KR20090078501A - Apparatus and method for prediction of influent flow rate and influent components using artificial neural network - Google Patents

Abstract

A prediction method of inflow flux and inflow component concentration of a sewage disposal plant by an artificial neural network and a device thereof are provided to predict inflow flux by using inflow data obtained from the sewage disposal plant and weather data only, thereby automatically coping with sewage. A data collecting unit(10) collects inflow water quality data of a sewage disposal plant and weather data. A data preprocessor(20) receives collected data sets, and converts the data sets into a standardized state, so that the converted sets are applied to an artificial neural network model. An artificial neural network unit(30) repeatedly performs signal transmission through a concealment layer, and controls a weighted value by conducting a learning process for minimizing an error. A data predicting unit inputs a variable according to the weighted value to predict inflow water.

Description

Apparatus And Method For Prediction of Influent Flow Rate and Influent Components Using Artificial Neural Network}

The present invention is a method for predicting the inflow flow rate and inflow component concentration of the sewage treatment plant by artificial neural network for measuring the inflow flow rate and the inflow component and predicting the future inflow drifting concentration for the effective process diagnosis and control of the sewage treatment plant. And to an apparatus.

As environmental regulations are gradually tightened, instrumentation, control and automation (ICA) technologies for the reliable and economical operation of sewage treatment plants have been rapidly delivered. With the development of this ICA technology, research on the model of sewage treatment process, which is a means for effective process diagnosis and control, has been actively conducted.

This model is a deterministic model that is basically based on the process-related mechanisms and mass balance equations. The black-box model generated by considering only the change pattern, and the gray-box model in which the white-box model and the black-box model are combined in series or in parallel are developed and developed by the IWA task group. Activated sludge model (ASM) and artificial neural network (ANN) models are representative examples of white-box and black-box models, respectively.

These models include process design, which means evaluation by simulation of various design plans for the construction of sewage treatment plants, process optimization, and improvement and control of processes that suggest appropriate alternatives in the event of process failure. It is useful in terms of use.

In terms of process optimization and control, the use of the model is limited to the application based on the present point of time, that is, to normalize after the process abnormality occurs. It is not used as a tool to prevent problems that can occur by detecting them in advance.

In order to overcome the limitations of the current existing model, it is necessary to develop an appropriate influent prediction model, and if the developed influent prediction model is combined with the existing process performance prediction model, information on the future state of the process will be provided in advance. It is anticipated that probable problems can be detected in advance and actions taken before process failures can contribute to the stable operation of the process.

Although some researchers have attempted to develop influent forecasting models, most of the models developed by them are deterministic models that have the disadvantage of being difficult to calibrate with many variables that change over time.

Of course, the results of using the black-box model, which is another type of model, have been reported, but this does not reflect changes in other influent components, which are predicted to be limited only to the influent flow rate, which may be a factor that causes the process to fail. Has a threshold.

Therefore, based on the recognition of the necessity of the influent prediction model, the present invention develops a predictive model that can overcome the limitations of the deterministic model while predicting the influent characteristics as well as the actual influent flow rate of the sewage treatment plant. The purpose is to provide a method and apparatus for predicting the inflow flow rate and inflow component concentration of a sewage treatment plant by an artificial neural network using an artificial neural network, which is a black-box model.

In order to solve the above problems, the present invention provides a device for predicting inflow flow rate and inflow component concentration of a sewage treatment plant, comprising: a data collecting unit (10) for collecting inflow water quality data and meteorological data of the sewage treatment plant; A data preprocessor 20 which receives the data sets collected by the data collector 10 and converts the data sets into a standardized state to be applied to an artificial neural network model; An artificial neural network unit 30 for performing standardized input values through the data preprocessing unit 20 and repeatedly performing signal transmission through the hidden layer to perform learning to minimize errors, thereby adjusting weights; A data prediction unit 40 for predicting the inflow of water by inputting a variable according to a weight established by the artificial neural network unit 30; It is achieved by the apparatus for predicting the inflow flow rate and inflow component concentration of the sewage treatment plant by the artificial neural network, characterized in that the display section 50 for outputting the calculated data of the data prediction unit 40 on the screen.

In addition, the present invention, the inflow water quality data of the sewage treatment plant is any one or more of flow rate, pH, DO, temperature, BOD, COD _Mn , SS, TN, TP, the weather data is preferably at least one of rainfall, humidity and temperature. Do

In addition, according to the present invention, the data collection unit 10 is constructing a database 11 to store the collected data, the pre-processed data sets during the training of the neural network or verifying the predicted performance of the trained neural network It is preferable to be used for either

According to the present invention, there is provided a method of predicting the inflow flow rate and inflow component concentration of a sewage treatment plant, comprising: a data collection step (S10) of receiving inflow water quality data and meteorological data from the sewage treatment plant at the data collection unit 10; A data preprocessing step (S20) for standardizing the collected data to apply to the artificial neural network model; Establishing a prediction model by performing learning through an artificial neural network on the data preprocessed variables (S30); Preferably, the inflow flow rate and inflow component concentration prediction step (S40) for evaluating the predictive performance of the trained neural network using the verification data set for the artificial neural network is preferable.

In addition, according to the present invention, the data preprocessing step (S20), the step of removing the outlier (Outlier) that can reduce the reliability of the data of the data secured from the sewage treatment plant; Reducing or eliminating short term variation in the long term data set to determine linear or nonlinear correlation between the measured time series data and to identify trends over time of successive data; It is preferable that the data is composed of the steps of normalizing so that all data distributed on different scales have a distribution of values between 0 and 1.

In addition, according to the present invention, it is preferable to use a control chart technique to remove the outliers of the data.

In addition, according to the present invention, it is preferable to use the 5-day moving average technique to determine the linear or nonlinear correlation between the measured time series data and the trend over time of consecutive data.

In addition, according to the present invention, in the prediction model establishment step (S30), the weight value to determine the neural network structure and epoch number through trial and error method, and to generalize the trained neural network based on the structure and epoch number of the selected neural network It is desirable to determine them.

Therefore, by the method and system for predicting the inflow flow rate and inflow component concentration of the sewage treatment plant by the artificial neural network of the present invention, using only the inflow data and meteorological data obtained from the sewage treatment plant, one day after and two days later, Three days later, the inflow flow rate, COD _Mn , SS, TN, etc. are predicted, and accordingly, appropriate response of a sewage treatment plant can be performed automatically.

Hereinafter, a method and system for predicting the inflow flow rate and the inflow component concentration of the sewage treatment plant by the artificial neural network of the present invention will be described in more detail with reference to the accompanying drawings.

1 is a block diagram of an apparatus for predicting inflow flow rate and inflow component concentration of a sewage treatment plant by an artificial neural network of the present invention.

As shown in the figure, the present invention includes a data collection unit 10, a data preprocessor 20, an artificial neural network unit 30, a data prediction unit 40, and a display unit 50.

The data collection unit 10 is an inflow water quality data (flow rate, pH, DO, temperature, BOD, COD _Mn , SS, TN, TP) and meteorological data (rainfall) in the sewage treatment plant installed and operated as a standard activated sludge process , Humidity, temperature, etc.).

The data collector 10 constructs a database 11 to store the collected data, and the 420 pre-processed data sets are used for developing a predictive model.

Of these, 210 data sets were used for the training of neural networks, and the remaining 210 data sets were used to verify the predictive performance of the trained neural networks.

The data sets collected by the data collector 10 are transferred to the data preprocessor 20 to become a standardized state for developing an artificial neural network.

To this end, first, by using control chart techniques to remove outliers that may reduce the reliability of data among the data obtained from the sewage treatment plant, values outside the management upper limit and the management descending line are considered as outliers. .

Secondly, to reduce the short-term fluctuations in the long-term data sets obtained by using the 5-day moving average technique to determine the linear or nonlinear correlation between measured time series data and the trend over time of consecutive data; Removal is performed.

Finally, for data standardization, which is the basic stage of data preparation of neural networks, all data () distributed on different scales () are standardized so that a value distribution between 0 and 1 can be obtained.

The input values standardized by the data preprocessor 20 allow a prediction model to be established through the artificial neural network unit 30.

The present invention uses a multi-layer feed-forward back-propagation network, so that neurons are arranged in each layer, namely input layer, hidden layer, output layer, and each layer. Neurons are connected by a connection weight.

That is, normalized input data is input to neurons of the input layer and passed through the hidden layer to the output layer.

The artificial neural network unit 30 determines a proper neural network structure and epoch number by a trial and error method, and weights enough to generalize the trained neural network based on the selected neural network structure and epoch number. Determine the values.

The neural network unit 30 performed all simulation tasks using neural network toolbox (The MathWorks Inc., 2000) provided by MATLAB, which is commercial software.

As the data predictor 40 establishes a prediction model through the artificial neural network unit 30, the data predictor 40 predicts the inflow flow rate and the inflow component concentration through the artificial neural network model.

The quantitative evaluation of the influent prediction performance of the target sewage treatment plant using neural network can be evaluated by the error between the predicted value and the measured value derived through the simulation, and the average relative error rate for the quantitative prediction performance of the neural network model developed in the present invention. (Average relative difference).

Figure 2 shows the overall flow chart of the prediction method of the inflow flow rate and influent component concentration of the sewage treatment plant by the artificial neural network of the present invention.

First, a data collection step (S10) of inputting data from an actual sewage treatment plant into the data collection unit 10 is performed.

For this purpose, B city A sewage treatment plant, which is installed and operated as a standard activated sludge process, was selected as the sewage treatment plant to be developed for the inflow forecasting model using artificial neural network.

Inflow water quality data (flow, pH, DO, temperature, BOD, COD _Mn , SS, TN, TP) and weather data (rainfall, humidity and temperature, etc.) measured from the actual sewage treatment plant from 2005 to 2006 It will be collected through the data collector 10.

The data set collected through the data collector 10 is subjected to data preprocessing, and 420 data sets which have been subjected to data preprocessing are used for the development of a prediction model. Of these, 210 data sets were used for the training of neural networks, and the remaining 210 data sets were used to verify the predictive performance of the trained neural networks.

Table 1 below shows the statistical characteristics of the influent sewage treatment plant.

Where Hum. : Humidity, Temp. : Temperature, Max. : Maximum value, Min. : Minimum value, S.D .: Standard Deviation

When the data collection step is performed, a data preprocessing step (S20) for standardizing the collected data is performed.

To this end, outliers, which may reduce the reliability of data among the data obtained from the sewage treatment plant, may be removed because they may hinder the development of accurate prediction models.

In the present invention, the control chart technique is used to remove the outliers.

As in Equation (1), upper control limit (UCL) and lower control limit (LCL) were set, and values outside the UCL and LCL were considered outliers and removed.

In this study, the constant was used as 3 with reference to the value proposed by Mjalli et al. (2007).

Second, various statistical processes are generally performed to identify linear or nonlinear correlations between measured time series data and trends of consecutive data.

In the present invention, one of a variety of statistical techniques used a 5-day moving average technique, which can be expressed as in Equation 2, and the short-term fluctuations in the long-term data sets obtained through this reduction and elimination was made. Where is the measurement variables and is the current time period.

The correlation between these preprocessed data was qualitatively evaluated, and the variables that had high correlation with inflows, COD _Mn , SS, TN, etc. were used as input variables of neural network.

 Finally, Eq. (3) is standardized so that all data () distributed on different scales have a distribution of values between 0 and 1 for data normalization, which is the basic stage of data preparation of neural networks. .

To develop each artificial neural network model for predicting inflow flow rate, COD _Mn , SS, and TN concentrations after 1 day from the current point of time, The appropriate input variables for the variable were selected.

5 shows an example of such a qualitative correlation evaluation of precipitation and inflow flow rate (a), the relationship between humidity and inflow flow rate (b), and the relationship between COD _Mn and other variables to be predicted (c, d). It is shown through a graph.

5, it can be seen that an increase in precipitation and humidity results in an increase in inflow flow rate and a decrease thereof results in a decrease in inflow flow rate. This means that the behavior of change in inflow, precipitation and humidity is comparable.

Through qualitative correlation evaluation through these graphs, humidity and precipitation were selected as input variables for the prediction of inflow flow, and inflow flow was selected one day before considering the time series characteristics.

In addition, as shown in FIG. 5, it can be seen that the change behaviors of other variables to be predicted, that is, the target variables and the inflow COD _Mn concentration, are also similar. Inflow COD _Mn rather than variables The concentration was chosen.

After performing the data preprocessing, a step of establishing a prediction model through an artificial neural network (S30) is performed.

The artificial neural network simplifies the biological neurons present in the human brain and their connection and mathematically models them to implement an intelligent form represented by the human brain.

Currently, the neural network is widely used in fields such as system identification, prediction, pattern recognition, classification, and process control. The neural network has a single layer feed-forward network, multi-layer feed-forward network, and feed according to its structure. It is classified into back-back network, recurrent network, and self-organized network, and the neural network structure of appropriate type is selected according to the purpose of neural network use.

In order to predict the influent, the present invention uses a multi-layer feed-forward back-propagation network which is most commonly used in modeling and prediction fields.

3 shows a structure of a general multi-layer feed-forward back-propagation network.

As shown in FIG. 3, neurons are arranged in respective layers, that is, an input layer, a hidden layer, and an output layer, and neurons of each layer are connected by connection weights.

Normalized input data is input to the neurons of the input layer and passed through the hidden layer to the output layer. The signal transfer from the neuron of the previous layer to the neuron of the next layer can be expressed as shown in FIG. 4, and in this study, the tanh-sigmoid function represented by Equation 4 is used as a transfer function for signal transmission. It was.

Learning in a neural network means minimizing the error between the measured value and the output value calculated through the neural network by adjusting the weights connected between the neurons.

The present invention overcomes the shortcomings of the back propagation algorithm instead of the back-propagation algorithm, which has the disadvantage of not being able to perform further learning when it falls into the local minimum in the error space. Levenberg-Marquart (LM) algorithm is used as a learning algorithm.

The main factors that determine the performance of a neural network are the structure of the neural network, the end point of learning of the neural network (ie, the number of epochs), and the weights given.

There is no absolute way to determine the number and weight of epochs for the structure and learning of neural networks.

Therefore, in the present invention, a proper neural network structure and epoch number are determined by a trial and error method, and weight values are sufficiently determined to generalize a trained neural network based on the selected neural network structure and epoch number. It was.

In the present invention, all simulation work was performed using neural network toolbox (The MathWorks Inc., 2000) provided by MATLAB, which is a commercial software.

FIG. 6 shows the structure of a neural network determined and determined by a trial and error method to predict an inflow flow rate after one day from the current time point.

As can be seen from Figure 6, the artificial neural network for the prediction of the inflow flow is composed of one input layer, two hidden layers, one output layer, four neurons in the input layer, each layer in two hidden layers Five neurons and one neuron were arranged in the output layer.

Table 2 below shows information about the structure of a neural network configured to predict COD, SS, and TN concentrations after 1 day from the current time point.

TABLE 2

Where, t: present, t-1: 1 day before the present, t + 1: 1 day after the present

The neural network having the structure as described in FIG. 6 and <Table 2> was fixed to the number of epochs to 100, and then the neural network was trained using the training data set by changing the weight, and trained using the verification data set. The predictive performance of the neural network was evaluated.

In the present invention, the reason why the number of epochs is 100 is that when the number of epochs is 100 or more, the problem of overfitting occurs due to excessive training regardless of the weight change.

When the step (S30) of establishing a prediction model through the artificial neural network is performed, the prediction step (S40) of the inflow flow rate and the inflow component concentration is performed through the artificial neural network model.

FIG. 7 shows the results of training a neural network using 210 data sets for prediction one day after the current time point of each target variable, and the results predicted by a neural network trained using the remaining 210 validation data sets. Shows.

Here, each graph is a figure regarding the prediction result of (a) outflow volume, (b) COD _Mn , (c) SS, and (d) TN.

The quantitative evaluation of the influent prediction performance of the target sewage treatment plant using neural network can be evaluated by the error between the predicted value and the measured value.

Therefore, for evaluating the quantitative prediction performance of the neural network model developed in the present invention, the average relative error represented by Equation (5) below was used. Where N is the total number of variables compared, i is the i th measurement value, and i is the i th prediction value. The average state error rate for each predictor is shown in Table 3.

TABLE 3

7 and Table 3, it was confirmed that the training of the artificial neural network was properly performed to predict each influx component one day after the current time point, and the artificial neural network trained using the data not used in the training As a result of verifying the performance of, the trained artificial neural network describes the behavior of the change of the measured values of each target variable well, and the relative error rate is 0.65% for flow rate, 1.69% for COD _Mn and 3.16% for SS. In the case of TN, it can be seen that the value is less than 5% which is included in the general 95% confidence interval as 1.96%.

When evaluating through this verification result, artificial neural network model developed to predict inflow flow rate and concentration of inflow components after 1 day from the present point of time can be generalized and used in the target sewage treatment plant.

As a result of the prediction after the current day 1, 2 days, 3 days after the current point of view, for the prediction of each inflow flow rate, COD, SS, TN after 2 days, 3 days after the current time point After fixing the epoch number to 100, the artificial neural network with the input variables and structure such as the artificial neural network used for prediction 1 day after the current point of view was made only by changing the weight value.

8 and 9 are prediction graphs for (a) inflow rate (b) COD _Mn , (c) SS, and (d) TN, respectively.

Tables 4 and 5 show the relative error rates used as quantitative indicators for evaluating the predictive performance of the developed artificial neural networks.

TABLE 4

TABLE 5

Through this, the artificial neural network model developed for prediction 2 days and 3 days after the present time point can predict the change behavior of each target variable properly without causing the problem of overfitting. .

Therefore, it can be said that an artificial neural network model for prediction after 2 days and 3 days after the current viewpoint, which has the same structure as the artificial neural network for prediction 1 day after the current point of view and differs only in the connection weight value, is appropriately developed. have.

1 is an overall configuration diagram of an apparatus for predicting inflow flow rate and inflow component concentration of a sewage treatment plant by an artificial neural network of the present invention;

2 is an overall flowchart of a method for predicting the inflow flow rate and the inflow component concentration of the sewage treatment plant by the artificial neural network of the present invention;

3 is a structure of an artificial neural network having a general multiple layer,

4 is a signaling diagram for each layer of the multiple artificial neural network of FIG.

5 is a graph showing the relationship between precipitation and inflow flow rate, humidity and inflow flow rate, other predicted variables and COD _Mn ,

6 is a structure of a neural network determined and determined by trial and error method in order to predict an inflow flow rate 1 day after the current time point,

7 is a graph comparing the results of the predictions of the target variables one day after the current time point;

8 and 9 are graphs comparing the results of the prediction after 2 days and 3 days after the current time point.

* Detailed Description of Drawings *

10: data collector 11: database

20: data preprocessing unit 30: artificial neural network

40: data prediction section 50: display section

Claims (8)

In the prediction apparatus of inflow flow volume and inflow component concentration of a sewage treatment plant, A data collector 10 for collecting inflow water quality data and weather data of the sewage treatment plant; A data preprocessor 20 which receives the data sets collected by the data collector 10 and converts the data sets into a standardized state to be applied to an artificial neural network model; An artificial neural network unit 30 for performing standardized input values through the data preprocessing unit 20 and repeatedly performing signal transmission through the hidden layer to perform learning to minimize errors, thereby adjusting weights; A data prediction unit 40 for predicting the inflow of water by inputting a variable according to a weight established by the artificial neural network unit 30; Predicting device of the inflow flow rate and inflow component concentration of the sewage treatment plant by the artificial neural network, characterized in that it comprises a display unit (50) for outputting the calculated data of the data prediction unit (40) on the screen. The method of claim 1, The inflow water quality data of the sewage treatment plant is at least one of flow rate, pH, DO, temperature, BOD, COD _Mn , SS, TN, TP, and the weather data is at least one of rainfall, humidity, and temperature. Prediction device of inflow flow rate and inflow component concentration of sewage treatment plant by The method of claim 1, The data collection unit 10 is building a database 11 to store the collected data, so that the pre-processed data sets are used for either the training of the neural network or the verification of the predicted performance of the trained neural network. A device for predicting the inflow flow rate and inflow component concentration of a sewage treatment plant using an artificial neural network. In the prediction method of the inflow flow volume and inflow component concentration of a sewage treatment plant, A data collection step (S10) of receiving inflow water quality data and meteorological data from the sewage treatment plant at the data collection unit 10; A data preprocessing step (S20) for standardizing the collected data to apply to the artificial neural network model; Establishing a prediction model by performing learning through an artificial neural network on the data preprocessed variables (S30); Inflow flow rate and inflow component of the sewage treatment plant by the artificial neural network, characterized in that it consists of the inflow flow rate and inflow component concentration prediction step (S40) for evaluating the predictive performance of the trained neural network using the verification data set for the artificial neural network Method of predicting concentration. The method of claim 4, wherein The data preprocessing step (S20) may include: removing an outlier that may reduce the reliability of data among the data secured from the sewage treatment plant; Reducing or eliminating short term variation in the long term data set to determine linear or nonlinear correlation between the measured time series data and to identify trends over time of successive data; A method for estimating the inflow flow rate and inflow component concentration of a sewage treatment plant by an artificial neural network, comprising: standardizing all data distributed on different scales so as to have a distribution of values between 0 and 1. The method of claim 5, A method of predicting inflow flow rate and inflow component concentration of a sewage treatment plant by an artificial neural network, using a control chart technique, to remove outliers of the data. The method of claim 5, Inflow flow rate and inflow of the sewage treatment plant by artificial neural network, characterized by using the 5-day moving average technique to determine the linear or nonlinear correlation between the measured time series data and the trend over time of the continuous data. Method of predicting component concentration. The method of claim 4, wherein In the predictive model establishment step (S30), the neural network structure and the number of epochs are determined through trial and error, and the weight values for generalizing the trained neural network are determined based on the selected neural network structure and the epoch number. Prediction method of inflow flow rate and inflow component concentration of sewage treatment plant by artificial neural network.

Images