A Water Quality Parameter Prediction Method Based on Sparse Labeled Long Short-Term Memory Networks

Abstract

This invention discloses a water quality parameter prediction method based on sparse labeling and Long Short-Term Memory (LSTM) networks. It selects the top *s* water quality indicators with the highest correlation to the indicator to be predicted, and further optimizes a subset of these indicators to form an input feature set *Si*. The hyperparameters of the water quality prediction model are varied, and the model is trained using the set *Si* to determine the optimal hyperparameters. The resulting water quality prediction model is then trained using the data corresponding to the input feature set *Si*, yielding the final model. In actual prediction, the water quality indicators from existing monitoring data that correspond to the input feature set are used as the input to the model, and the output is the indicator to be predicted. This invention can predict the water quality data of this indicator over a future period with relatively low water quality testing workload, enabling real-time closed-loop control in wastewater treatment processes and improving control accuracy.

Description

Technical Field

[0001] This invention relates to improvements in water quality testing technology, specifically to a method for predicting water quality parameters based on sparse labeling long short-term memory networks, belonging to the field of water quality testing technology. Background Technology

[0002] The scientific and rational development, utilization, and protection of water resources is a crucial component of the national ecological strategy. Wastewater treatment and domestic sewage treatment are concrete manifestations of water conservation and environmental protection policies. Obtaining water quality data is a vital aspect of wastewater treatment, forming the basis for closed-loop control, improved treatment efficiency, and energy savings. Currently, some important water quality parameters are difficult to obtain in real time using sensors, requiring manual water sample collection and laboratory analysis. The water quality indicator testing process mainly includes sampling, sample registration, sample testing, record keeping, and report submission, consuming significant time, manpower, financial resources, and management costs. Furthermore, these parameters have long time intervals and exhibit temporal sparsity, making them unsuitable for supporting closed-loop real-time control in wastewater treatment.

[0003] Considering the limitations of existing water quality indicator testing, the advantages of water quality change trend prediction technologies based on artificial intelligence and machine learning become apparent. The aim is to analyze water quality parameters that can be obtained in real time through sensors, uncovering their correlation with the parameters requiring laboratory testing, and thus obtaining predicted values ​​for wastewater treatment. LSTM (Laser-Surveyed Mechanism) is a recurrent neural network suitable for processing and predicting events with relatively long intervals and delays in time series. Using LSTM to build a prediction model can solve the delay problem caused by the sampling and testing process, reduce the number of water quality tests, alleviate the workload of water quality testing, obtain relevant parameters for wastewater treatment in a timely manner, improve wastewater treatment efficiency, and transform water pollution from a traditional reactive approach to a proactive preventative approach. This is of great significance for water pollution prevention. Summary of the Invention

[0004] To address the shortcomings of existing water quality testing methods, such as low timeliness and high acquisition costs for some water quality indicators, the present invention aims to provide a water quality parameter prediction method based on sparse labeling long short-term memory networks. This invention can predict the water quality data of the indicator for a future period with a lower workload of water quality testing, realize real-time closed-loop control in the wastewater treatment process, improve control accuracy, and save energy.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The method for predicting water quality parameters based on sparsely labeled long short-term memory networks comprises the following steps.

[0007] 1) Data preprocessing; preprocessing the historical monitoring data of several water quality indicators obtained in advance;

[0008] 2) Determine the relevant water quality indicators; calculate the correlation between the indicator Xi to be predicted and the indicators involved in the preprocessed data, select the s water quality indicators with the highest correlation, and form a set of relevant water quality indicators S={Xi(1),Xi(2),…Xi(s)}, where Xi(s) represents the s-th water quality indicator associated with Xi; the indicator Xi to be predicted is one of the indicators involved in the preprocessed data;

[0009] 3) Select some water quality indicators from the associated water quality indicator set S to form the input feature set Si of the water quality prediction model; the indicator to be predicted Xi is the output of the water quality prediction model; the input feature set Si of the water quality prediction model includes the indicator to be predicted Xi.

[0010] 4) Based on the water quality prediction model input feature set Si obtained in step 3), change the hyperparameters of the water quality prediction model to establish water quality prediction models with different hyperparameters. Then, use the water quality prediction model input feature set Si to train these water quality prediction models with different hyperparameters respectively, and use the corresponding test set to test them. Compare the prediction errors in the water quality prediction models with different hyperparameters. The hyperparameters of the model with the smallest RMSE are the hyperparameters of the water quality prediction model.

[0011] 5) Based on the water quality prediction model obtained in step 4) and the input features of the water quality prediction model obtained in step 3), the water quality prediction model is trained using the data corresponding to the input feature set Si to obtain the final water quality prediction model. During training, the time interval T between the measured values ​​of the input features corresponding to the indicator to be predicted is first set. For the input features corresponding to the indicator to be predicted, the measured value of the indicator is input at time t+n*T, where n is a natural number. The predicted value is used as the input for the next time step at other times. For other input features, the measured value is input at any time step.

[0012] 6) In actual prediction, the water quality indicators in the existing water quality indicator monitoring data that correspond to the input feature set of the water quality prediction model are used as the input of the water quality prediction model, and the output of the water quality prediction model is the indicator Xi to be predicted. During prediction, if there is a measured value of the input feature at the current time, the measured value is used as the input of the input feature; otherwise, the predicted value is used as the input of the input feature.

[0013] The data preprocessing described in this invention involves, on the one hand, filling in missing data using linear interpolation; and on the other hand, normalizing the dataset using a normalization method so that all water quality indicators are between 0 and 1.

[0014] The correlation is expressed as the absolute value of the correlation coefficient calculated using the Spearman coefficient. Water quality indicators with an absolute value of correlation coefficient greater than a set value are selected to form a set S of related water quality indicators.

[0015] Furthermore, step 3) involves selecting a subset of water quality indicators from the associated water quality indicator set S to form the input feature set Si of the water quality prediction model, as follows:

[0016] 3.1) First, construct a water quality prediction model, and then initialize the relevant parameters of the water quality prediction model;

[0017] 3.2) Based on the required number of input indicators for the water quality prediction model, randomly select multiple water quality indicators from the associated water quality indicator set as input features for the water quality prediction model to obtain multiple training sets composed of different combinations of input features;

[0018] 3.3) Based on the multiple training sets obtained in 3.2), input them into the water quality prediction model with the same hyperparameters for training. Test the training results using the test set and use RMSE to evaluate the error between the true value and the predicted value. Compare the prediction errors of different combinations of input features in the water quality prediction model with the same model parameters. The combination of input features with the smallest RMSE constitutes the input feature set Si of the water quality prediction model.

[0019] In step 3.1), the water quality prediction model is constructed based on the LSTM neural network model. The initialization of model-related parameters includes setting the time step d, batch size, number of training epochs, and hidden layer dimension hidden_dim.

[0020] Step 5) When training the water quality prediction model,

[0021] a. If T=0, all Xi in the input of the water quality prediction model are measured values, and the water quality prediction model does not need to be trained;

[0022] b. If T≠0, the water quality prediction model needs to be trained. During training, the input features and model parameters remain unchanged. At time t+n*T, the measured values ​​of the index Xi in the training set are input, and the predicted values ​​are input at other times. After the model converges, the final water quality prediction model is obtained.

[0023] When T is infinite, the input value at each moment is the predicted value of the water quality prediction model, and there is no need for manual acquisition of the actual value of the indicator to be predicted; the time interval T can be extended or shortened as needed.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] This invention uses LSTM to predict water quality data over a relatively long period, which can reduce the number of water quality tests, reduce the workload of water quality testing, and save manpower and time costs. Compared with other typical water quality prediction methods, the prediction model can handle the sparse labeling problem and can combine predicted values ​​with measured values ​​to obtain water quality data at future time points. It can be applied to fields such as closed-loop control of wastewater treatment and has good application value.

[0026] Traditional water quality testing methods rely on sampling and analysis, which are sparse in the time domain and unsuitable for real-time control in wastewater treatment. This invention addresses the shortcomings of traditional water quality testing methods in wastewater treatment, such as low timeliness and high testing costs. It utilizes sparse labels generated from water quality testing results to establish an LSTM-based prediction model to obtain real-time water quality parameters. This solves the timeliness problem in the sampling-then-analysis process and can be used for real-time closed-loop control in wastewater treatment, improving control accuracy and saving labor and energy costs. Attached Figure Description

[0027] Figure 1 -Flowchart for predicting water quality parameters in this invention.

[0028] Figure 2 - A schematic diagram of the present invention where measured values ​​are input once every time interval T.

[0029] Figure 3 -Graph showing the total nitrogen test results at different time intervals in this embodiment of the invention. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] The basic idea of ​​this invention for water quality prediction is as follows: Preprocess historical monitoring data of real-time water quality indicators obtained from sensors and historical data of water quality indicators obtained through laboratory testing. Calculate the correlation between each water quality indicator and the indicator to be predicted (i.e., the indicator to be obtained through laboratory testing) based on the correlation between the indicators. Select a set of water quality indicators with a high correlation to the indicator to be predicted and construct a dataset, dividing it into training and testing sets in a certain ratio (usually 8:2). Establish an LSTM water quality prediction model and initialize relevant parameters, perform feature selection and model parameter optimization, and find the input features and model parameters with the best prediction effect. Then train the model, inputting the measured value of the water quality indicator every time interval T, along with other real-time collected data, to obtain the predicted value of the water quality indicator at the next time moment. The final prediction model is obtained through training. Based on the final prediction model, input relevant indicator data from the existing data into the prediction model to output the water quality indicator to be predicted, achieving the required prediction. This reduces the detection frequency and workload of the indicator to be predicted, and may even eliminate the need for actual detection.

[0032] The specific steps of this invention are as follows, and see also... Figure 1 :

[0033] (1) Check whether there are missing data in the water quality index monitoring data obtained by the sensor. If the data is missing, use linear interpolation to fill the missing data, that is, use the average value of the data before and after the missing value to fill the missing value. If there is no missing data or the missing data has been filled, use the maximum and minimum value normalization method to normalize the dataset so that the water quality index is between 0 and 1, which is convenient for subsequent application.

[0034] (2) After the data preprocessing in step (1), the correlation between the index Xi to be predicted and all the indices is calculated using the Spearman coefficient. According to the absolute value of the correlation coefficient, the correlation is classified as follows: 0.8-1.0 is extremely strong; 0.6-0.8 is strong; 0.4-0.6 is moderate; 0.2-0.4 is weak; and 0-0.2 is extremely weak or no correlation. Water quality indices with an absolute value of correlation coefficient greater than 0.4 are selected to form a water quality index set S={Xi(1),Xi(2),…Xi(s)}, where Xi(s) represents the s-th water quality index associated with Xi.

[0035] (3) In order to obtain the best prediction results, it is necessary to select some water quality indicators from S to form the input feature set Si. The specific steps include:

[0036] 3.1) Construct a water quality prediction model based on the LSTM neural network model, initialize the relevant model parameters, and set the values ​​of time step d, batch size, number of training epochs, and hidden layer dimension hidden_dim;

[0037] 3.2) Feature selection is performed on the indicators in S. Based on the number of indicators required for the input water quality prediction model, multiple input features are randomly selected as input features of the water quality prediction model. N training sets are formed by different combinations of input features.

[0038] 3.3) Based on the N training sets obtained in 3.2), input them into the water quality prediction model with the same hyperparameters for training. Test the training results using the test set and use RMSE (root mean square error) to evaluate the error between the true value and the predicted value. Compare the prediction errors of different combinations of input features in the water quality prediction model with the same model parameters, and find the set of results with the smallest RMSE. This set of input features constitutes the input feature set Si.

[0039] (4) Based on the input feature set Si obtained in step (3), change the d, batch_size, epoch and hidden_dim of the LSTM neural network to establish water quality prediction models with different hyperparameters. Then, use the input feature set Si to train these models with different parameters respectively, test them with the test set, compare the prediction error of the input feature set Si in the water quality prediction models with different parameters, find the group with the smallest RMSE, and the model parameters corresponding to this group are the hyperparameters of the water quality prediction model with the best prediction effect.

[0040] (5) Based on the above steps, the hyperparameters and input characteristics of the water quality prediction model are obtained. Then, according to the actual engineering needs, the time interval T of the measured values ​​of the input index is set, that is, the measured value of the input index is at time t+n*T (n=0,1,2,3...), and the predicted value is used as the input for the next time at other times. After setting T, the model needs to be trained according to different T values, specifically in the following cases:

[0041] a. If T=0, all Xi values ​​in the input of the water quality prediction model are measured values, and the model does not need to be retrained;

[0042] b. If T≠0, the model needs to be retrained. During retraining, the input features and model parameters remain unchanged, but the measured values ​​of the index Xi in the training set are input at time t+n*T (n=0,1,2,3...), and the predicted values ​​are input at other times. After the model converges, the final water quality prediction model is obtained.

[0043] (6) In actual prediction, the water quality indicators in the existing water quality indicator monitoring data that correspond to the water quality prediction model input feature set are used as the input of the water quality prediction model, and the output of the water quality prediction model is the indicator Xi to be predicted. When predicting, if there is a measured value of the input feature at the current time, the measured value is used as the input of the input feature; otherwise, the predicted value is used as the input of the input feature.

[0044] In actual prediction, the time interval T for manual testing can be extended or shortened according to engineering needs. When T is infinite, the input value at each moment is the predicted value of the model, that is, there is no need to manually test the true value of the indicator to be predicted; thus achieving the goal of greatly saving water quality testing costs and reducing the number of tests.

[0045] This invention relates to a water quality parameter prediction method based on a sparsely labeled Long Short-Term Memory (LSTM) network. For water quality parameters that require sampling and testing but lack real-time accuracy, a neural network model is constructed for prediction to meet real-time application needs. Since these water quality parameters are obtained through testing, typically every few hours or once a day, the sample labels are sparse. This invention addresses this problem by establishing an LSTM network to predict water quality parameters. The input includes other real-time parameters obtainable by sensors and the water quality parameter from the previous time step. The output is the predicted value of the water quality parameter for the next time step. During model training, at time points where actual test results (measured values) are available for the water quality parameter to be predicted, the input information uses the measured values; at other time points, the predicted values ​​are used. The loss function is calculated based on the error between the predicted and measured values ​​at the time points of the measured parameters. During actual prediction, the input values ​​are consistent with those used in the model training phase. If a measured value exists, it is used to replace the predicted value at the corresponding time point to correct the prediction result and improve accuracy.

[0046] The present invention will be further described below with reference to a non-limiting embodiment.

[0047] Please refer to Tables 1, 2 and 3 for the relevant data involved in this embodiment.

[0048] Table 1 Correlation Analysis (Absolute Values)

[0049]

[0050] Table 2 Prediction errors for typical input features

[0051]

[0052] Table 3. Prediction errors of different model parameters for total nitrogen prediction

[0053]

[0054] This invention is a method for predicting water quality indicators based on LSTM neural networks. It comprehensively utilizes methods such as average value filling, data normalization, correlation calculation, input feature selection, and hyperparameter setting to establish a prediction model based on historical monitoring data of water quality indicators, enabling accurate prediction of the indicator Xi to be predicted. Figure 1 , Figure 2 The prediction flowchart shown, the correlation analysis shown in Table 1, the prediction errors shown in Tables 2 and 3, and Figure 3 The prediction results shown illustrate the present invention in detail, and the specific steps are as follows:

[0055] (1) To better illustrate the predictive effect of the present invention, this embodiment predicts the total nitrogen water quality parameter, denoted as Xi. Historical monitoring data of water quality indicators are released by the China National Environmental Monitoring Centre. The released indicators include nine monitoring indicators: water temperature, pH, dissolved oxygen, conductivity, turbidity, potassium permanganate, ammonia nitrogen, total nitrogen, and total phosphorus. Historical monitoring data of the Yingugong section of the Yellow River Basin in Yinchuan City, Ningxia Hui Autonomous Region, are selected, totaling 1857 data points. Since water quality indicators cannot be monitored during the maintenance of the monitoring station, there are data gaps. Linear interpolation is used to fill the missing data, and then normalization is performed using formula (1), where The value of a certain water quality indicator at a certain moment. , These represent the maximum and minimum values ​​of the indicator, respectively. This represents the result after normalization;

[0056] (1)

[0057] (2) The Spearman coefficient was used to calculate the correlation between total nitrogen and other indicators (the correlation between total nitrogen and itself is always 1, so no calculation is needed). The calculation formula is shown in (2), where n represents the sample size. and Let X and Y represent the magnitudes of water quality index X and water quality index Y at time i, respectively.

[0058] (2)

[0059] The correlation analysis results are shown in Table 1. Based on the absolute value of the correlation coefficient, 0.8-1.0 indicates extremely strong correlation, 0.6-0.8 indicates strong correlation, 0.4-0.6 indicates moderate correlation, 0.2-0.4 indicates weak correlation, and 0-0.2 indicates extremely weak correlation or no correlation. The set of water quality indicators with an absolute correlation coefficient greater than 0.4 with total nitrogen is also included. ={Total nitrogen, water temperature, dissolved oxygen, conductivity, ammonia nitrogen, pH};

[0060] (3) Input feature selection

[0061] a. Initialize the prediction model parameters, setting the time step d to 2, the batch size to 70, the number of training epochs to 200, and the hidden layer dimension hidden_dim to 256;

[0062] b. From The input features for predicting total nitrogen are selected, and typical input features and training sets are constructed based on the types, number, and relevance of the indicators included in the input features.

[0063] c. Train the initial model using different training sets, predict the training results using the test set, and evaluate the prediction error using RMSE. The RMSE calculation is shown in Equation (3). The prediction errors for different input features are shown in Table 2. When the input features are dissolved oxygen, ammonia nitrogen, pH, water temperature, and total nitrogen, the prediction effect of total nitrogen is the best. This feature set is denoted as Si.

[0064] (3)

[0065] (4) Change the d, batch_size, epoch and hidden_dim of the LSTM neural network to establish water quality prediction models with different parameters. Input the predicted total nitrogen (Xi) input features (Si) obtained in step (3) into multiple water quality network models with different model parameters for training and prediction. The prediction errors of different model parameters are shown in Table 3. It can be seen from Table 3 that the prediction error of total nitrogen is the smallest when d is 3, batch_size is 70, epoch is 400 and hidden_dim is 256. Therefore, the hyperparameters of the water quality prediction model are determined to be d is 3, batch_size is 70, epoch is 400 and hidden_dim is 256.

[0066] (5) Based on the above steps, the hyperparameters and input features are determined to form a water quality prediction model. The value of T is adjusted as needed according to the actual situation, and then the water quality prediction model is retrained. During retraining, the input features and model parameters remain unchanged, but the measured values ​​of the indicators in the training set are input at time t+n*T (n=0,1,2,3…), and the predicted values ​​are input at other times, such as… Figure 2 As shown; train the model until it converges to obtain the final water quality prediction model;

[0067] (6) The final water quality prediction model, after testing on the test set, can predict water quality indicators relatively accurately. In actual prediction, the model input is discretized, and the prediction results for each input of the measured value Xi at time intervals T=0-5 are shown below. Figure 3 In steps (a)-(f), when T=0, the measured value of the previous Xi is input for each prediction of Xi; when T=1, the measured value of total nitrogen is input once every two predictions, and the predicted value of the previous prediction is input at other times; when T=2, the measured value of total nitrogen is input once every three predictions, and the predicted value is input at other times; when T=3, 4, 5... and so on, thus obtaining the predicted value of total nitrogen at different times, including the predicted values ​​at other times between existing measured values, as well as the predicted values ​​at different times within a future time period. In practical applications, the time interval T of manual testing can be extended or shortened according to engineering needs, thereby greatly saving water quality testing costs and reducing the number of tests.

[0068] The above embodiments of the present invention are merely illustrative examples and are not intended to limit the implementation of the invention. Those skilled in the art can make other variations and modifications based on the above description. It is impossible to exhaustively list all possible implementations here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A method for water quality parameter prediction based on a sparse label long short-term memory network, characterized in that: The steps are as follows, 1) data preprocessing; the historical monitoring data of several water quality indicators obtained in advance are preprocessed; 2) determine the associated water quality indicators; calculate the correlation degree of the to-be-predicted indicator Xi and the indicators involved in the preprocessed data, select the top s water quality indicators in the correlation degree, and constitute the associated water quality indicator set S={Xi(1), Xi(2), … Xi(s)}, wherein Xi(s) represents the s-th water quality indicator associated with Xi; the to-be-predicted indicator Xi is one of the indicators involved in the preprocessed data; the correlation degree is expressed by the absolute value of the correlation coefficient calculated by the Spearman coefficient, and the water quality indicators with an absolute value of the correlation coefficient greater than a set value are selected to constitute the associated water quality indicator set S; 3) select part of the water quality indicators from the associated water quality indicator set S to constitute the water quality prediction model input feature set Si; the to-be-predicted indicator Xi at the next moment is the output of the water quality prediction model; the water quality prediction model input feature set Si contains the to-be-predicted indicator Xi at the current moment; the water quality prediction model is constructed according to the LSTM neural network model; 4) according to the water quality prediction model input feature set Si obtained in step 3), change the hyperparameters of the water quality prediction model, establish water quality prediction models with different hyperparameters, then use the water quality prediction model input feature set Si to train these water quality prediction models with different hyperparameters, test them with the corresponding test set, compare the prediction errors of the water quality prediction models with different hyperparameters, and the model hyperparameters corresponding to the minimum RMSE are the hyperparameters corresponding to the water quality prediction model; 5) according to the water quality prediction model obtained in step 4) and the water quality prediction model input feature obtained in step 3), train the water quality prediction model using the data corresponding to the water quality prediction model input feature set Si to obtain the final water quality prediction model; during training, first set the time interval T of the measured value of the input feature corresponding to the to-be-predicted indicator; for the input feature corresponding to the to-be-predicted indicator, input the measured value of the indicator at t+n*T, wherein n is a natural number, and the predicted value is used as the input at the next moment; For other input features, input the measured value at any time; 6) during actual prediction, input the water quality indicators corresponding to the water quality prediction model input feature set in the existing water quality indicator monitoring data into the water quality prediction model, and the output of the water quality prediction model is the to-be-predicted indicator Xi; during prediction, if there is a measured value for the input feature at the current moment, use the measured value as the input of the input feature, otherwise use the predicted value as the input of the input feature.

2. The sparse label-based long short-term memory network water quality parameter prediction method according to claim 1, characterized in that: The data preprocessing, on the one hand, fills the missing data by linear interpolation method; on the other hand, normalizes the data set by normalization method, so that the water quality indicators are between 0 and 1. 3.The sparse label based long short term memory network water quality parameter prediction method of claim 1, wherein: The step 3) of selecting part of the water quality indicators from the associated water quality indicator set S to constitute the water quality prediction model input feature set Si is as follows, 3.1) first construct a water quality prediction model, and then initialize the parameters of the water quality prediction model; 3.2) According to the number of water quality prediction model index input requirements, randomly select multiple water quality indicators in the associated water quality indicator set as the input features of the water quality prediction model, obtain multiple training sets composed of different input feature combinations; 3.3) According to the multiple training sets obtained in 3.2), input into the water quality prediction model with the same hyperparameters for training, test the training results with the test set, and use RMSE to evaluate the error between the true value and the predicted value. Compare the prediction errors of different input feature combinations in the water quality prediction model with the same model parameters, and the input feature combination with the smallest RMSE constitutes the water quality prediction model input feature set Si.

4. The sparse label-based long short-term memory network water quality parameter prediction method according to claim 3, characterized in that: In step 3.1), initializing model-related parameters includes setting the time step d, the batch size batch_size, the training round epoch, and the hidden layer dimension hidden_dim. 5.The sparse label based long short term memory network water quality parameter prediction method of claim 1, wherein: Step 5) When training the water quality prediction model, a. If T=0, all Xi in the input of the water quality prediction model are measured values, and the water quality prediction model does not need to be trained; b. If T≠0, the water quality prediction model needs to be trained; when training, the input features and model parameters remain unchanged, the measured value of the index Xi in the training set is input at t+n*T time, and the predicted value is input at the remaining time. After the model converges, the final water quality prediction model is obtained; When T is infinite, the input value at each time is the predicted value of the water quality prediction model, so there is no need to manually obtain the true value of the predicted index; the time interval T is extended or shortened according to the demand.

Images