0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)
22 Ansichten11 Seiten
This paper deals with the control of fixed bed bioreactors. The dynamics of these processes are described by a nonlinear distributed parameter model. An adaptive controller is then designed, which is based on the nonlinear discretized model. Its performance is illustrated by simulation results on a fixed bed anaerobic waste water treatment process.
This paper deals with the control of fixed bed bioreactors. The dynamics of these processes are described by a nonlinear distributed parameter model. An adaptive controller is then designed, which is based on the nonlinear discretized model. Its performance is illustrated by simulation results on a fixed bed anaerobic waste water treatment process.
This paper deals with the control of fixed bed bioreactors. The dynamics of these processes are described by a nonlinear distributed parameter model. An adaptive controller is then designed, which is based on the nonlinear discretized model. Its performance is illustrated by simulation results on a fixed bed anaerobic waste water treatment process.
Printed in Gr eat Britain. 0005-1098/92 $5.1]0 + 0.00 Pergamon Press Lid 1~) 1992 International Federation of Aut omat i c Control Modelling and Adaptive Control of Nonlinear Distributed Parameter Bioreactors via Orthogonal Collocation* D. DOCHAI N, t ~t J. P. BABARY[ I and N. TALI - MAAMAR[ [ Adapt i ve linearizing control can be synthesized f or distributed paramet er bioreactors based on an orthogonal collocation approximation model and its simulation perf ormance can be shown to be effective. Key Words--Distributed parameter systems; nonlinear systems; orthogonal collocation; biotechnol- ogy; adaptive control. Abstract--This paper deals with the control of fixed bed bioreactors. The dynamics of these processes are described by a nonlinear distributed parameter model. In a first step, the partial differential equations of the model are reduced to ordinary differential equations by using an orthogonal collocation method. A nonlinear adpative controller is then designed, which is based on the nonlinear discretized model. Its performance is illustrated by simulation results on a fixed bed anaerobic waste water treatment process. 1. INTRODUCTION OVER THE LAST f ew decades, t her e has been a growi ng and wi despr ead devel opment of in- dustrial bi ot echnol ogi cal processes. The oper a- t i on in St i r r ed Tank React or s ( STR) has been ( and is still) a wi del y used t echnol ogy in f er ment at i on processes. But , mor e and mor e, ot her t echnol ogi es ar e consi der ed f or bi opr oc- esses oper at i on, such as fl ui di zed bed, packed bed or ai r lift r eact or s, whi ch pr es ent several advant ages over t he "cl assi cal " STRs (e. g. Schtirgerl, 1989; Moser , 1988). For i nst ance, hi gher cell mass, mor e efficient pr oduct ext rac- tion and hi gher flow r at es ar e of t en ment i oned as advant ages of fl ui di zed and packed bed bi oreac- tors. Typi cal exampl es ar e fl ui di zed or packed * Received 7 February 1990; revised 29 January 1991; revised 26 September 1991; received in final form 8 February 1992. The original version of this paper was not presented at any IFAC meeting. This paper was recommended for publication in revised form by Associate Editor A. Bagchi under the direction of Editor P. C. Parks. $ Laboratoire d'Automatique, Dynamique et Analyse des Syst~mes, Universit6 Catholique de Louvain, B~itiment Maxwell, Place du Levant 3, 1348 Louvain-La-Neuve, Belgium. 1: Chercheur qualifi6 FNRS, Belgium. To whom correspondence should be addressed. II Laboratoire d'Automatique et d'Analyse des Syst~mes du CNRS, 7, avenue du Colonel Roche, 31077 Toulouse, France. 8 7 3 bed anaer obi c di gest ers (e. g. Poncel et et al., 1985; Poncel et , 1986) and et hanol pr oduct i on with al gi nat e ent r apped yeast s (e. g. Williams and Munnecke, 1981). In t he cont ext of t he moni t or i ng and cont r ol of f er ment at i on processes, t he use of fixed bed bi or eact or s i nt r oduces new met hodol ogi cal pr obl ems t o be sol ved. The behavi our of t hese r eact or s ar e t hen expect ed t o be mor e l i kel y in plug flow condi t i ons t han in compl et el y mi xed ones. And, strictly speaki ng, t hei r dynami cs cannot be descr i bed by or di nar y di fferent i al equat i ons ( ODEs ) as in STRs, but by part i al deri vat i ve equat i ons ( PDEs ) . Ver y f ew paper s have been concer ned in t he l i t er at ur e with t he cont r ol of di st r i but ed pa r a me t e r bi or eact or s, with t he not abl e except i on of Lut t ma n et al. (1985) (see also Isaacs et al., 1986) f or t he case of an air-lift bi or eact or . Maendeni us et al. (1987) also consi der t he cont r ol of a packed bed bi or eact or (in cascade with a Cont i nuous STR ( CSTR) ) , but t hey t r eat ed it as a CSTR, whi ch seems fai r in t hat appl i cat i on wi t h r espect t o t he small size of t he bi or eact or . In this paper , we shall concent r at e on fixed bed bi or eact or s, whi ch present s t he fol l owi ng charact eri st i cs: t he act i ve bi omass is kept wi t hi n t he vessel and t he di fferent subst rat e(s) and pr oduct ( s) flow t hr ough it. Fur t he r mor e , diffusion phe nome nons will be consi der ed her e as negligible. The di st ri but ed par amet er model of t he pr ocess will t hen be const i t ut ed of fi rst -order hyper bol i c PDEs . Ther e are basically t wo di f f er ent appr oaches to t r eat t he cont r ol of di st r i but ed par amet er systems (e. g. Ray, 1981). In t he first one, t he cont rol design is based on t he di st r i but ed 874 D. DOCHAIN et al. paramet er model and then eventually reduced to a finite dimension control solution. However , the usually very large uncertainty on the knowledge about the kinetics of fermentation processes (due to the high complexity of processes involving living organisms) renders this approach questionable in practice. Therefore we made up our mind for a second approach, the early lumping: in a first step, the PDEs are reduced to ODEs by using orthogonal colloca- tion and in a second step, the finite dimension ODEs are considered for the control design. The objective of this paper is to propose control algorithms for fixed bed bioreactors, which are based on the reduced nonlinear ordinary differential equation system. The central question we address in this paper is to suggest how to extend the application of adaptive linearizing control schemes to nonlinear distributed paramet er bioprocesses. In that sense, we believe that our paper is a maybe modest, but yet very important, step towards the extension of the very promising adaptive linearizing control strategies to processes of higher and higher practical importance and complexity. We shall concentrate here on the regulation of one out put , i.e. of one component concentration at the out put of the bioreactor. The performance of the controller will be illustrated on a typical case of application: anaerobic waste t reat ment process. The design of the proposed control schemes is in line with recent works on the control of STR bioreactors (Dochain and Bastin, 1984; Dochain, 1986; Bastin and Dochain, 1990) and presents the following basic features: it is based on the well-known nonlinearities of the process and is adaptive in order to deal with the paramet er uncertainty. An important section is dedicated to the modelling of the dynamics of fixed-bed bioreactors in absence of diffusion. The pro- posed model fits into the General Dynamical Model framework introduced in Bastin and Dochain (1990) for Stirred Tank Bioreactors. The paper is organized as follows. Section 2 introduces the distributed paramet er dynamical model of fixed bed bioreactors. Its reduction to ODEs by orthogonal collocation is present ed and discussed in Section 3. Section 4 is dedi cat ed to the design of an adaptive linearizing control law. In Section 5, the performance of the adaptive controller is illustrated by simulation experiments. 2. DYNAMI C AL MODE L OF F I XE D B E D B I OR E AC T OR S Let us consider a fixed bed bi oreact or (Fig. 1) in which m biochemical reactions with n t F FIG. 1. Sc he ma t i c vi ew of a fi xed bed r eact or . components t ake place. Among these reactants, nl are micro-organisms ent rapped or fixed on some support and which remain within the reactor, and n2 ot her component s (essentially substrates and products) flow through the reactor. For simplicity, we also consider the cross-section of the bi oreact or to be constant and equal to A. From mass balance considerations and under the assumptions that diffusion is negligible, we can deduce, in line with the general model formulation in Bastin and Dochain (1990), the following dynamical model: cafg, = g , ~ ( ~ , , ~ 2 ) , ( 1 ) 9 t a~2= F a~2 - - - - + ( 2 ) 3t A 0z following boundary conditions: ~2(t, z = 0) = ~2.i,(t). above equations, ~1 is the with the (3) In the biomass concentration vect or ( d i m~ l =n l ) , ~2 is the other component s concentration vect or (dim ~2 = n 2 ) , ~2, i n( t ) is the influent concentration of ~2 (which is different from zero only for external substrates), ~(_~1, ~2) is the reaction rate vect or (dim ~ = m), K1 and/ ~2 are the yield coefficient matrices (dim/~1 = nl x m; dim K 2 = n 2 X m), F the hydraulic flow rate and z is the space variable (z e ]0, HI). Remark. It is well known (e.g. Hol mberg, 1982; Bastin and Dochain, 1990) that, in bioprocesses, the reaction rates 9, which represent the kinetics of the process, are generally nonlinear, but also often badly known and ill-defined, functions of the state of the bioreactor. A simple example Consider a bioreactor with two reactions described by the schemes below: (1) an autocatalytic growth reaction with one limiting substrate S and one biomass population X: S --~X, (4.a) Modelling and control of distributed par amet er bioreactors 875 (2) a deat h reaction of t he micro-organisms: X ~ Xd, (4.b) where X o is t he non-active biomass concentration. If we assume that the non-active micro- organisms leave t he bioreactor, t he dynamics of the above process will dear l y be described by the equations (1) and (2) with: ~ = X , ( 5 . a ) = I S T , ( 5 . 6 ) ~2,in = [Sin 0] T, (5. C) = [/.JX k d X ] T, ( 5 . d ) where /~ is the specific growth rate, kd is the death coefficient and kt is a yield coefficient. 3. REDUCTI ON OF THE DI STRI BUTED PARAMETER MODEL TO ORDI NARY DI FFERENTI AL EQUATI ONS It is clear from the above model (1), (2) that the state variables ~1 and ~2 are functions of time and space, i.e. ~l(t, z) and ~2(t, z). We shall now see how to approximate the PDEs of each component of ~ and ~2 by a finite number (equal to p + l ) of ODEs at p + l discrete spatial positions along t he bioreactor. There exist several met hods to reduce the above PDEs to ODEs (e.g. Ray, 1981). They consist of expanding the variables as a finite sum of products of time functions and space functions. In t he orthogonal collocation met hod (e.g. Villadsen and Stewart, 1967; Villadsen and Michelsen, 1978), the state variables will be written as follows: p +l ~a(t, z ) = ~ f l i ( z ) ~ l a ( t ) , (6.a) j =0 p + l ~2(t, z ) = ~ f l j ( z ) ~ 2 a ( t ) , (6.b) j=o where ~la(t) and ~2,j(t) are t he value of ~ and ~2 at some discrete spatial positions (called collocation points) along the bioreactor, i.e.: ~la(t) -- ~l.j(t, z = zj), (7.a) ~2d(t) = ~z, i (t , z = z j ) , (7.b) and the basis functions f l j ( z ) are chosen as orthogonal functions (e.g. Lagrange poly- nomials), such that: {~ if i = j flj(Z~) = if i =/:j" The integer p in equation (6) corresponds to the number of interior collocation points, z = Zo and z = zp+l corresponds to t he input (z = 0) and the output (z = H) of the bioreactor. The choice of the orthogonal collocation met hod for t he space discretization of the fixed bed bioreactor model has been dictated by two main reasons. First of all, this met hod is largely used and accepted in chemical engineering (e.g. Georgakis e t a l . , 1977; Jensen and Ray, 1982; Jorgensen, 1986) for the reduction of dynamical models of tubular reactors (which presents great similarities with the same kind of processes used in biotechnol- ogy). Secondly, orthogonal collocation is known to be an efficient and powerful reduction method. Villadsen and Michelsen (1978) have shown that the collocation met hod and the Galerkin met hod (often considered as the best approximation met hod) will give similar results if some precautions are t aken for the implementa- tion of the first one (calculation of quadrat ure points with a formula including a weight function W ( z ) = z P( 1 - z)% and for which t here exist an optimal choice). And at the same time, t he collocation met hod offers the advantages that its implementation is easier and that the nat ure and dimension of the state variables remain un- changed after the reduction procedure (in our case, the state variables will remain concentra- tions in the discrete ODE model). Moreover, it has been shown that orthogonal collocation methods preserves mass balances (e.g. Cho and Joseph, 1983). Now we can write t he partial derivative of ~2 with respect to z appearing in equation (2) as follows: , 9 ~ 2 P X ~ I d f l j ( z ) = L':'o ( 8 ) Let us note: b , j = d E ( z = z , ) d z ( 9 ) By introducing (6)-(9) into equations (1)-(3), each PDE is t ransformed into p + 1 differential equations at t he p interior collocation points and at the output of the reactor. We obtain the following ODEs system of order n ( p + 1): d x 1 dt = K l q g ( X l ' x 2 ) ' (10) dx2= - - - ; + F~ + Kz q o ( x l , x2), (11) dt 2' t 876 D. DOCHAIN et al. wi t h [ ~ 1 , ] ~ 1 , 2 X l = . , ~ l , p + l [ ~ ( ~ I , I , ~ 2 , 1 ) 1 q g ( X l ' X 2 ) ~ - - - [ ~ ( ~ 1 ' 2 ' ~ 2 ' 2 ) 1 ' t - { ~ ( ~ l , p + l , ~ 2 , p + l) J g , o . . . . . . o l 0 /'(l 0 " - " 0 J K 1 ~ . . . . . ;' 0 0 . . . . . . /{1 g2 o . . . . . . 0 - I 0 K2 0 - ' . ..o / . : , o o LJ B = [Bo]ia=,top+l wi t h e b? FR = ~ ~2 in(t) wi t h /;i = d i a g { - b ~ o } , i = l t o p + l , K 2 --~ ~ 2 ,1 1 X 2 : ~ 2 , 2 , ~ 2 , p + 1 . . I B/j = di ag {bq}, a nd t he f ol l owi ng d i me n s i o n s f or t h e di f f e r e nt ve c t or s a nd mat r i ces : di m (~i) = ni (p + 1); di m (Ki) = ni ( p + 1) m( p + 1), di m q~ = m( p + 1); di m (Bij) = di m ( bi ) = n2 n2. i = 1 , 2 , A simple example (continued) Co n s i d e r t he e x a mp l e o f Se c t i on 2 ( b i o r e a c t o r wi t h gr owt h a nd d e a t h r e a c t i ons ) . As a ma t t e r o f i l l ust r at i on, l et us s e e h o w t he r e d u c e d mo d e l will l o o k l i ke in t he s i t ua t i on wh e r e onl y t wo i nt er nal c ol l oc a t i on poi nt s a r e c o n s i d e r e d ( p = 2). Le t us first de f i ne t he v a l u e s o f t h e c o n c e n t r a t i o n o f X, S a nd Xa a nd o f t he speci f i c gr owt h r a t e # at t he c ol l oc a t i on poi nt s a n d at t h e r e a c t o r o u t p u t z; (i = 1, 2, 3): X i = X ( z = z i ) , S = 5 ( 2 ~ = z i ) , X d . i ~ - X d ( Z = Z i ) , U i = u ( X i , S i , X d , i ) . No t e al s o t hat he r e : F Bl l B12 B13 1 g = / B 2 1 B 2 2 B 2 3 / LB31 B32 B33_1 "bl l 0 b12 0 bl l 0 b21 0 b22 0 b2t 0 b31 0 b32 0 b31 0 m -b,0 0" 0 blo F b2o 0 F ~ = - ~ 0 b2o b3o 0 0 b3o m si nce Xa, i. = 0 f r om 0 b13 0 - b12 0 b13 0 b23 0 b22 0 b23 0 b33 0 b32 0 b33 [ Si. ] = F b 2 o ] S i n Xa,i.-I - - A 0 I b3o I 0 I D t he b o u n d a r y c ondi t i ons . Th e n t he r e d u c e d O D E s o f t h e pr oc e s s will b e e qua l t o: d [ X l] x2 X3 i 1 - 1 0 0 0 i ] = 0 0 1 - 1 0 0 0 0 0 1 - m $1 X d 1 dt Xa,2 F A F A "~ , x T I k d X , I #2X2 I kdX2 I ' / A 3 X 3 [ k d X 3 / 0 b21 0 b31 0 blo 0 b2o 0 b3o ~ 0 0 b12 0 b13 0" b . 0 baz 0 b13 0 b22 0 b23 0 b21 0 b22 0 b23 0 b32 0 b33 0 b31 0 b32 0 b33 . " & "l X d , l [ Xd2l S31 X a , d Modelling and control of distributed paramet er bi oreact ors 877 "_.kl 0 0 + 0 0 0 m Remarks 0 0 0 0 6 1 0 0 0 0 0 - k I 0 0 0 0 0 1 0 0 0 0 0 - k l 0 0 0 0 0 1 . m ].~ 1X1 k d Xl # 2 K2 kdX2 3X3 kdX3 (a) It is obvious that the choice of the basis orthogonal function and of the number and position of the collocation points can be a critical question in the procedure of reducing the PDE to ODE. This mat t er will be further discussed along with the simulation experiments in Section 5 (see also Babary et al., 1991). (b) It is worth to not e the similarity bet ween the reduced model of fixed bed bioreactors and the General Dynami cal Mode l of stirred tank bioreactors present ed in Bastin and Dochai n (1990), the main differences lying in the reaction rate vect or (which is now of dimension m( p + 1)) and the presence of the matrix B. 4. THE CONTROL ALGORI THM 4.1. St at ement of the cont rol pr obl em We shall concent rat e in this paper on the control of the concentration of one component at the out put of the bioreactor, under the following conditions: (C1) The flow rate F is the control input; (C2) The concentration of the controlled com- ponent is measured not only at the out put of the bi oreact or but also at each interior collocation point, and at the reactor input (in case of an external substrate); (C3) The yield coefficients are positive, constant and unknown; (C4) m l -< m reaction rates are unknown; (C5) The system (10), (11) is minimum phase (Byrnes and Isidori, 1985; Isidori, 1989). Let us denot e y the concentration along the reactor of the component which has to be controlled, y,. the value of y at the collocation points z =z,- and Y the value of the controlled component (at the react or output): yi = y ( z = z~) i = l t o p , Y = y ( z = zp+t ) Then, by definition, we can write Y as a linear combination of the state variables xt and x2, or equivalently as a linear combination of the component concentration x2 at the out put position Zp 1: Y a - c T [ X l ] , ( 1 2 . a ) LX2J ~.. T - - C2 ~2,p+l- (12.b) The dynamical equation of the out put Y is readily obtained from (11), (12): dY F F T ~ d--t = - . . CTBx2 + A C2b.+l~:2,in "]- CIK2(P(~I,p+I, ~2, p+l ) - (13) It is straightforward to check that the t erm CTBx2 is a linear combination of only the variables Yi at the different collocation points z,. By considering condition C4, the last term of the above equation can be rewritten as follows: CT cp = K21~1 + K22~2 - - ' K21I ' J r - [ 0 2 " ' " Om_m,+l][2,1''" (P2,m_m,+l] T = 0T$, (14.a) where ~1 and ~2 contain the unknown and known reaction rates, respectively, and 0 and are then equal to: 0 T= [K21@~, 02 . . . . . 0,,-,,,,+1], (14.b) $ r = [1, @2a, - . , @2,,,-,n,+1], (14.C) i.e. 0 contains all the unknown parameters. Therefore, the out put dynamical equation takes the following form: d Y F CTBx 2 F - d--7 = - A + A CTb"+*tJ2"" + 0To" (15) Equation (15) is the basis for the derivation of the controller. 4.2. Adapt i ve linearizing cont rol al gori t hm Assume that the control obj ect i ve is to have a desired closed loop dynamics equal to the following linear first-order equation: dY d t = A(Y* - Y), ). _> 0, (16) where Y* is the desired value of Y. This objective can be achieved by implementing the following model reference linearizing control law (see e.g. Bastin and Dochain, 1988, 1990), which is readily obtained by combining equations (15) and (16): F - - A )'(Y* - Y) - 0 r 0 c T~p +l ~2 , i n - - CTBx 2 . (17) From Conditions C3 and CA, the paramet ers 0 are assumed to be unknown: we shall then AUTO 28:5-B 878 D. DOCHAIN et al. estimate them via some adaptation schemes, e.g. a least squares estimation algorithm with a forgetting factor ), (e.g. Bastin and Dochain, 1990): dqJ - - = - t o w + q), (18.a) dt dqJ - t oqJ o+ t o y + F T- F d----t-= A C2bp+l~2'in- CTBx2' (18.b) dO d---t = r w ( Y - qJ0 - qJ +0) , ( 18. c) dF d--~ = -FqjWTF + yF, F(0) > 0, 0 < ), - 1. (18.d) The adaptive version of the linearizing control equation is then equal to: F = A ~'(Y* - T) - OTq) + - . ( 1 8 . O C 2 b p + l ~ 2 , i n - - Cr Bx2 Co mme n t . The above adaptive linearizing con- trol algorithm (18) is in line with similar adaptive linearizing control schemes (e.g. Bastin and Dochain, 1988) for which theoretical stability properties have been emphasized. 5, SIMULATION RESULTS Simulation experiments have been performed on the example present ed above. Assume here that we wish to control the substrate concentra- tion S at the output of the bioreactor, i.e. Y = Sp+l. Its dynamics is equal to: dSp+l = bp+,,oS~. + ~ bp+, , iS, j d t A i=l - k d z p +l Xp +, , (19) with ~up+~ =/~(Xp+~, Sp+a). The linearizing con- trol algorithm (17) is here written: A~, ( S * - St,+, ) + k d h , +, X p +, F (20) p + l - b p + l , 0 S i n - E b , + x , i S i i=1 Assume also that the deat h paramet er kd is known (e.g. from some batch experiments), and that the yield coefficient k~ and the specific growth rat e/ ~ are unknown. In this example, let us further introduce the following definitions. (1) = where p,+~ is an unknown positive function of the bioreactor state (this definition only implies, in accordance with the physical reality (there is no growth without substrate, see t he above reaction scheme), that /Up+l =0 if Sp+l =0) (see also Bastin and Dochain, 1990). (2) The auxiliary variable Z = S, +] + k , X , + l . The dynamical behaviour of Z is given by: d Z F bp+l , 0Si n + Z bp+l,iSi dt A i=1 - k d ( Z - Sp+I). (21) With the above definitions of p and Z, t he output dynamical equation is t hen equal to: [ ] dSp+, = _ - - F bp+l oSi. + ~'~ bp+l,iSi dt A ' i=1 - - s +o. ( 22) The adaptive version of the control scheme (20) is then given by: F = a Z( S* - Sp+~) + ~Sp ~( Z - Sp+l), (23) p + l -bp+l,oSin - E bp+l,iSi i=1 where the estimate of pp+~ is updat ed by t he least squares algorithm (18a-d) and Z is estimated by using the dynamical equation (21), i.e.: dZ F [ p + l ] d--7 = -~-[bp+x. 0Si. + ~--~a bp+,,iSi - kd(Z -- Sp+, ). (24) The adaptive controller (18), (23), (24) is schematized in Fig. 2. It has been i mpl ement ed on the above example. The paramet ers of t he system have been based and derived from a fixed bed anaerobic digestion pilot plant (Guiot e t al . , 1988). Here the substrate concentration is t he organic mat t er concentration expressed in COD (Chemical Oxygen Demand) units and X is the concentration of active biomass (in VSS (Volatile Suspended Solids) units). In the simulations, the paramet ers have been set to t he following values: A = 0 . 0 2 m 2 , H= l m, k1=0. 4, kd = 0.05 h -1. For simulation purposes, we have chosen a FIG. 2. Scheme of the closed loop system. Modelling and control of distributed paramet er bi oreact ors 879 Contois model for the specific growth rate #: #m~S # = K c X + S ' with #max = 0.35 h -1 and Kc = 0.4. As been pointed out in Section 3, the choice of appropriate basis orthogonal functions and number and position of the p interior collocation points is a very important question in the reduction procedure of the distributed paramet er model. As a mat t er of fact, a certain number of papers deal with that subject (e.g. Georgakis e t a l . , 1977; Wysocki, 1983; Babary e t a l . , 1991). For instance, Wysocki (1983), in his paper on first order hyperbolic systems (i.e. the kind of equations we deal with in this paper) suggests choosing the collocation points as the zeros of Legendre polynomials. Moreover, his conclu- sions are that three or four interior collocation points are sufficient for the simulation and control of these systems. In our simulations, we have been testing different solutions. Following Michelsen and Villadsen (1972) and a number of applications (e.g. Hariri, 1985), the Lagrange orthogonal polynomials, which are assumed to be reliable and are easy to comput e, have been chosen as the orthogonal basis functions. The collocation points have been chosen as the zeros of Jacobi polynomials P( p~' O) ( z) (of which the Legendre polynomial is a particular case) (see also Michelsen and Villadsen, 1972). The Jacobi polynomials can be comput ed by using the following recursive expression: P~"t 3)(z ) = [ z - g , ( a , /3, n ) ] P ( , ~ ) ( z ) - h , ( o 6 / 3 , n ) P ~ O z ) ( z ) n = 1 to p, with 0, / 3>- 1, P_a( z) =O, Po ( z ) =l , ha=O, h n - - (n - 1)(n + a ~ - 1)(n +/3 - 1)(n + a~+ fl - 1) ( 2n+ a ' + f l - 1)(2n + t r + / 3 - 2)2(2n + t r + / 3 - 3) ' for n > l 1 [ a' 2-/ 32 ] f o r n > O. g , = ~ 1 ( 2 n + ~ + ~ 1 ) 2 _ 1 , Different values of the paramet ers a and /3 of the Jacobi polynomial, and different number of collocation points have been considered. From this study, we found out that the best choice of a and/ 3 is: a = O, / 3=4. Moreover, from the many simulations we have been performing, we also found out that, with the above choice of er and/ 3, a number p = 4 of 0.8.5- 0.3 0. ~ 0.2 S (gCOD/I) ~1~ 1 .. ~ j (ii) : ct = l$=2 I~.~ ~ ' ~ ( i i i ) : o ~ = O , 13 = 4 0 time (h) 1 ~6 . 3 . 80 Op e n l oop s i mul a t i on wi t h di f f er ent a a n d b. interior collocation points is sufficient for correctly simulating the process and that larger values of p do not significantly improve the quality of the simulations. The simulation study in open loop is summarized in Fig. 3, which shows one typical simulation experiment, i.e. the response of the system to a step of the influent substrate concentration Sin (from 5 to 6g l -a) at time t = 5 h, for three different sets of values of tr and r : (i): te = fl = 0 (Legendre polynomial) (ii): a~ = 2, fl = 2 (iii): t r - - 0, fl = 4 (the best choice). We have also compared the orthogonal colloca- tion approximation with a finite difference method, an also often used reduction met hod where the time and space derivatives are approximated by finite differences: a t A t ' O z A z As pointed out in the literature (e.g. Goodson and Polls, 1978), in order to have a satisfactory accuracy, a high number of discretization points may be necessary when using finite difference approximation which may then require excessive comput er time. This phenomenon has been observed here: we had to go up to 500 space discretization points in order to obtain a behaviour which tends to l ook as good as the one obtained by the orthogonal collocation met hod, and this resulted in an increase of computer time larger than one order of magnitude. The results with the finite difference met hod for three values of space discretization points (N = 50, 200, 500) and corresponding to the same simulation experiment shown in Fig. 3 are shown in Fig. 4. From this exhaustive simulation study, we decided, for bot h the simulation of the process and the controller design, to use the orthogonal collocation met hod with the Jacobi polynomial 880 D. DOCHAIN e t al. 0 . 4 0 . 3 5 0. 3 0 . 2 5 S (gCOD/1) t ,.., I N=200 N=500 0 . 2 t ~ o ' ~b ~o ~o so t i me ( h) FIG. 4. Op e n l oop s i mul a t i on wi t h a fi ni t e di f f er ence a ppr oxi ma t i on. for t he above avalues of or, fl and p (0~ = 0, f l = 4 , p =4) . The four i nt eri or col l ocat i on points are t hen pl aced at t he fol l owi ng spatial positions: z~ = 0.31213, z2=0. 5789, z3 = 0.81289, z4 = 0.9627. The values of t he par amet er s bi/ are expressed hereaft er: [ b i j ] i = 1 t o 5 , j = 0 t O 5 "--0.42276 0.07134 = -0. 02323 0.00986 -0. 01429 B The initial - 5. 53249 11.9772 -1. 17311 - 3. 77846 0.28574 -1. 95405 - 0. 11059 0.59899 0.15738 - 0. 82149 -13. 95651 21.36328 9.34809 - 11. 33316 -4. 51751 13.27289 -3. 35601 -14. 97153 4.04364 -40. 36524 si mul at i on - 13.42872- 6.86530 - 7. 06384 17.82928 37 condi t i ons have been chosen so as t o correspond t o a st eady-st at e: Sin(0 ) = 5 g COD I-~,F(0) = 2 1 h -' Xl (0) = 29.4026gVSS I-~, S, (0) = 1. 9602gCOD 1-' X2(0) = 13.2069 g VSS 1- t, S:(0) = 0.8805 g COD l - Xd0) = 6.5457 g VSS 1-~, $3(0) = 0. 4364g COD 1-1 X4(0) = 4.1758 g VSS 1- ~, $4(0) = 0.2784 g COD l - Xd0) = 3.734 g VSS 1-1, S5(0) = 0.2489 g COD V ~. In Fig. 5, t he cont rol l er ( 20) , (25) has been i mpl e me nt e d under t he f ol l owi ng i ni ti al condi - ti ons and design par amet er s: S * = $ 5 ( 0 ) , 3. =2, ~5( 0) =0. 181gCOD- ~h -1, 2( 0) = 1.59 g COD 1 -~ 0. 3 0.28 0 . 2 6 0.24 0. 2 0.003 0 . 0 0 2 5 0 . 0 0 2 0. 0015 S ( g COD/ I ) AS i n 0 ~ ' ' IQ0 ~10 t i me ( h ) F ( m 3 / h ) 100 ' time(h) 0.22 0.21 l 0. 2 0 . 1 9 0. 18 - 0.17 0 a ( l / g COD/ h ) t i me ( h ) 1.9- 1.8" 1.7" 1 . 6 - 1. 5 1.4 z (g COD/D Y t i me ( h ) d Fr o. 5. Ad a p t i v e cont r ol of a f i xed bed bi or eact or . Modelling and control of distributed paramet er bi oreact ors 881 ( \ g C O D / l ) t = 0 : _ - - " * ~ t=40 :+++ * , t "\X -- t - - 250 : / i = J i i J 0 0.2 0.4 0 . 6 0 . 8 1 z (m) FIG. 6. Spatial repart i t i on of t he subst rat e concent r at i on S. 0 . 2 5 1 l S (gCOD/I) " 1 P l r e g u l a t o r : / A d a p t i v e l i n e a r i z i n g c o n t r o l l e r : - - - 0 . 2 5 1 0 . 2 4 9 - 0 . 2 4 5 t 0 . 2 4 7 [ , , , , , , , , , o ~o ~o tim e(h ) FiG. 8. Compari son wi t h a PI regul at or. This choice corresponds to a 10% error on the initial values of bot h Z and p. Figure 5 shows a simulation experiment in which the reference out put S* has been set to 0.26 g COD 1-1 at time t = 24 h, and a ramp of the disturbance input Sin has been applied bet ween time t = 99h and t = 100 h from 5 to 6 g COD 1-1. The controller proves successful t o drive the system t o the new steady state and to compensat e the effect of the disturbance, this in spite of wrong initial values for the paramet er p and the auxiliary variable Z. The wrong initialization explains the variation of the out put $5 from the its desired value S* at the beginning of the simulation. Not e that, because of the on-line estimation scheme, the closed-loop behavi our does not exhibit a first-order dynami- cal behavi our (as in the "i deal " (non-adaptive) situation of equation (16)) but a second-order one with damped oscillations. In Fig. 6, the spatial repartition of the substrate concentration S(t, z) along the react or is shown at three different instants ( t = 0, 40, 250 h). We have tested the robustness of the adaptive control law in presence of uncertain paramet er values. Figure 7 shows one typical simulation 0.3 1 S (gCODB) 1 0.28 ASin _ _ i 0.24 ! 0.22 ] , , , , . . . . . 0 I00 200 tim e(h ) PIG. 7. Robust ness of t he adapt i ve cont r ol l er wi t h respect t o model inaccuracies. experiment where a wrong value for the deat h paramet er kd (kd = 0.06, i.e. a 20% error) has been introduced in the control algorithm (while the other simulation conditions are those of Fig. 5). Not e the excellent performance of the adaptive controller in presence of this model uncertainty, which is very close to the one of Fig. 5. We have also compared our control algorithm with a simple classical PI regulator which computes the value of the influent flow rate F from the regulation error (S*-Sp+O at the reactor out put and which is characterized by the following transfer function: Hpi(S) = Kp 1 + where s is the Laplace variable. Figures 8 and 9 show the performance of the PI when a ramp of the disturbance input Sin is applied bet ween time t = 39 h and t = 40 h from 5 to 6 g COD 1-1. We have not ed that, although reasonable sets of design parameters have been considered, the performance of the PI may be degrading quite easily (Fig. 8; Kp =0. 1, T~ =2 ) and that the system can even be led to instability (Fig. 9, 0.3" s (gCOD/l) 0.28 0.26 . : : ~ , j . . . ~ v~.,:':~ 0.$4 t 0 1 0 0 2 0 0 tim e(h) FIG. 9. Compari son wi t h a PI regulator: unst abl e behavi our. 882 D. DOCHAIN et al. Kp = 0 . 1 5 , T/ = 0 . 8 ) . Thi s s hows t ha t a car ef ul c ont r ol des i gn ba s e d i n pa r t i c ul a r on t he we l l - known nonl i ne a r a nd di s t r i but e d p a r a me t e r f eat ur es o f t he bi opr oc e s s ma y i mp r o v e sig- ni fi cant l y t he c ont r ol p e r f o r ma n c e . Acknowledgements--This work has been initiated during the 4 month stay (from March to June 1989) of D. Dochain as "chercheur associ6" of the CNRS (France) at the LAAS, which is gratefully acknowledged. The authors would also like to thank Dr M. Perrier, Prof. S. B. Jorgensen and Prof. A. Munack for stimulating comments about this work. 6. CONCLUSIONS I n t hi s p a p e r , we deal t wi t h t he c ont r ol o f fi xed bed bi or e a c t or s . Th e mode l l i ng a nd c ont r ol desi gn has c o n c e n t r a t e d o n fi xed b e d b i o r e a c t o r s in t he a bs e nc e o f di f f usi on ef f ect s. A ge ne r a l dyna mi c a l mo d e l f or fi xed be d bi or e a c t or s has be e n i nt r oduc e d. Re d u c t i o n of t he di s t r i but e d p a r a me t e r mo d e l t o o r d i n a r y di f f er ent i al e qua - t i ons has b e e n p e r f o r me d by usi ng an o r t h o g o n a l col l ocat i on me t h o d . Th e r es ul t i ng dyna mi c a l mode l has s e r ve d as a basi s f or t he des i gn o f an a da pt i ve c ont r ol a l gor i t hm whi ch is b a s e d o n t he we l l - known nonl i ne a r s t r uc t ur e of t he mo d e l a n d do not r equi r e a ny a p r i o r i k n o wl e d g e f or t he pr oces s ki net i cs. Th e c ont r ol p e r f o r ma n c e has be e n i l l ust r at ed a nd di s cus s ed t h r o u g h s ever al s i mul at i on e xpe r i me nt s . I n par t i cul ar , its pe r f or - ma nc e has b e e n c o mp a r e d wi t h t hos e o f a cl assi cal PI . I t is wo r t h not i ng t ha t t he a p p r o a c h p r o p o s e d in t hi s p a p e r can be easi l y e x t e n d e d t o o t h e r t ypes o f di s t r i but ed p a r a me t e r bi or e a c t or s ( ai r lift bi or e a c t or s ( e. g. Lu t t ma n et al . , 1985) a nd f l ui di zed b e d r e a c t or s ( t he e xt e ns i on is st r ai ght - f or wa r d if t he b e d por os i t y e is a s s u me d t o be cons t ant ; see Ponc e l e t , 1986)) a n d t o bi or e a c t or s whe r e t her e exi st l ongi t udi nal di f f usi on ef f ect s (see e. g. Mos e r , 1988; Lu t t ma n et al . , 1985). Bes i des , it is wel l k n o wn t ha t di s t r i but e d p a r a me t e r r e a c t or s ma y exhi bi t a n o n - mi n i mu m p h a s e dyna mi c a l be ha vi our . I n o r d e r t o deal wi t h t hat p h e n o me n o n , t he p r o p o s e d c ont r ol l aw ma y a p p e a r t o be uns t abl e a nd a modi f i e d ver s i on of t he l i near i zi ng c ont r ol l aw has t o be i nt r oduc e d. I n Ba b a r y et al. ( 1990) , we ha ve p r o p o s e d t o modi f y it in a wa y whi ch is i n l i ne wi t h o t h e r cl assi cal wor ks on t he c ont r ol o f n o n - mi n i mu m pha s e s ys t ems ( e. g. Cl a r ke , 1984). As a ma t t e r o f f act , t he c ont r ol l e r des i gn is t h e n b a s e d o n t he mi ni mi zat i on, wi t h r es pect t o t he i nput F, of t he f ol l owi ng cos t f unct i on: J = {(Y* - Yt+~) - , k( r * - lit)} + q { F t - F t _ l } 2, q>--O. No t e t ha t it is exact l y t he Ge n e r a l i z e d Mi n i mu m Va r i a nc e ( GMV) c ont r ol l e r if ). = 0 ( e. g. Cl a r ke , 1984 and ( f or an a ppl i c a t i on t o n o n l i n e a r bi ochemi cal s ys t ems ) Do c h a i n a nd Bas t i n, 1984) and t ha t it r e duc e s t o a di s cr et e- t i me ver s i on o f t he c ont r ol l e r (17) if q = 0. REFERENCES Babary, J. P., N. Tali-Maamar and D. Dochain (1990). Modelling and control of distributed parameter bioreac- tors. Proc. 1st Int. Conf. on Modelling and Control of Biotechnol., Ecolog. and Biomed. Systems, Varna, Bulgaria, 20-25 September. Babary, J. P., T. Damak and N. Tali-Maamar (1991). On simulation of distributed parameter bioreactors. In R. Vichnevetsky and J. J. H. Miller (Eds), Proc. 13th 1MACS World Congress on Computation and Applied Mathematics, Dublin, 22-26 July, Voi. 3, pp. 1498-1499. Bastin, G. and D. Dochain (1988). Nonlinear adaptive control algorithms for fermentation process. In W. J. Book (Ed.), Proc. ACC, Invited Session on Advances in Control of Biotechnical Processes, pp. 1124-1128. IEEE Publ. Serv. Bastin, G. and D. Dochain (1990). On-line Estimation and Adaptive Control of Bioreactors. Elsevier, Amsterdam. Byrnes, C. I. and A. Isidori (1985). Global feedback stabilization of nonlinear systems. Proc. 24th I EEE CDC, Ft Lauderdale, pp. 1031-1037. Cho, Y. S. and B. Joseph (1983). Reduced-order steady-state and dynamic models for separation processes. AI ChE J., 29, 261-269. Clarke, D. W. (1984). Self-tuning control of nonminimum- phase systems. Automatica, 20, 501-517. Dochain, D. (1986). On-line parameter estimation, adaptive state estimation and adaptive control of fermentation processes. Doctoral Dissertation, Univ. Cath. Louvain, Belgium. Dochain, D. and G. Bastin (1984). Adaptive identification and control algorithms for nonlinear bacterial growth systems. Automatica, 20, 621-634. Georgakis, C., R. Aris and R. Amundson (1977). Studies in the control of tubular reactors--I. General considerations. Chem. Eng. Sci., 32, 1359-1369. Goodson, R. E. and M. P. Polis (1978). Identification of parameters in distributed systems. In W. H. Ray and D. G. Lainiotis (Ed.), Distributed Parameter Systems, Identification, Estimation and Control, pp. 47-133. Marcel Dekker, New York. Guiot, S. R., A. Pauss, D. Bourque, M. El Housseini, L. Lavoie, C. Beaulieu and R. Samson (1988). Effect of upflow liquid velocity on granule size distribution in an upflow anaerobic bed-filter (UBF) reactor. Proc. 5th Int. Syrup. on Anaerobic Digestion, pp. 121-124. Bologna, Italy. Hariri, A. (1985). Mod61isation et commande optimale de processus hyperboliques. Application aux processus thermiques h contre courant. Doctoral Dissertation, LAAS no 85125, Universit6 Paul Sabatier, Toulouse, France. Holmberg, A. (1982). On the practical identifiability of microbial growth models incorporating Michaelis-Menten nonlinearities. Math. Biosch., 62, 23-43. Isidori, A. (1989). Nonlinear Control Systems: An Introduction, 2nd ed. Springer Verlag, Berlin. Isaacs, S., A. Munack and M. Thoma (1986). Use of orthogonal collocation approximation for parameter identification and control optimization of a distributed parameter biological reactor. AI ChE meeting, Miami Beach. Jensen, K. F. and W. H. Ray (1982). The bifurcation behavior of tubular reactors. Chem. Eng. Sci., 37, 199-222. Jorgensen, S. B. (1986). Fixed bed reactor dynamics and control--a review. Proc. IFAC Control of Distillation Columns and Chemical Reactors, Pergamon, Oxford, pp. 11-24. Modelling and control of distributed parameter bioreactors 883 Luttman, R., A. Munack and M. Thoma (1985). Mathematical modelling, parameter identification and adaptive control of single cell protein processes in tower loop bioreactors. Advances in Biochemical Eng., Springer Vedag, Berlin, 32, 95-205. Maendenius, C. F. , B. Mattiasson, J. P. Axelsson and P. Hagander (1987). Control of an ethanol fermentation carried out with alginate entrapped Saccharomyces cereviaiae. Biotech. and Bioeng., 29, 941-949. Michelsen, M. L. and J. Villadsen (1972). A convenient computational procedure for collocation constants. Chem. Eng. J., 4, 64-68. Moser, A. (1988). Bioprocesses Technology. Kinetics and Reactors. Springer Verlag, New York. Poncelet, D. (1986). Analyse fondamentale des lits fluidis(~s biologlques. Ph.D. Thesis, Univ. Liege, Belgium. Poncelet, D. , R. Binot, H. Naveau and E. J. Nyns (1985). Biotechnologie des lits fluidis(~s en r(~acteurs cyfindriques et tronconiques. Trib. C~b~deau, 38(494), 3-12; 38(497), 33-48. Ray, W. H. (1981). Advanced Process Control. McGraw- Hill, New York. Schiigerl, K. (1989). Biofluidization: appfication of the fluidization technique in biotechnology. Can. J. Chem. Eng., 67, 178-184. Villadsen, J. V. and W. E. Stewart (1967). Solution of boundary-value problems by orthogonal collocation. Chem. Eng. Sci., 22, 1483-1501. Villadsen, J. and M. L. Michelsen (1978). Solution of Differential Equation Models by Polynomial Approxima- tion. Prentice-Hall, Englewood Cliffs, NJ. Williams, D. and D. M. Munnecke (1981). The production of ethanol by immobilized yeast cells. Biotechn. and Bioeng., 23, 1813-1825. Wysocki, M. (1983). Application of orthogonal collocation to simulation and control of first order hyperbolic systems. Math. and Comp. in Simulation, 25, 335-345.