Você está na página 1de 8

INTELLIGENT METHODS IN DOSING CONTROL OF WATER TREATMENT

Esko Juuso*, Katja Viirret and Marjatta Piironen Kemira Chemicals Oy, Oulu Research Center PL 171, 90101 Oulu, Finland Phone: + 358 10 861 214 email: forename.surname@kemira.com *Control Engineering Laboratory, Infotech Oulu and Department of Process Engineering, University of Oulu, PL 4300, 90401 Oulu Phone: +358 8 553 1011, Fax +358 8 553 2466 email: esko.juuso@oulu.fi

ABSTRACT: Water treatment includes many complex phenomena, such as coagulation and flocculation. These reactions are hard or even impossible to control satisfyingly by conventional methods. That is the main reason why the applicability of the intelligent methods to the control of wastewater purification chemicals is discussed in this paper. There exist research articles concerning this subject. According to these articles the field test results have been promising. Linguistic equations method, which is chosen as the method in our case, is a new intelligent method in this line of process industry. For this research, data were recorded from a real water purification process. Two kinds of models were derived for the process by using this information, namely a steady-state and a dynamic model. When the quality or amount of incoming water changes crucially, a new operating point can be defined by the static model. The dynamic model for its part predicts a reasonable dosing rate for the chemicals in the current working point.

KEYWORDS: Linguistic equations, intelligent methods, water treatment

INTRODUCTION
The optimum chemical dosage rate is presumably the goal of every chemical water treatment plant. However, due to difficulties in on-line measurement and complexity of chemical reactions, the optimum dosage is very hard to obtain by conventional methods. Therefore the chemicals are usually dosed on the basis of different laboratory tests based on offline measurements. Thus, operators' experience and their attention affect the control actions. Due to the facts, mentioned above the chemicals are often over-dosed. These problems have been handled by intelligent methods. The applications have been implemented in different water treatment plants, e.g. drinking water and municipal wastewater purification plants. The information gained from these research works can also be utilised in developing the control systems of wastewater treatment processes in other fields. Neural networks were used as the intelligent method almost in every article. The backpropagation algorithm seems to be used in most of the cases as the training method. Fuzzy systems where also used successfully in a couple of cases. In one of the applications a genetic algorithm was used for parameter tuning of the fuzzy model. The papers dealing with this subject are listed in the reference list at the end of this paper. Linguistic equations have been used in direct process control but not yet in dosing control of water treatment processes. We chose the linguistic equation method for our case. Unlike fuzzy systems, linguistic equations do not necessarily need expert knowledge. On the other hand, data sets do not have to be as large as for neural networks. This feature is important in our case because the data we handled was quite small. Linguistic equations coped well with our small set of information, especially because the results can be assessed by expertise in various stages of system development.

LINGUISTIC EQUATIONS
Linguistic equations have a close connection to fuzzy systems. Originally, linguistic equation (LE) systems were used in tuning of fuzzy systems. Recently, LE systems are used more and more in direct applications. The knowledge is represented by an interaction matrix, and the nonlinearities are handled by membership definition. In this way, the systems can be adapted to various operating conditions. The rules of fuzzy systems can be considered as a drawback especially when tuning a fuzzy control system. Linguistic equations method has such advantages, which eliminate the problems, related to this rule-baseness. The rule base of fuzzy logic controller is build by interviewing the process experts e.g. operators. It can be quite tough job because of the human factors. When using the linguistic equations the software program sketches scenarios to which the operators are supposed to answer. The control actions can also be extracted from the real control situations. Thus, the experts do not have to express their operating actions in a form of rules. LINGUISTIC EQUATION CONTROLLER Lets consider a fuzzy PI type controller which has two input variables e = error, e = change of error and one output variable u = change of control. The rulebase of a PI fuzzy controller is shown in the Table 1 and in Figure 1. In this case, the linguistic values are based on the following symbols: P = positive, N = negative, ZE = zero, B = big and S = small. 1. Rules of a fuzzy controller Rule e e u 1 IF PB AND NB THEN ZE 2 IF PB AND NS THEN PS 25 IF NB AND PB THEN ZE

PB PS

ZE NS NB NB NB NB

PS ZE NS NB NB NS

PB PS ZE NS NB ZE

PB PB PS ZE NS PS

PB PB PB PS ZE PB

ZE NS NB

e
Figure 1. Rule base of a fuzzy controller The rules can be replaced by linguistic relations and further by linguistic equations. Originally, the linguistic values NB, NS, ZE, PS and PB of the variables were replaced by integer numbers -2, -1, 0, 1 and 2. And so we get a matrix, where the rule base is represented by these integers. See Figure 2. The tuning stage is improved by using real-valued linguistic equations, and recently, an interaction matrix with real values has been introduced. More generally, a set of linguistic rules or relations can be changed into a compact equation

A
j =1

ij

X j = 0,

(1)

2 1

0 -1 -2 -2 -2 -2

1 0 -1 -2 -2 -1

2 1 0 -1 -2

2 2 1 0 -1 1

2 2 2 1 0 2

0 -1 -2

Figure 2. Rulebase presented by integers. where Xj is a linguistic level for the variable j, j=1m, and m is the number of variables, coefficients Aij {-1,0,1} define the directions of the interactions. For example, when linguistic value for e is NB and PS for e then u can be calculated as follows using the previous Equation (1):

( 1) X u + 1 X e + 1 X e = 0, ( 1) X u + 1 ( 2) + 1 (1) = 0,
X u = 1 .
Several sets of linguistic relations can be combined by the following matrix presentation

(2) (3) (4)

AX = 0.
The linguistic equation of a PI controller,

(5)

u = e + e,
is a special case of the matrix equation (5) with the interaction matrix A = [1 1 -1], and X = [e e u]T.

(6)

LE controller can be made adaptive and it can be tuned with neural networks. The amount of data needed in this method is small and the operation of the controller is clear. Each parameter has a special meaning for the overall control task.

PROCESS DESCRIPTION
In this water treatment process amount of some impurities and turbidity of the outlet are diminished by dosing certain chemicals. The water treatment process consists crudely of wastewater input, the chemicals, which are dosed into this flow and the outlet of the separation process. Fig. 3. Chemicals L Wastewater Water treatment process Treated water

Figure 3: Process description

PROCESS MODELLING
Process data were collected from the water treatment process described above. The data were used for building two models, namely a steady-state and a dynamic model. The steady-state model defines a new operating point every time the flow rate or turbidity of incoming water changes crucially. The dynamic model predicts the effects of the chemical dosage in the stable operating area. STEADY-STATE MODEL The steady-state model has four inputs: flow rate of incoming water, its turbidity and the amount of chemical added. The output is turbidity of treated water. The model can be used as a feedforward controller. The controller defines the operating point on the basis of the quality of wastewater when the operating conditions change drastically. Defining the membership definitions was the first step in building a model using the linguistic equations method of the recorded data. Figure 4 shows the membership definition curve for flow rate. The curve is defined by five parameters. Flow rate

Figure 4: Membership definitions An example set of membership functions is shown in Figure 5. They are derived from the membership definitions. The locations of the membership functions are defined manually for some variables. In three variable case, this is done for two variables, and the LE model defines the membership functions for the remaining (third) variable.

Figure 5: Membership functions generated of membership definitions A four variable linguistic equation was derived for the process. Figure 6 shows the actual and calculated values for the flow rate of incoming water. The surface in Figure 7 describes the interactions between chemical dosage and turbidities of incoming and treated water as they should be in feedforward control. Since Figure 7 is based on normal flow rate of the incoming water, the interaction coefficient of the flow rate is not shown.

120 100 virta 80 60 40 20 Error 0 -20 Fuzziness 2 0 -2 0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30

Figure 6. The actual and the calculated values for flow rate.

[0.41392 -0.44325 -1]

high 0.5
0.45 0.4 0.35 pam+ 0.3 0.25 0.2 0.15

Figure 7. Model surface low 0


100 2500 200 3000 3500

low 0.1 2000 low

high 4000

300 high

Turbidity of the outlet sameus2

Turbidity sameus1 of the inlet Figure 7. Model surface

DYNAMIC MODEL Deriving the dynamic model has exactly the same stages as deriving the static model. However, the data had to be preprocessed to bring out the dynamics of the data. The chemical dosing rate at the moment k+n and turbidity of the outlet at the moment k+n were chosen as the inputs of the model. The output was turbidity of the outlet at the moment k+(n+1). Naturally, the delay of the process had to be taken into account. Figure 8 shows both the actual and the calculated values for outlet turbidity at the moment k+(n+1). And the Figure 9 shows the interaction surface of the inputs and the output of the model

1 0.8 0.6 0.4 0.2 0 0.5 0 -0.5 0.05 0 -0.05 Figure 8. The actual and the calculated values of outlet turbidity at the moment k+(n+1).

high Turbidity(k+(n+1)) low low high Turbidity(k+n) high low Chemical dosing rate Figure 9. The interaction surface of the inputs and the output of the model.

DOSING CONTROL
The controller that predicts the optimum dosing rate for the chemical consists of two parts as can be seen in Figure 10. The other part is a feedforward controller that defines a new operating point on the basis of measurements of incoming water. Thus, the change to the chemical dose can be made before large changes in turbidity values can be seen in the outlet. After definition of the operating point, the minor changes in turbidity values are measured from the outlet and the feedback controller fine-tunes the chemical dose. The working point model and local dynamic models will be used for detecting changes in operating conditions. Chemicals Turbidity Flow rate Wastewater Water treatment process Turbidity Treated water

Figure 10. Dosing control The steady-state model succeeded quite well. The linguistic equation model was able to calculate a good prediction for every variable in the model by using the information of other variables. Specialised models were developed for high and low level flow rates of the incoming wastewater since detecting some interactions was disturbed by effects of the flow level in a small data set. The dynamic model proves that the linguistic equations method is an efficient technique also for dynamic modelling. The model can predict the turbidity of outlet at the moment k+(n+1), when turbidity and chemical dosage are given to the model at the moment k+n (Figure 8). Nevertheless, our simulation experiment brought out the fact that the dynamic model needs improvement. The turbidity of the incoming wastewater must be included to the model. Its also clear that more frequent measurements are needed for control development because the delay of the process is less than the online measurement frequency.

CONCLUSIONS
Linguistic equations are suitable for both steady-state and dynamic modelling of water treatment processes. Linguistic equations provide a feasible basis for adaptive dosing control. The steady-state model can be directly used as a feedforward controller. The dynamic model needs improvement before using it in the controller design. Also the frequency of on-line measurements should be increased.

REFERENCES
Baba, K., Enbutsu, I., Matuzaki, H., Nogita, S. 1990. Intelligent Support System for Water and Sewage Treatment Plants which Includes a Past History Learning Function Coagulation Injection Guidance System Using Neural Net Algorithm. Enbutsu, I., Baba, K., Hara, N., Waseda, K., Nogita, S. 1993. Integration of multi AI Paradigms for Intelligent Operation Support Systems Fuzzy Rule Extraction from Neural Network. In: Water Sci Technol, vol. 28, no. 11-12, p. 333-340. Evans, J., Enoch, C., Johnson, M., Williams, P. 1998. Intelligent Based Auto-coagulation Control Applied to a Water Treatment Works. Gagnon, C., Grandjean, B.P.A., Thibault, J. 1997. Modelling of Coagulant Dosage in a Water Treatment Plant.

Han, T-H., Nahm, E-S., Woo, K-B., Kim, C.J., Ryu, J-W.1997. Optimization of Coagulant Dosing Process in Water Purification System. Juuso, E. K. 1999. Fuzzy control in Process Industry: The Linguistic Equation approach. In:Verbuggen, H. B., H.-J. Zimmermann and R. Babuska, editors, Fuzzy Algorithms for control, International Series in intelligent Technologies, pp. 243-300. Kluwer, Boston. Juuso, E. K., Balsa, P., Leivisk, K. 1997. Linguistic Equation Controller Applied to a Solar Collector Field, Proceedings of the European Control Conference - ECC97, Brussels, July 1-4, 1997, volume 5, TH-E 14, Paper 267, 6 pp. Nahm, E-S., Lee, S-B., Woo K-B., Lee, B-K., Shin, S-K. 1996. Developement of Optimum Control Software Package for Coagulant Dosing Process in Water Purification System. Riissanen, T., Eklund, P. 1996. Working within a Fuzzy Control Appliation Developement Workbench: Case Study for a Water Treatment Plant. Zhang, Q., Stephen, J.S. 1999. Real-Time Water Treatment Process Control with Artificial Neural Networks.

Você também pode gostar