JPH06328092A - Data collecting method in control system of biological treatment system - Google Patents

Abstract

PURPOSE:To provide a teacher data collecting method in the case of applying a neural net control system for controlling a treatment system to a real plant. CONSTITUTION:In the case of SRT control for a treatment system performed by a neural net control system, the return sludge flow rate QR is controlled to various values by a QR control section 60 in the state of controlling the excessive sludge drawn solid amount Mw to the given value by an Mw constant control section 30. At that time, data such as the sludge capacity index SVI, the average sludge retention time SRT, the return sludge flow rate QR, the inflow water quantity Qs and the excessive sludge flow rate Qw are collected, and teacher data is generated from these data items.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data collecting method in a control system of a biological treatment apparatus, and more particularly to a data collecting method in a filamentous bulking neural network control system which is a control system in biological treatment of wastewater and wastewater. It is a thing.

[0002]

2. Description of the Related Art Generally, wastewater is aerobically treated by activated sludge microorganisms (including anaerobic processes when denitrification and dephosphorization are necessary, etc.). At the time of this treatment, the activated sludge microorganisms proliferate. Therefore, in order to maintain the treatment system stable (steady), it is necessary to discharge a part of the activated sludge to the outside of the treatment system as excess sludge.

As the surplus sludge amount control (excess sludge control), which is the control of the discharge amount in this case, the following has been proposed up to the present, and a part thereof has been put into practical use.

(1) A method for setting a target amount of sludge to be drawn per day and pulling it out until the surplus integrated flow rate reaches its target value (quantitative pulling control), (2) A certain ratio of the amount of sludge in the aeration tank is changed every day. How to pull out (Sludge day (S
(A) control), (3) a method of extracting a certain proportion of the sludge amount in the treatment system every day (average sludge retention time (SRT) control), and the above-mentioned surplus sludge amount control, the best management method is currently It is known to be SRT control.

FIG. 6 shows an example of a configuration for performing this SRT control. In FIG. 6, the detected values of the sludge volume (SV) meter 2 installed near the output of the aeration tank 1 and the activated sludge floating volume (MLSS) meter 3 are respectively sludge volume index (S
VI) It is introduced into the total 4 and the sludge capacity index SVI is calculated. The value of the calculated sludge capacity index SVI, the inflow water amount Qs by the inflow water amount (Qs) meter 5, and the return (surplus)
Based on the amount of returned sludge obtained by the sludge concentration (CR) meter 61, the final sedimentation tank 1 is calculated by the calculation processing (WS) unit 7.
The amount of sludge in 0 is calculated. Further, the calculated sludge amount MF in the final settling tank, the surplus sludge flow amount Qw from the surplus sludge flow amount (Qw) meter 11, and the above-mentioned return (excess) concentration CR are input to the SRT control unit 8. Then, based on these inputs, the SRT control unit 8 calculates the excess sludge amount, and ON / OFF controls the excess sludge pump 9 for extracting the excess sludge in accordance with this excess sludge amount to discharge a predetermined amount of excess sludge. Therefore, the processing system is kept stable.

[0006]

However, when the SRT control is realized as a control system by the conventional method described above, in order to grasp the sludge amount MF in the final settling basin, W
The S section 7 has a drawback that it requires a workstation-level computer. Therefore, the above control system could not be practically introduced without such computer equipment. Further, even in a treatment facility having this computer equipment, since the conventional calculation of the sludge amount MF in the final settling basin is based on a mathematical model, great effort is required to identify (fix) the model parameter. there were.

Therefore, the present applicant has previously proposed a control system using a neural network (neural net model) (Japanese Patent Application No. 4-247450). This control method is
In short, the set value of the average sludge retention time SRT, the value of the sludge volume index SVI, the inflow water amount Qs, and the returned sludge flow rate QR are used as input variables to the neural network model, and the surplus sludge flow rate Qw is used. The neural network control unit including the output variable is provided in place of the arithmetic processing unit and the SRT control unit in the conventional configuration. Further, in this case, the teacher data necessary for learning is created using the steady analysis program, and the parameters included in the neural model are identified using the teacher data. This identification is performed by repeating the operation of correcting the above parameters based on the difference between the output and the teacher data until the predetermined learning pattern ends. Further, by increasing the number of repetitions, the sum of squares of the error is reduced and the neural network output value and the teacher data value can be matched. According to this control method, since the neuro model is used, for example, the error back propagation method (Error) is used.
Model parameter identification is facilitated by learning such as Backpropagation Method).

However, this method is premised on that teacher data for learning has been obtained, and the collection of this teacher data must be determined in consideration of input / output variables to the neural network system. . In the above method, the teacher data is created by simulation, but in reality, the internal state of the processing system cannot be measured.

In order to solve the above-mentioned problems, an object of the present invention is to collect learning data (teacher data), which is effective when the control system based on the neural network as described above is applied to an actual plant. To provide.

[0010]

A data collection method in a control system for a biological processing apparatus according to the present invention is used for identifying a parameter in a neural network control system for controlling a processing system by the biological processing apparatus. A method of collecting teacher data, wherein the flow rate of returned sludge Q in a state in which the amount Mw of excess sludge drawn-out solid matter is constantly controlled
The teacher data is collected under the condition of controlling R.

The control of the processing system is, for example, SRT control. The collection items of the teacher data are, for example, input variables such as the sludge capacity index SVI, the average sludge retention time SRT, the return sludge flow rate QR, and the inflow water amount Qs, and the output variable of the surplus sludge flow rate Qw.

[0012]

[Function] With the constant control of the amount Mw of the excess sludge drawn solid matter, the returned sludge flow rate QR is controlled to various values, and items such as input / output variables in the control system obtained at that time are collected to create teacher data. .

By controlling the amount of solid sludge withdrawn Mw to be constant as described above, it is possible to accurately control the amount of excess sludge solid even under the condition that the sludge volume index SVI changes. On the other hand, for example, when the amount of drawn sludge per day is controlled to be constant, the actual amount of excess sludge drawn solids per day decreases because the drawn-out sludge concentration Cw decreases as the SVI increases. As a result, there are drawbacks such as an increase in the amount of sludge in the treatment system.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows an example of the previously proposed neural network control system, and FIG. 2 shows teacher data used in this control system. The respective configurations of the embodiments of the data collection method of the present invention for obtaining the data are shown.

First, the neural network control system of FIG. 1 will be described. The main difference between this control system and the conventional control system shown in FIG.
The neural network control unit 20 is provided instead of the T control unit 8.

Since the manipulated variable in the SRT control is the surplus sludge amount (usually the surplus sludge amount per day), the input item (input variable) of the neural network control influences this surplus sludge amount under the SRT control. Factors were selected, specifically, sludge volume SV30 and activated sludge floating amount MLSS
Sludge volume index SVI calculated from concentration, set value (SRTset) of average sludge retention time SRT, returned sludge flow rate Q
There are four items, R and inflow water amount Qs. In addition, output items (output variables) that are manipulated variables in neural network control
Is the excess sludge flow rate Qw.

Next, a method of collecting teacher data in the embodiment of the present invention will be described. In addition, more specifically, the teacher data are the above-mentioned four input variables (sludge capacity index SVI, set value SRTset of average sludge retention time SRT, return sludge flow rate QR, inflow water amount Qs) and surplus output variables. It is data regarding the relationship with the sludge flow rate Qw.

In the data collection method of this embodiment, items to be examined are (1) operation control method at the time of data collection, and (2)
It is a method of setting variable parameters (controllable parameters) at the time of data collection.

First, regarding the overall configuration of the operation control system at the time of collecting the learning data, considering the above input / output items,
It becomes what was shown in FIG. This operation control system is similar to the control system of FIG.
Instead, a constant Mw (amount of excess sludge drawn solids) control unit 30 is used. In addition, the excess concentration (C) for detecting the concentration of the excess sludge extracted by the excess sludge pump 9
w) A total of 12 and a turbidity (Tb) total 13 for detecting the concentration of sludge (SS) in the discharged water from the final settling tank 10 are provided, and these detected values are the detected values of the Qw total 11. At the same time, it is input to the Mw constant control device section 30. Further, the return sludge pump 91, the control valve 92, and the return sludge flow rate (QR) control unit 60 are provided, and the detected value of the return sludge flow rate (QR) meter 6 is input to the QR control unit 60 to set the return sludge flow rate. QR control unit 6 based on the value QRset
The return sludge flow rate QR is adjusted by controlling the opening and closing of the control valve 92 by 0.

In this operation control system, the Mw constant control device section 30 has the following functions.

(1) A function of setting a set value Mwset (Kg-SS / day) of the target excess sludge drawn-out solids amount Mw per day, (2) a function of setting a starting time of the excess sludge pump 9, and Excess sludge pump 9 at this set time
To output the start command to start the (3)
A function of adding up the product of the excess sludge flow rate Qw and the excess sludge concentration Cw from the start-up of the excess sludge pump 9 to obtain the integrated value Mwcal of the actual excess amount of extracted solid matter, where this integration is Mwcal = ∫ It is performed by Qw (t) × Cw (t). When the detection signals from the Qw meter 11 and the Cw meter 12 are digital signals, the integration is ΣQw (k) × Cw
(K).

(4) Set value Mwset and integrated value Mwc
Compare with al to see if the actual amount of withdrawal is the same as the set value.
Or, when it becomes more than that, a function to stop the excess sludge pump 9, and when the amount of sludge discharged from the final settling tank 10 is large, it flows out of the treatment system as discharged water per day. By obtaining the integrated value of the amount of sludge to be discharged and the sum of the amount of sludge in the discharged water and the amount of surplus sludge, the total amount of sludge drawn out of the treatment system can be accurately grasped. That is, in this case, the integrated value M
wcal is Mwcal = Σ (Qw (k) × Cw (k)
+ Qs (k) × Ce (k)). Here, the sludge concentration Ce in the discharged water can be calculated by the Tb meter 13 in terms of sludge concentration conversion.

Next, an example of a variable parameter setting method at the time of data collection will be described together with a teacher data collection method. Variable parameters are set values for constant Mw control,
It is necessary to consider two items, the set value of QR control.
Note that the inflowing wastewater amount may be a variable parameter in some cases.

First, the set value Mwset and the set value QRset in the Mw constant control and the QR control are set to Mw (1) and QR (1) respectively, and the operation is performed for a certain fixed period.
The fixed period is usually about one week to one month. still,
The larger the set value Mwset, the larger the amount of surplus sludge discharged to the outside of the treatment system, so that the treatment system is in a steady state in a short period of time. Then, when the steady state is reached, the sludge capacity index SVI, the inflow water amount Qs, the returned sludge flow rate QR, the average sludge retention time SRT, and the surplus sludge flow rate Qw, which are five input / output items to the neuro control system, are obtained. .
Here, the value of the average sludge retention time SRT is calculated from the amount M of sludge in the treatment system and the amount Mw of excess sludge drawn-out solids per day.
Calculation is performed using the formula SRT = M / Mw. Also. The value of the excess sludge flow rate Qw is an integrated value of the excess sludge amount discharged as excess sludge per day under the constant Mw control.

Next, the set value QRset of the flow rate of the returned sludge in the QR control is sequentially changed from QR (1) to QR (2) and Q.
R (3), ... Change in several steps as necessary, and the above five input / output items are obtained each time. Alternatively, even when the sludge capacity index SVI changes, the above five input / output items are obtained each time.

The above series of data is for the case where the set value Mwset in the constant Mw control is Mw (1), but if necessary, the set value Mws is added to Mw (2) and Mw (3).
Change et, and follow the same procedure for teacher data (1) to (5)
To collect.

In the control system of FIG. 1, the parameters included in the neural network model are identified by using the teacher data obtained as described above, for example, in the following procedure. Here, as the neural net model, the three-stage hierarchical model of the input layer, the intermediate layer, and the output layer shown in FIG. 3 was used. The calculation procedure up to output Y31 in this hierarchical model is as follows.

(1) X1i (i = 1 to 4), which is a value obtained by dividing the input variables SVI, Qs, QR, and RST by the full-scale value, is input to the neurons in the input layer. This is directly output to the intermediate neuron.

(2) The neuron in the intermediate layer, as shown in equation (1), outputs X1i from the input layer and weighting factor W2.
The product sum of j and 1i is calculated, and the offset value θ is added to this product sum value.
An internal variable U2j is obtained by adding 2j.

[0030]

[Equation 1]

(3) The output from the neurons in the middle layer is
Generally, as shown in the equation (2), it is obtained by converting the internal variable by a sigmoid function.

[0032]

[Equation 2]

(4) Further, in the neuron of the output layer, the internal variable U31 is obtained from the equation (3) and the output Y31 is obtained from the equation (4) by the same procedure as the neuron of the intermediate layer.

[0034]

[Equation 3]

[0035]

[Equation 4]

Next, the output Y31 of the neural network model obtained above is learned. The above-mentioned error back-propagation method is generally used for this learning. This is because the error between this output Y31 and the teacher data (here, the excess sludge flow rate Qw) corresponding to the input data at that time is obtained under various input conditions, and the sum of squares of this error is used as the evaluation function to obtain this value. This is a method of correcting the parameters (weighting coefficient W, offset value θ) included in the neural control so as to decrease. Further, the steepest descent method is usually used for the parameter correction algorithm in this case. In the present embodiment, the probabilistic descent method by the remainder (Hideki Aso "Neural network information processing" industry book 1990) was used.

FIG. 4 shows a specific procedure for the above learning. First, parameters (weighting coefficient W, offset value θ)
Are initialized with random numbers 0 to 1, for example. Then, based on the neural network model, internal variables and output values in the intermediate layer and output layer units are calculated from the input data. Further, the parameters of the output layer and the intermediate layer are corrected based on the error between the output Y31 from the output layer and the teacher data T31. Then, the above operation is performed by the learning pattern (for example, 33
Repeat until one set of learning patterns) is completed. When this learning pattern is completed, the learning pattern is executed the set number of times.

FIG. 5 illustrates the above learning result. N is the number of repetitions described above, and (A) to (D) of FIG.
N = 1, N = 53, N = 500, N = 1002
It shows the case of. In these figures, the solid line is the teacher data, and the dotted line is the output Y of the neural network model.
31 is shown.

From this result, the parameters (weighting coefficient W, offset value θ) are set so that the sum of squares of error SUMX decreases as the number of iterations increases, and the output value of the neural network and the value of the teacher data match. It turns out that has been updated.

The control system of FIG. 1 can be applied to various processes such as the oxidation ditch method and the A / O method other than the standard activated sludge method described above.

[0041]

As described above, according to the present invention, since the teacher data is collected under the constant control of the excess sludge drawn solid amount Mw and the control of the returned sludge flow rate QR, the control by the neural network is performed. The teacher data for learning when the system is applied to an actual plant can be collected easily and accurately.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an example of a neural network control system according to the present invention.

FIG. 2 is a configuration diagram showing an embodiment of a data collection method of the present invention.

FIG. 3 is an explanatory diagram of a hierarchical neural network model used in the control system.

FIG. 4 is a flowchart showing a learning procedure in the control system.

FIG. 5 is a characteristic diagram of an experimental example showing a learning result in the control system.

FIG. 6 is a configuration diagram of a conventional SRT control device.

[Explanation of symbols]

 1 ... Aeration tank 2 ... Sludge volume (SV) total 3 ... Activated sludge floating amount (MLSS) total 4 ... Sludge volume index (SVI) total 5 ... Inflow water amount (Qs) total 6 ... Return (excess) sludge concentration (CR) Total 9… Surplus sludge pump 10… Final sedimentation tank

Claims (3)

[Claims] 1. A method of collecting teacher data used for parameter identification in a neural network control system for controlling a treatment system by a biological treatment device, wherein a fixed amount of excess sludge drawn solid matter (Mw) is controlled. A method for collecting data in a control system of a biological treatment apparatus, comprising collecting the teacher data under the condition of controlling the return sludge flow rate (QR) in the above state. 2. The data collection method according to claim 1, wherein the control of the treatment system is an average sludge retention time (SRT) control. 3. The collection items of the teacher data are input variables such as sludge capacity index (SVI), average sludge retention time (SRT), return sludge flow rate (QR), and inflow water amount (Qs), and excess sludge flow rate. 3. The data collection method according to claim 1, which is an output variable of (Qw).