You are on page 1of 6

Engineering Applications of Articial Intelligence 15 (2002) 385390

A fuzzy model for an anaerobic digester, comparison with experimental results


M. Polit*, M. Estaben, P. Labat
! de Perpignan, Av. de Villeneuve, 66860 Perpignan Cedex, France LP2A et LEA/SICA, Universite

Abstract This paper presents a fuzzy dynamic model for anaerobic digestion processes. A mass-balance model is developed with numerical values coming from the literature. The output gaseous ow rates are calculated with this model. In order to take into account the temperature and pH inuence that is not present explicitly, the growth rate for each biomass is multiplied by a fuzzy coefcient function of these two parameters. The model is tested on a set of experimental data and against a different model. The comparison is good regarding the aims of the model. r 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Fuzzy logic; Anaerobic processes; Modelling; Mass-balance equation; Waste-water-treatment

1. Introduction The internal working conditions and dynamics of wastewater treatment processes are badly grasped and many problems of methodology in modelling remain to be solved. It is difcult to take into account the numerous factors that can inuence the specic bacterial growth rate and the yield coefcients that characterise microorganism growth. It should also be noted that the lack of accuracy in the measurementsand some times the lack of measurementoften leads to identication problems. Mathematical models used for control engineering can be classied in two classes: those based on differential equations derived from physical laws without any use of measurements of the systems (white box model) and those entirely derived from data using no physical insight whatsoever (black box model). In fact, if the black box modelling is combined with some physical knowledge from the system a model referred commonly as grey box model will be obtained. For the reasons previously evoked, this solution seems to be the most appropriate in the modelling of real-life processes (Lindskog, 1997).

It is necessary to model the anaerobic digestion for almost two aims:


*

To estimate, from indirect measurements, quantities for which no sensor is available (or too expensive). To control processes (regulation around some nominal set point, optimal controly).

Whatever this aim will be, it should be made explicit, because it should very strongly inuence the modelling procedure. In this work, the model is elaborated in order to allow simulations for testing and tuning a controller and to calculate parameters that are not measured on-line but are useful for verifying process working. The model is constructed using rst, mass balance equations and modelling some parameters with fuzzy logic. The obtained results are compared with a set of experiences performed on a xed bed reactor and with the results of an other model.

2. The model description 2.1. Mass balance equations

*Corresponding author. Tel.: +33-4-68-66-17-55,; fax: +33-4-6866-22-87. E-mail address: polit@univ-perp.fr (M. Polit).

Using the work of Denac, et al. (1988), a dynamic model for a xed bed reactor or an UASB one is elaborated. Concentrations in the liquid phase of the

0952-1976/02/$ - see front matter r 2003 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 2 - 1 9 7 6 ( 0 2 ) 0 0 0 9 1 - X

386

M. Polit et al. / Engineering Applications of Articial Intelligence 15 (2002) 385390

reactor are supposed to be constant. The microbial kinetics are described by the law of Monod (1942):
*

biomass Xi growth rate: rXi mi kdi ; 1

A mathematical non-linear model with 10 state variables (5 substrates and 5 biomasses) is obtained. It is linearized around an equilibrium point. 2.3. The fuzzy part of the model 2.3.1. Modication of m As some parameters like temperature, pH and alkalinity are not constant over a long period of time, their inuence on the process evolution has to be taken into account. No parameter in the model is explicitly varying with pH or temperature. A possible solution is to consider that m varies with those two parameters. Bastin and Dochain (1990) expresses m as the multiplication of individual terms, each of them referring to one of the inuencing factors. mt mS mX mpHmT mP::: 8

substrate Si consumption rate: Xi rSi mi ; Yi mi Si mimax Si Si KSi

2 3

with mi max being the maximal growth rate for the biomass i (h1); KSi the saturation constant (g COD/1); Kdi the death rate organisms (h1); Yi the conversion yield of biomass into substrate (g biomass/g COD). For each substrate and each biomass involved in the anaerobic digestion (i.e. glucose, acetic, butyric and propionic acids and hydrogen and the associated biomasses), a mass balance equation is written as substrate Si : r X dSi Qin Sin Si aj r S j 4 dt VT j 1 with Qin being the input liquid ow rate (in l/h), Sin the input concentration (in gCOD/l); VT the total volume of the reactor (in l); r the number of anaerobic reactions; aj the stoichiometric coefcients. biomass Xi : dXi r xi : 5 dt

with P being the synthesis product concentration in the liquid phase. All the factors m; excepted mS; are between 0 and 1. Rozzi (1984) uses, for example, a parabolic law derived from experimental data, to treat the inuence of the pH: mpH a pH2 b pH c 9

with a; b; c being the constants. Topiwala and Sinclair (1971) modelled the inuence of temperature by an Arrhenius type law mT 8 > 1 =RT < a1 exp E a2 exp E2 =RT b > : 0

if T1 pT pT2 ; if T oT1 or T > T2

10

2.2. The gaseous ow rates The gaseous ow rates are determined by the following equations: r X QCH4 a1 j rSj ;
j 1

with E1 ; E2 being the activation energy, R the gas constant; a1 ; a2 ; b the constants; T1 273 K; T2 320 K: In order to obtain a model taking into account the existing experimental results and easy to modify, a fuzzy coefcient m% (Pedrycz, 1993) function of temperature and pressure is introduced. Each mi Si is modied, being multiplied by m% : mi Si -mi Si m% with

QCO2 QH2

r X j 1

a2 j rS j ; 6 7

m% 0:5 mf and 0:5pmf p0:5: To calculate the coefcient mf, using fuzzy logic, the experimental data from the LBE/INRA (Narbonne) were used to design the fuzzy subsets. The rst considered equilibrium point is given by pH 7; T 351C; Qin 8:33 l=h; Sin 10 g COD=l:

r X j 1

a3 j rS j ;

Qgas QCH4 QCO2 QH2 ; a1 j; a2 j a3 j

where and are the stoichiometric coefcients for the chemical reactions in which methane (CH4), carbon dioxide (CO2) and hydrogen (H2) are produced. The parameter and the stoichiometric coefcient values come from the literature (Denac et al., 1988).

M. Polit et al. / Engineering Applications of Articial Intelligence 15 (2002) 385390

387

The values of Qin and Sin are chosen in order to correspond to a residence time y of 5 days. pH and temperature are varying around this equilibrium point as 6:1ppHp7:9 and 13pT p57o C: For each pH value and each value of temperature, we calculate the corresponding value of mf : pH and temperature are fuzzyed as shown in Fig. 1. The results of Eqs. (9) and (10) are used to build the bound limits of the output fuzzy subsets (see Fig. 2). In Figs. 1 and 2, the different fuzzy subsets are named: LN, large negative; MN, medium negative: SN, small negative. ZE, zero.
pH 1 LN MN SN ZE 0 6.1 6.32 6.55 6.78 7 T 1 LN MN SN ZE 0 13 18.5 24 29.5 35 40.5 46 51.5 57 T (C) SP MP LP 7.22 7.45 7.67 7.9 pH SP MP LP

LP, large positive; MP, medium positive; SP, small positive. The shape and the number of those subsets were chosen in order to optimize the results in term of precision and simplicity. The 49 inference rules that connect the input fuzzy subsets of the pH and T with the output subsets of mf are given in Table 1. An example of using this table will be as follows: If pH is SN AND if T is MP THEN mf will be MP... . 2.3.2. Values of m The inference rules give a resulting membership amf for the output variable mf : To convert this fuzzy information into a precise value, the centroid of the membership function amf Dmf is calculated. After that, the values of mf are calculated. 2.4. Calculations of the gaseous ow rate As T and pH are varying, mi is varying too, by mf : The new values of mi are introduced in the preceding equation system and the gaseous ow rate variations, function of pH and temperature are calculated. The results and those that are obtained with the values of mi from Eqs. (9) and (10), are plotted in Fig. 3, they are not very different.

Fig. 1. Fuzzy subsets for the pH and the temperature.

Table 1 Inference rules T pH LN MN LN LN MN ZE SP ZE LN SN LN LN MN SP LP MP LN ZE LN LN MN SP LP MP LN SP LN LN MN SP LP MP LN MP LN LN MN ZE SP ZE LN LP LN LN LN LN SN MN LN

f 1 LN MN SN ZE 0 -0.5 -0.45 -0.3 -0.15 0 0.15 0.3 0.45 0.5 f SP MP LP


LN MN SN ZE SP MP LP

Fig. 2. Fuzzy subsets for mf :

LN LN LN LN SN MN LN

Gaseous flow rate

Gaseous flow rate 7.5 7 pH 6.5 6 0 20 40 Temp (C)

8 6 4 2 0 8

8 6 4 2 0 8

60

7.5 pH

6.5

20

40 Temp (C)

60

(a)

(b)

Fig. 3. Gas ow rate variations with m from Eqs. (9) and (10) (a) and from this calculation (b).

388

M. Polit et al. / Engineering Applications of Articial Intelligence 15 (2002) 385390

80 Gas flow rate (l/h) 70 60 50 40 30 20 60 70 80 90 100 110 120 130 140 150 Time (h) 350 Gas flow rate (l/h) 300 250 200 150 100 700
Qgascalculated Qgas-measured Qgascalculated Qgas-measured

130 Gas flow rate (l/h) 120 110 100 90 80 70 60 320 340 360 380 400 420 Time (h) 400 350 300 250 200 150 100 50 0 780
Qgas-measured Qgascalculated

Gas flow rate (l/h)

Q gascalculated Q gas-measured

720

740

760

780

800

800

820 Time (h)

840

860

Time (h)

Fig. 4. Comparison between calculated and experimental values of Qgas :

Simulated total gas flowrate (l/h)

350 300 250 200 150 100 50 0 0 50 100 150 200 250 300 Measured total gas flow rate (l/h)

Fig. 5. Comparison between calculated and experimental values of Qgas set points.

compared with the experimental values in Fig. 4. In Fig. 5, the different mean values of the calculated values of Qgas function of the measured values are plotted. In this comparison, the agreement between the two values is good when Qgas is constant (see Fig. 5). When Qgas is varying quickly, the transition between a set point and another one corresponds to a change in the set of parameters in the model (see Fig. 4). The dynamic of this transition is not fast enough and the settling time of the model is too high in comparison with the experiment. The improvement of the model dynamic is explained in part 4.

4. Tuning of the model and results 3. Comparison with the experimental results 4.1. Tuning of the settling time In order to compare the results of the model with experimental results obtained at the LBE of INRA (Narbonne), for the different working points (different values of Qin ; Sin ; pH and T ) the CH4, the CO2 and the H2 ow rates are calculated. The experimental set-up from the LBE is, for example, described in Dochain et al. (2000). The comparison between the experimental and the calculated values for those ow rates shows that the inuence of the pH and T is not expressed enough through the values of m: Two fuzzy coefcients (one for CH4 and one for CO2), function of pH and T are introduced to take this inuence more into account. The new values calculated for the gaseous ow rate are In order to ameliorate the settling time of the model, two new parts are introduced:
*

The rst one in order to take into account the inuence of Sin on Qgas ; a retroaction is introduced. In minimising the difference between Qgas-calculated and Qgas-measured ; Sin is sometimes modied (Rivoire and Ferrier, 1990). The second one is a dynamic part that is introduced when the input variables are signicantly varying, in order to evaluate the output gaseous ow rate evolution.

M. Polit et al. / Engineering Applications of Articial Intelligence 15 (2002) 385390

389

Qin Sin pH System

Qgas-measured

T Q QC02 + + gas-calculated QH2 + QCH4 +

Model Sin

Fig. 6. Principle of the model.

Qgas-measured

Qgas-fuzzy

Qgas-mod1

Qin

350 300 Gas flow rate (l/h) 250 200 150 100 50 0 780

100 80 60 40 20 0 850 Qin (l/h) Qin (l/h)

790

800

810

820

830

840

Time (h)

Qgas-measured

Qgas-fuzzy

Qgas-mod1

Qin

120 110 Gas flow rate (l/h) 100 90 80 70 60 50 40 350 360 370 380 Time (h) 390 400

50 40 30 20 10 0 410

Fig. 7. Comparison between the fuzzy model, the Bernard et al. model (mod 1) and the experimental data.

The system is then considered as a rst order. The gain and the settling time are evaluated with fuzzy logic from the experimental data and a fuzzy coefcient is introduced like in the static part. This modication allows the introduction of a faster change from a set point to another.

4.2. Results and discussion Bernard et al. (2001) have developed two steps (acidogenesismethanization) mass balance model. It incorporates electrochemical equilibria in order to include the alkalinity. A step-by-step identication procedure to estimate the model parameters was made. An example of the comparison for the gaseous ow rate between our model, the experimental data and the Qgas calculated with the dynamical model of Bernard et al. (2001) is given in Fig. 7. The comparison is now

The principle of this new model is depicted in Fig. 6. The gaseous ow rate is calculated with the modied model.

390

M. Polit et al. / Engineering Applications of Articial Intelligence 15 (2002) 385390

very good even when the gaseous ow rate is varying fast. The advantage of this model in comparison with theoretical model is that its parameters can be easily adapted to different evolutions of the process and to different processes. A short adaptation period is only necessary.

wood industry in order to improve the process efciency (AMOCO). It does not necessary reect its views and in no way anticipates the Commissions future policy in this area. References
Bastin, G., Dochain, D., 1990. On-line Estimation and Adaptive Control of Bioreactors. Elsevier, Amsterdam. Bernard, O., Hadj-sadok, Z., Dochain, D., Genovesi, A., Steyer, J.P., 2001. Dynamical model development and parameter identication for an anaerobic wastewater treatment. Biotechnology and Bioengineering 75 (4), 424438. Denac, M., Miguel, A., Dunn, I.J., 1988. Modelling dynamic experiments on the anaerobic degradation of the molasses wastewater. Biotechnology and Bioengineering 31, 110. * A., Roca, Dochain, D., Steyer, J.P., Genovesi, A., Franco, A., Punal, ! rez, D., 2000. E., Lema, J.M., Polit, M., Lapa, R., Canals, J., Pe Proceedings of the Fifth International Symposium on Systems Analysis and Computing in Water Quality Management, Watermatex, Gent, Belgium, 2000. Lindskog, P., 1997. In: Helledoorn, H., Driankov, D. (Eds.), Fuzzy Model Identication. Springer, Berlin. Monod, J., 1942. Recherches sur la croissances des cultures bact! riennes. Hermann, Paris. e Pedrycz, W., 1993. Fuzzy Control and Fuzzy System. Research Studies Press Ltd, Taunton Somerset, England. Rivoire, M., Ferrier, J.L., 1990. Commande par calculateur identicateur. Eyrolles, Paris. Rozzi, A., 1984. Modelling and control of anaerobic digestion processes. Transactions of the International Measurement and Control. 6, 153159. Topiwala, H., Sinclair, C.G., 1971. Temperature relationship in continuous culture. Biotechnology and Bioengineering 13, 795813.

5. Conclusion In the paper, a real-world modelling situation was exposed. The process, as almost always in such situation, was not completely known. Uncertain factors affecting the model were modelled using fuzzy logic and expertise on the process. The obtained results are in good agreement with the experimental ones and with those obtained by a theoretical method. The next step in the study will be to use a method to estimate parameters that are not measured on-line in the process, such as the alkalinity or the VFA. So this kind of model will be used to evaluate in real time the quantities that are necessary to control the process and to establish fault diagnosis.

6. Acknowledgements This study has been carried out with nancial support from the Commission of the European Communities, Agriculture and Fisheries (FAIR) specic RTD programme, CT96-1198, Advanced Monitoring and Control of the operation of wastewater treatment for the

You might also like