historical influent and effluent data to form a dataset, and divide the dataset into training, testing, and validation sets.

[0036] In the two-dimensional water inlet matrix, each element in the same row represents the value of multiple water inlet input indicators at the same historical time point, and the historical time points of each row are different. The multiple water inlet input indicators are air temperature, precipitation, water inlet chemical oxygen demand (water inlet COD), water inlet pH value, water inlet ammonia nitrogen, water inlet flow rate, aeration time, average aeration dissolved oxygen, and average aeration dissolved oxygen variance. Each indicator can be added or deleted according to the actual production situation.

[0037] In the effluent two-dimensional matrix, the historical time points of each row correspond one-to-one with the historical time points of each row in the influent two-dimensional matrix. Moreover, each element in the same row represents the value of multiple effluent input indicators at the same historical time point. These multiple effluent input indicators are effluent chemical oxygen demand (effluent COD), effluent pH value, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen. Each indicator can be added or deleted according to the actual production situation.

[0038] like Figure 2 As shown, taking the influent two-dimensional matrix and effluent two-dimensional matrix both using the values ​​of multiple influent input indicators (influent data) and multiple effluent input indicators (effluent data) at 24 historical time points (t1~t24) as input, and predicting the multi-indicator wastewater quality forecast data (effluent data) at one future time point (t25) as an example, and as... Figure 3 As shown, in this example, the multiple influent input indicators are air temperature, precipitation, influent chemical oxygen demand (COD), influent pH, influent ammonia nitrogen, influent flow rate, aeration time, average dissolved oxygen (DO), and average DO variance. The multiple effluent input indicators are effluent COD, effluent pH, effluent ammonia nitrogen, effluent total phosphorus (TNP), and effluent total nitrogen (TN). The predicted wastewater quality data for these multiple indicators are the values ​​of effluent COD, effluent pH, effluent ammonia nitrogen, effluent TNP, and effluent TN. Details are as follows:

[0039] The input to the multi-network combined model consists of a two-dimensional matrix composed of water inflow and outflow data from the past 24 points, specifically:

[0040]

[0041]

[0042] in, Let represent the values ​​of temperature, precipitation, chemical oxygen demand (COD) of influent, pH value of influent, ammonia nitrogen of influent, influent flow rate, aeration time, average dissolved oxygen concentration (DOC) and variance of average dissolved oxygen concentration (DOC) at the ti-th historical time point, respectively. These represent the values ​​of chemical oxygen demand (COD), pH, ammonia nitrogen, total phosphorus, and total nitrogen in the effluent at the ti-th historical time point, respectively.

[0043] The output of the multi-network combined model is a one-dimensional matrix composed of effluent water quality data at the 25th time point, specifically:

[0044]

[0045] in, These represent the values ​​of chemical oxygen demand (COD), pH, ammonia nitrogen, total phosphorus, and total nitrogen in the effluent at the 25th historical time point.

[0046] Dataset partitioning: The dataset is divided into training, test, and validation sets in a 6:2:2 ratio. The training set is used to train the model and find the optimal weights and biases; the test set is used to filter the trained model and find the optimal hyperparameters; the validation set is used only for performance evaluation of the trained model.

[0047] Step 3: Construct a multi-network combination model.

[0048] The multi-network combined model in this embodiment adopts a multi-input (historical inflow data, historical outflow data) and single-output (future outflow data) neural network model, using a combination of gated recurrent neural networks (GRU), convolutional neural networks (CNN), and fully connected neural networks (DNN). A pre-defined training set is input into the model for training, and a test set is input into the model after each training iteration. The specific parameters of the model are determined by comparing the model metrics output from the training set and the test set.

[0049] like Figure 4 As shown, the multi-network combined model includes two parallel CNN-GRU networks and a DNN network connected to the two parallel CNN-GRU networks. The CNN network is used to obtain the weights in the data and input them into the GRU network. The GRU network is used to obtain the water quality time series and input it into the DNN network. The DNN network performs effluent water quality prediction and outputs multi-index wastewater water quality prediction data for future time points. The multi-index wastewater water quality prediction data includes the values ​​of effluent chemical oxygen demand, effluent pH, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen.

[0050] When training the multi-network combined model, the inflow and outflow two-dimensional matrices are first output to two parallel CNNs (Convolutional Neural Networks) to obtain the weights in the data, then input to GRUs (Gated Recurrent Neural Networks) to obtain the temporal relationships, and finally input to DNNs (Fully Connected Neural Networks) to output the outflow prediction data. The specific neural network usage is as follows:

[0051] First, the influent two-dimensional matrix and the effluent two-dimensional matrix are input into a CNN network. Each CNN network contains two convolutional layers. In this invention, both convolutional layers are one-dimensional convolutions, performing convolution only on a single water quality index within the input matrix, i.e., the matrix only undergoes vertical convolution. The kernel size is 1×N. Figure 5 As shown, N is the time length of a convolution in a single sequence of the input matrix.

[0052] The feature extraction data after convolutional layers is input into two GRUs respectively to obtain the water quality time series. A single GRU is like... Figure 6 As shown, its main components are the reset gate and the update gate. The formula for the reset gate is:

[0053] r t =σ(W r x t +U r h t-1 +b r )

[0054] Where σ is the sigmoid function, which can be used to transform the input data between 0 and 1, and W... r U r Both are weight matrices; b r This is the bias matrix.

[0055] The updated gate formula is:

[0056] z t =σ(W z x t +U z h t-1 +b z )

[0057] Where σ is the sigmoid function, W z U z Both are weight matrices; b z This is the bias matrix.

[0058] First, we use the state h transmitted from the previous file. t-1 and the input x of the current node t We obtain two gating states. Then, we determine the amount of past information to forget by resetting the gates, and determine which information needs to be discarded and which new information needs to be added by updating the gates. Finally, we perform a state update, and the updated formula is shown below:

[0059]

[0060]

[0061] Among them, W hU h Both are weight matrices; b h This is the bias matrix.

[0062] The water quality time series output by the two GRUs are input into a DNN to predict future water quality. The DNN outputs the future water quality to be predicted.

[0063] Step 4: Train the multi-network combination model constructed in Step 3 to obtain the trained model, and then validate the trained model.

[0064] In this step, the inflow and outflow 2D matrices from the training set are input into the two parallel CNN-GRU networks of the multi-network ensemble model for training, to find the optimal weights and biases. After each training iteration, the inflow and outflow 2D matrices from the test set are input into the two parallel CNN-GRU networks of the multi-network ensemble model to find the optimal hyperparameters, loss function, optimizer, and activation function. The hyperparameters include kernel size, number of hidden layers in each neural network, learning rate, and number of training epochs. The loss function is the absolute mean error, the optimizer is adaptive matrix estimation, and the activation function is the tanh function. After multiple rounds of training and testing, the trained model is validated using a pre-defined validation set. The evaluation metric used in the validation process is goodness of fit (R²). 2 ), mean squared error (RMSE), and mean percentage error (MAPE).

[0065] R 2 The closer the value is to 1, the better the regression equation fits the observed values; conversely, the closer the value is to 1, the worse the regression equation fits the observed values.

[0066] RMSE, or Root Mean Square Error, is the square root of the ratio of the sum of squares of the deviations between the predicted and actual values ​​to the number of observations. It is used to measure the degree of deviation between the predicted and observed values.

[0067] MAPE stands for Mean Percentage Error, which is the process of comparing the error with the original data value. The closer the result is to 0, the more perfect the model is, and the greater than 100% is the poor model.

[0068] Step 5: Use the model trained in Step 4 to predict the historical water quality data to be predicted, and obtain the multi-index wastewater quality prediction results.

[0069] Following steps 1 to 2, the historical water quality data to be predicted is processed into a matrix that can be input into a multi-network combined model. The matrix is ​​then input into the model, and the model finally outputs multi-index wastewater quality prediction results. These multi-index wastewater quality prediction results include the predicted values ​​of effluent chemical oxygen demand, effluent pH, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen at the predicted future time points.

[0070] The following example, using the application of this invention in a wastewater treatment plant in Jilin Province, illustrates the technical solution and predicted effects of this invention.

[0071] Step 1: Data Preprocessing. The wastewater treatment plant data includes data from January 1, 2020 to June 30, 2020, totaling 8,640 time points. This includes effluent data (effluent COD, pH value, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen) and influent data (temperature, precipitation, influent COD, influent pH value, influent ammonia nitrogen, influent flow rate, aeration time, average dissolved oxygen concentration (DOC), and average DOC variance).

[0072] Remove outliers (excessively large or negative values) from the data and leave gaps.

[0073] Then, linear interpolation is used to fill in the missing values.

[0074] Data normalization: In order to accelerate the gradient descent speed, the water quality parameters are normalized according to the formula, and the max-min normalization is used to scale all input parameters to the range of [0,1].

[0075] Step two, take the following steps: Figure 2 The diagram shows multiple indicator inputs and outputs, using historical data from 24 time points as input to predict one future time point (output); it employs methods such as... Figure 3 The data shown includes multiple influent input indicators (air temperature, precipitation, influent chemical oxygen demand, influent pH value, influent ammonia nitrogen, influent flow rate, aeration time, average dissolved oxygen content and average dissolved oxygen content variance) and multiple effluent input indicators (effluent chemical oxygen demand, effluent pH value, effluent ammonia nitrogen content, effluent total phosphorus content and effluent total nitrogen content).

[0076] In this example, the time interval is 1, and the generated input and output data includes two tensor inputs containing 8617 two-dimensional matrices and a tensor output containing 8617 one-dimensional matrices.

[0077] Dataset partitioning: The dataset is partitioned into training, test, and validation sets in a 6:2:2 ratio. This partitioning of the tensors results in a training set with tensor inputs of 5170 two-dimensional matrices and tensor outputs of 5170 one-dimensional matrices. A test set contains tensor inputs of 1723 two-dimensional matrices and tensor outputs of 1723 one-dimensional matrices. A validation set contains tensor inputs of 1724 two-dimensional matrices and tensor outputs of 1723 one-dimensional matrices.

[0078] Step 3: Construct a multi-network combination model.

[0079] First, the two input matrices are fed into a CNN (Convolutional Neural Network) to obtain the weights in the data. Then, they are fed into a GRU (Gated Recurrent Neural Network) to obtain the temporal relationships. Finally, they are fed into a DNN (Fully Connected Neural Network) to output the water outflow prediction data. The specific neural network usage is as follows:

[0080] The combined model of this invention first inputs two input matrices into a CNN containing two convolutional layers. The one-dimensional convolution used in this invention performs convolution only on a single water quality indicator, i.e., the matrix only undergoes vertical convolution. The kernel size is 1×3, and the convolution operation within a single indicator is performed over a time interval of size 3. Figure 5 As shown.

[0081] The feature extraction data after convolutional layers are input into two GRUs respectively to obtain the water quality time series.

[0082] The loss function selection formula for MAE (Mean Absolute Error) is as follows:

[0083]

[0084] Where n is the sequence length, f(x) is the predicted value obtained through the neural network, and y is the true value.

[0085] The optimizer selects the Adam (Adaptive Moment Estimation) adaptive matrix estimate as follows:

[0086] m t := beta1*m t-1 +(1-beta1)*g

[0087] v t := beta2*v t-1 +(1-beta2)*g*g

[0088]

[0089] Where beta1 and beta2 are the exponential decay rates, which are generally close to 1 and default to 0.9, and g is the loss function θ.t Find the partial derivative of m. t and v t They are first-order and second-order variables, respectively, lr t This is the learning rate.

[0090] The activation function chosen for each layer is the hyperbolic tangent function (tanh function).

[0091] Step 4: Model Training: The previously divided training set is input into the model for training. After each training session, the test set is input into the model to select appropriate model hyperparameters. The hyperparameters obtained in the experiment are as follows: convolutional kernel size 1x3, number of hidden layers in CNN layer 128, number of hidden layers in GRU layer 128, number of hidden layers in DNN layer 128, learning rate 0.0001, and number of training epochs 1000.

[0092] The model was validated using a pre-defined validation set, and the evaluation metric was goodness of fit (R²). 2 The metrics include mean squared error (RMSE) and mean percentage error (MAPE). A comparison of relevant metrics between existing methods and the method proposed in this invention is shown in Table 1:

[0093] Table 1 Comparison of relevant indicators between existing methods and the method proposed in this invention.

[0094] Existing methods 0.73 0.43 0.37 This method 0.85 0.24 0.21

[0095] As shown in Table 1, the multi-network combination model based on parallel CNN-GRU network in this invention has a good effect on wastewater effluent prediction. By constructing dual inputs of influent and effluent data, the prediction error is reduced and the prediction accuracy is improved.

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0097] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A multi-index wastewater quality prediction method based on a parallel CNN-GRU network, characterized in that, Includes the following steps: Step 1: Obtain historical influent and effluent data from the wastewater treatment plant, and perform data preprocessing on the historical influent and effluent data, including outlier detection, missing value filling, and data normalization. Step 2: Construct two-dimensional matrices for inflow and outflow based on the processed historical inflow and outflow data to form a dataset, and divide the dataset into a training set, a test set, and a validation set. In the two-dimensional water inflow matrix, each element in the same row represents the value of multiple water inflow input indicators at the same historical time point, and the historical time points of each row are different. The multiple water inflow input indicators are air temperature, precipitation, water inflow chemical oxygen demand, water inflow pH value, water inflow ammonia nitrogen content, water inflow flow rate, aeration time, average aeration dissolved oxygen content, and average aeration dissolved oxygen content variance. In the effluent two-dimensional matrix, the historical time points of each row correspond one-to-one with the historical time points of each row in the influent two-dimensional matrix, and each element in the same row is the value of multiple effluent input indicators at the same historical time point. The multiple effluent input indicators are effluent chemical oxygen demand, effluent pH value, effluent ammonia nitrogen, effluent total phosphorus and effluent total nitrogen. Step 3: Construct a multi-network combined model, which includes two parallel CNN-GRU networks and a DNN network connected to the CNN-GRU network. The CNN network is used to obtain the weights in the data and input them into the GRU network. The GRU network is used to obtain the water quality time series and input it into the DNN network. The DNN network performs effluent water quality prediction and outputs multi-index wastewater water quality prediction data for future time points. The multi-index wastewater water quality prediction data includes the values ​​of effluent chemical oxygen demand, effluent pH value, effluent ammonia nitrogen, effluent total phosphorus, and effluent total nitrogen. Step 4: Input the inflow 2D matrix and outflow 2D matrix from the training set into the two parallel CNN-GRU network of the multi-network combination model to train the model and find the optimal weights and biases of the network. After each training iteration, the ingress and egress two-dimensional matrices from the test set are input into the two parallel CNN-GRU networks of the multi-network ensemble model to find the optimal hyperparameters, loss function, optimizer, and activation function. The hyperparameters include the kernel size, the number of hidden layers in each neural network, the learning rate, and the number of training epochs. The loss function is the absolute mean error, the optimizer is the adaptive matrix estimation, and the activation function is the tanh function. After multiple rounds of training and testing, the trained model is validated using a validation set. The evaluation metrics used in the validation process are goodness of fit, mean squared error, and mean percentage error. Step 5: Use the model trained in Step 4 to predict the historical water quality data to be predicted, and obtain the multi-index wastewater quality prediction results.

2. The multi-index wastewater quality prediction method based on parallel CNN-GRU network according to claim 1, characterized in that, The data preprocessing process includes the following steps: Outlier detection: Remove excessively large and negative values ​​from the data and leave the corresponding gaps; Filling in missing values: Using linear interpolation to fill in missing values; Data normalization: Use max-min normalization to scale all data to the range [0,1].

3. The multi-index wastewater quality prediction method based on parallel CNN-GRU network according to claim 1, characterized in that, Each CNN network contains two convolutional layers, both of which are one-dimensional convolutions. They perform vertical convolutions only on either the ingress or egress two-dimensional matrix, and their kernel size is 1×N. ， Where N is the time length of a convolution in a single sequence of the ingress or egress two-dimensional matrix.

4. The multi-index wastewater quality prediction method based on parallel CNN-GRU network according to claim 1, characterized in that, The hyperparameters are: kernel size 1x3, number of hidden layers in CNN layer 128, number of hidden layers in GRU layer 128, number of hidden layers in DNN layer 128, learning rate 0.0001, and number of training epochs 1000.

5. The multi-index wastewater quality prediction method based on parallel CNN-GRU network according to claim 1, characterized in that, The training set, test set, and validation set are divided in a 6:2:2 ratio.

Images