UKACC International Conference on Control 2012 Cardiff, UK, 3-5 September 2012

Bilinear approach employed for modelling of continuous stirred tank reactor processes
Tomasz Larkowski, Leszek Koszalka and Keith J. Burnham
Control Theory and Applications Centre Coventry University, CV15FB Coventry, UK, Email: [email protected] Computer Systems and Networks Wroclaw University of Technology, Wroclaw, Poland,

AbstractIn this paper feasibility of modelling approach based on a bilinear system approximation is demonstrated on one of the most frequently met processes in chemical engineering, namely a continuous stirred tank reactor. Selected examples of such systems from the literature are considered and modelled with a use of dynamic bilinear systems. Advantages of this approach are presented and discussed.

II. M ODEL

STRUCTURES

HBS structure belongs to a sub-class of so-called output afne models, i.e. models that retain afnity w.r.t. the output signals, see [8]. It comprises of a cascade connection of a static (memoryless) nonlinearity followed by a dynamic timeinvariant afne BS and is given by
na nb na nb

I. I NTRODUCTION Chemical reactor is often described as the most important unit operation in a chemical process, see [1]. The task of its modelling and control is commonly encountered in the literature representing a practical nonlinear industrial problem, see [2], [3], [4]. A popular model of a chemical reactor is the so-called continuous stirred tank reactor (CSTR). CSTR consists of a closed tank to which an input stream is fed in and the output stream fed out in a continuous manner, whilst a content is constantly stirred. Modelling of CSTRs is challenging mainly due to i) possibility of occurrences of rapid reactions (the so-called ignition-extinction phenomena), hence fast changing process gain and dynamics, and also due to ii) nonlinear steady-state behaviour, see [1] and [5]. This paper demonstrates that the CSTRs can be modelled by employing an approach based on bilinear system (BS) description. In order to increase modelling exibility of BSs, an extension consisting of a static nonlinearity that transforms the input signal is proposed. Such a structure, referred to as a Hammerstein-bilinear system (HBS), see [6], [7], is considered and compared to BS and Hammerstein system (HS) models. The use of BS based approach is motivated twofold. First, BSs retain a close structural relationship with linear models, hence standard well understood notions from classical linear system theory such as system time constants, damping/natural frequency and steady-state gain are to large extent retained. This follows from the property that BS structure can be interpreted as a linear time-varying system, which also greatly facilitates the control design. Second, BSs preserve linearity w.r.t. the parametrisation, which aids in their identication by allowing for standard parameter estimation methods to be used. In this paper three different CSTR models are considered. Two of the models are isothermal, whilst the third model is an example of a diabatic CSTR, see [1]. yk =

aj ykj +
j=1 i=1

bi vki +
j=1 i=1

ij vki ykj +c (1) (2)

vk = f (uk )

where aj , bi , ij and c are model parameters. The bilinearity is dened as a product between system output yk and the intermediate input variable vk , and f () denotes a general scalar static nonlinear function. Note that not all bilinear coefcients must necessarily be present in (1), hence a particular structure can be obtained by setting selected ij to zero. The HBS can be interpreted as a generalisation of both of its constituent subsystems, i.e. HS and BS models. In particular, a BS is obtained from (1)-(2) by setting uk = vk , i.e. by selecting f (x) = x, which gives
na nb na nb

yk =
j=1

aj ykj +
i=1

bi uki +
j=1 i=1

ij uki ykj + c (3)

Similarly, a HS is obtained by setting ij = 0 i, j in (1)-(2), which leads to

yk =
j=1

aj ykj +
i=1

bi vki + c

(4) (5)

vk = f (uk )

Also, a linear (or more precisely an afne) structure is obtained by imposing both restrictions simultaneously, i.e. ij = 0 i, j and uk = vk . In this paper, for simplicity, it is assumed that the input static nonlinearity is modelled as a polynomial of order n , i.e.
n

f (x) =
l=1

l xl

(6)

978-1-4673-1558-6/12/$31.00 2012 IEEE

473

Consequently, a particular HBS structure is given by a quadruplet that denes the number of a, b, and coefcients, i.e. HBS(na ,nb ,n ,n ). Moreover, the sum of all coefcients plus unity (to account for an offset), corresponds to the total number of degrees of freedom (DoF) in a given structure. An analogous notation is used w.r.t. other structures that can be derived from the HBS model. III. PARAMETER ESTIMATION The HBS structure is bilinear in terms of parametrisation, due to the products between and a, and also between and coefcients. A well known approach to solve such problems is to use the so-called bilinear parametrisation method (BPM), see [9]. The BPM solves the estimation problem in a two step manner, where in the rst step parameters are xed and a, b, parameters are calculated, whilst in the second step a, b and remain xed and parameters are computed [10]. Because the two subproblems separately are linear w.r.t. the unknowns, the ordinary least squares algorithm [9] can be applied, which renders the overall procedure numerically efcient, and this is the approach used here. An analogous technique can also be applied to HS models, see [11], whilst the parameters of afne and BS structures can be estimated by using a single ordinary least squares technique. IV. S IMULATION STUDIES A. Performance criteria and experimental setup In order to quantify the accuracy of models obtained, two performance criteria are used. Namely, the coefcient of determination and the (normalised) integral of absolute error, dened, respectively, as follows R2 = 100 1 T IAE = 1 N
N

the identication experiment more realistic, the measured output is assumed to be contaminated with an additive, normally distributed, white and zero-mean disturbances such that the resulting signal-to-noise ratio is approximately 37dB. The second validation data set comprises of a monotonic staircase input, which allows the performance of the identied models to be evaluated with the emphasis placed on the steady-state behaviour. Also, to provide an indication of the complexity of the models, the corresponding DoF are considered. B. Isothermal CSTR with a rst-order irreversible reaction 1) System description: The rst isothermal CSTR model considered, referred to as the CSTR1, is given by the following equations, see [1] for details, i.e. dCA (t) F (t) F (t) = CAf + k CA (t) dt V V F (t) dCB (t) = CB (t) + kCA (t) dt V
k

(9) (10)

yy yy

2 2 2 2

(7) (8)

|yk yk |
k=1

where y and y denote vectors composed of the measured (noisy) system outputs and outputs generated by the estimated model, respectively, and y is the mean value of y. The notation 2 denotes the Euclidean norm. Because the main interest of experiments lies in the modelling capabilities of the model structures considered, relatively long input-output data consisting of N = 20, 000 samples are used. Three data sets are considered, i.e. identication and two validation data sets. The sampling time is chosen as 0.1s. In the case of the identication data set and the rst validation data set, the input signal is generated as a series of uniformly distributed steps between the minimal and maximal range for a given system. The probability of transition to a different level is selected randomly with a uniform switching probability of 10%, providing a reasonable compromise between the content of transient and steady-state data. Additionally, to ensure that the input is sufciently exciting, a normally distributed, white and zero-mean noise sequence of comparatively small variance is added. To render

and describes a rst-order irreversible reaction A B where k is the reaction rate per unit volume. The remaining variables are: CA (t), CB (t) - concentrations of substances A and B inside the tank of a constant volume V , respectively, F (t) - inow/outow mass rate, CAf - inow concentration of substance A. Only the substance A is present in the inow stream, and inow and outow mass rates are equal. The actual units are unimportant and hence are not included. It is assumed that the manipulated variable is F (t) and that CB (t) is the output of interest. The task consists of identifying a model between F (t) and CB (t). The values of the parameters were chosen as V = 1, k = 0.2, CAf = 1 and the initial states of the process as Ca0 = Cb0 = 0.5, where the subscript zero denotes the initial value. The input F (t) is in range of (0, 1]. It is observed that a product, i.e. bilinearity, between the input F (t) and the state CA (t) occurs in (9) and that an analogous product between the input F (t) and the state CB (t) is also present in (10). The steady-state characteristic of the CSTR1 model is plotted in Figure 1, where it is observed that the curve resembles the type of steady-state characteristics typical of BSs. Therefore, these observations substantiate the usage of a bilinear based modelling approach. 2) Identication results: Selected identication results are given in Table I. First, it is observed that the model is considerably nonlinear because the third order afne structure resulted in R2 below 90% for all three data sets. Further increase of T the order of the afne structure does not lead to any signicant improvements in modelling performance. HS models show clear improvement, allowing for R2 of approximately 98% T in the case of the HS(1,1,4) to be achieved for all three data sets. Further increase of the order of the input polynomial does not lead to signicant improvements in model tting. This is due to the fact that the nonlinear steady-state characteristic is not a complex function, and it is rather the changing system dynamics that is not captured by the HS type structures. When considering the results obtained from BS models, an evident

474

1 0.9

Input
0.8 0.6 0.4 0.2 0 300

0.7

CBss

0.6 0.5 0.4 0.3 0.2 0 0.2 0.4

F (t)

0.8

400

500

600

Output
0.8

CB (t)

0.6 0.4 0.2 300 400 500

700 800 actual afne(3,3) HS(1,1,4) BS(2,1,1)

Fss

0.6

0.8

Fig. 1.

CSTR1 - A steady-state characteristic.

time [s]

600

700

800

Fig. 2. CSTR1 - Selected representative results of identication on validation data set 1 in the interval [300, 800]s. structure DoF afne(3,3) HS(1,1,2) HS(1,1,3) HS(1,1,4) BS(1,1,1) BS(2,1,1) HBS(1,1,1,2) HBS(1,1,1,3) 7 5 6 7 4 5 6 7 Id. data set R2 IAE T 103 80.40 54.49 96.70 24.04 97.81 15.51 98.10 13.75 99.97 2.265 99.98 1.845 99.97 2.275 99.99 1.421 Val. data set 1 R2 IAE T 103 80.08 55.27 96.10 25.13 97.83 17.16 98.10 15.68 99.97 2.276 99.98 1.851 99.97 2.304 99.97 2.260 Val. data set 2 R2 IAE T 103 88.11 74.10 94.78 31.00 96.92 19.39 97.49 17.57 99.97 2.716 99.99 1.983 99.97 2.748 99.99 1.751

Input
0.8

F (t)

0.6 0.4 0.2 20 40 60 80 100

Output
0.8

TABLE I CSTR1 - Q UANTIFIED IDENTIFICATION AND VALIDATION RESULTS FOR MODEL STRUCTURES CONSIDERED .

CB (t)

0.6 0.4 0.2 20 40 60

actual afne(3,3) HS(1,1,4) BS(2,1,1)

time [s]

80

100

Fig. 3.

CSTR1 - Selected results of identication on validation data set 2.

improvement in the approximation performance is noted, i.e. the R2 of almost 100% is obtained with the IAE criterion T decreasing by approximately 7 times, when compared to the best HS model. This means that both, the nonlinear steadystate characteristic and the changing system dynamics are approximated well by the BS structures with only 4 or 5 DoF. Because the only source of nonlinearity in the underlying process equations arises from product terms, cf. (9)-(10), this result could have been anticipated. Representative graphical results of the identication are given in Figures 2 and 3, showing the performances of the selected estimated models on arbitrarily chosen intervals of the validation data sets 1 and 2, respectively. It is observed that in the case of both gures the actual system output is virtually undistinguishable from that generated by the identied BS(2,1,1) structure. Consequently, it is concluded that a BS structures are appropriate for modelling the CSTR1 process.

C. Isothermal CSTR with the Van de Vusse reaction 1) System description: The second isothermal CSTR model considered, see [1], referred to as the CSTR2, is given by dCA (t) dt dCB (t) dt dCC (t) dt dCD (t) dt F (t) 2 CAf CA (t) k1 CA (t)k3 CA (t) V F (t) = CB (t) + k1 CA (t) + k2 CB (t) V F (t) = CC (t) + k2 CB (t) V F (t) 1 2 = CD (t) + k3 CA (t) V 2 = (11) (12) (13) (14)

with the behaviour governed by the so-called Van de Vusse reaction kinetics. The reactions A 1 B 2 C 2A D
k3 k k

(15) (16)

475

0.45 0.4

Input
8

0.35 0.3

F (t)

6 4 2 100 200 300 400 500

CBss

0.25 0.2 0.15 0.1

Output
0.5

CB (t)

0.05 0 0 1 2 3 4 5 6 7 8 9

0.4 0.3 100 200

Fss

Fig. 4. structure DoF afne(3,3) 7 HS(1,1,4) 7 8 HS(1,1,5) HS(1,1,6) 9 10 HS(1,1,7) HBS(1,1,1,3) 7 HBS(1,1,1,4) 8 HBS(1,1,1,5) 9

CSTR2 - A steady-state characteristic. Id. data set R2 IAE T 103 10.92 44.68 84.07 15.67 91.48 10.70 94.99 8.552 96.43 6.591 92.96 10.72 98.11 4.335 98.34 3.257 Val. data set 1 R2 IAE T 103 7.724 46.06 84.03 16.86 91.16 12.38 94.56 8.803 96.09 6.800 92.78 11.43 98.11 4.726 98.33 3.551 Val. data set 2 R2 IAE T 103 5.539 54.19 89.50 25.10 94.40 16.13 97.32 11.28 98.61 8.118 98.23 11.31 99.86 3.000 99.60 3.747

time [s]

300

actual afne(3,3) HS(1,1,7) 500 400 HBS(1,1,1,4)

Fig. 5. CSTR2 - Selected representative results of identication on validation data set 1 in the interval [100, 500]s.

Input
8 6

F (t)

4 2 20 40 60 80 100

TABLE II CSTR2 - Q UANTIFIED IDENTIFICATION AND VALIDATION RESULTS FOR MODEL STRUCTURES CONSIDERED .
0.4

Output CB (t)
actual afne(3,3) HS(1,1,7) HBS(1,1,1,4)
20 40

0.2 0

are irreversible and described by the reaction rate constants k1 , k2 and k3 . The remaining variables are: CA (t), CB (t), CC (t), CD (t) - concentrations of substances A, B, C and D inside the tank of constant volume V , respectively, F (t) - inow/outow mass rate, CAf - inow concentration of substance A. Only the substance A is present in the inow stream, and inow and outow mass rates are equal. It is assumed that the manipulated variable is F (t), whilst CB (t) is the output of interest. Therefore, the modelling task consists of identifying a model between F (t) and CB (t). The values of the parameters were chosen as: V = 1, k1 = 5/6, k2 = 3, k3 = 10/6, CAf = 10 and the initial states of the process are Ca0 = Cb0 = 0. The input F (t) is postulated to vary between (0, 9]. It is observed that a bilinearity between the input F (t) and the state CA (t) is present in (11) and an analogous bilinearity between the input F (t) and the state CB (t) appears in (12). However, in contrast to (9) corresponding to the CSTR1 model, here the product term is not the only source of nonlinearity in the evolution of CA (t). This is due to the 2 presence of the expression CA (t) in (11), whose inuence is controlled by the rate constant k3 . Consequently, it is concluded that although the underlying system equations indicate that a bilinear type behaviour is present, it may be infeasible to model the process by a BS

time [s]

60

80

100

Fig. 6.

CSTR2 - Selected results of identication on validation data set 2.

only over the entire range of its operation. This hypothesis is conrmed by considering the steady-state characteristic of the CSTR2 process given in Figure 4, where it is observed that the system exhibits the input multiplicity (IM) property [3]. Such static behaviour cannot be captured by a BS, due to its inherent structural limitations. This stands in contrast to HS and HBS that are structurally capable of IM, hence are anticipated to be more appropriate for modelling of the CSTR2 process. 2) Identication results: The identication results are given in Table II, where a low value of R2 for all three data sets in T the case of the afne model indicates that the overall process is considerably nonlinear. Further increase of the order of the afne structure does not result in noticeable improvements in tting. The performance criteria of BS structures are not included in the table because it was not possible to t such models, i.e. the corresponding R2 values were negative. This T result conrms the inappropriateness of using BS structures for approximating processes exhibiting IM. A considerable

476

improvement is achieved in the case of HS structures, when compared to the afne model, with R2 of about 97% on averT age for all three data sets obtained by HS(1,1,7) with 10 DoF. The results obtained by the HBS(1,1,4) and HBS(1,1,5), with 8 and 9 DoF, respectively, show further tting improvements with both models achieving R2 of approximately 98% for the T identication data set and the validation data set 1, and almost 100% for the validation data set 2. Compared to the HS(1,1,7) the IAE was reduced by approximately threefold in the case of the rst two data sets, and fourfold in the case of the third data set. Representative graphical results obtained from the identication procedure are presented in Figures 5 and 6, showing the performances of the selected estimated models on arbitrarily chosen intervals of the validation data set 1 and 2, respectively. Whilst in the case of Figure 5 slight discrepancies between the actual output and that generated by the estimated HBS(1,1,1,4) model are observed, the two corresponding curves are virtually indistinguishable in the case of Figure 6. Also, it is noted that the system exhibits a non-minimum phase behaviour when F (t) is high and changes to a lower value. This behaviour, which is observed to be manifested by spikes in the CB (t) signal, increases the difculty of obtaining an acceptable approximation. Consequently, despite such a challenging task, the modelling results obtained by the HBS(1,1,1,4) structure can be treated as very satisfactory in overall. D. Diabatic CSTR 1) System description: The considered diabatic CSTR model, see [1] and [5], referred to as the CSTR3, is given by dCA (t) F (t) = CAf CA (t) k0 r(t)CA (t) dt V dT (t) F (t) H = Tf T (t) + k0 r(t) dt V cp US T (t) Tj (t) V cp (17)

9 8 7

CAss

6 5 4 3 2 275 280 285 290

Tjss

295

300

305

310

315

Fig. 7. structure DoF afne(3,3) HS(1,1,3) HS(1,1,5) HS(1,1,7) HS(2,2,3) HS(2,2,5) HS(2,2,7) BS(1,1,1) BS(2,1,1) BS(2,2,2) HBS(1,1,1,3) HBS(1,1,1,5) HBS(1,1,1,7) HBS(2,2,2,3) HBS(2,2,2,5) HBS(2,2,2,7) 7 6 8 10 8 10 12 4 5 7 7 9 11 10 12 14

CSTR3 - A steady-state characteristic. Id. data set R2 IAE T 103 88.86 99.53 96.92 49.26 97.31 45.51 97.89 39.41 97.19 45.61 97.61 41.52 98.16 35.46 98.98 27.52 98.90 28.35 99.16 23.35 99.13 25.63 99.19 24.61 99.28 22.50 99.39 20.96 99.46 19.53 99.56 17.25 Val. data set 1 R2 IAE T 103 88.84 99.38 96.88 49.52 97.30 45.63 97.86 39.74 97.15 45.88 97.57 41.80 98.12 35.77 98.96 27.09 98.88 28.75 99.13 23.90 99.11 26.00 99.18 24.98 99.27 22.97 99.37 21.40 99.44 20.05 99.54 17.84 Val. data set 2 R2 IAE T 103 90.73 96.42 96.65 54.41 97.20 49.03 98.00 40.15 98.46 40.48 98.79 36.90 99.16 31.27 99.20 23.82 99.14 25.08 99.58 16.72 99.15 24.77 99.21 23.58 99.36 20.16 99.52 18.68 99.58 16.96 99.70 13.52

TABLE III CSTR3 - Q UANTIFIED IDENTIFICATION AND VALIDATION RESULTS FOR MODEL STRUCTURES CONSIDERED .

(18) Considering equations (17) and (18), it is noted that in each case bilinearities are present, i.e. products between F (t) and the state CA (t) in the rst equation and between F (t) and the state T (t) in the second equation. These, however, are clearly not the only contributions that render nonlinearity of the overall behaviour. This is due to the presence of nonlinear relationships involving an exponent of T (t) that appear in both equations. A steady-state characteristic of the CSTR3 model, given in Figure 7, shows the presence of the output multiplicity (OM) property [3]. Because non of the model structures investigated in this paper is structurally capable of OM, see [8], only a restricted range of the process operation is considered, i.e. the range of Tj [273, 306) within which OM is absent. Consequently, because the operating range is limited, it might be possible that a bilinear type behaviour will, in fact, be prevailing. The values of the parameters were chosen as: F (t) = 1,

where the rst order reaction rate per unit volume is given by the so-called Arrhenious expression, i.e. r(t) = exp E RT (t) (19)

The other variables are: CA (t) - concentration of substance A inside a tank of the constant volume V , F (t) - inow/outow mass rate, CAf - inow concentration of substance A, k0 pre-exponential factor, R - ideal gas constant, E - activation energy, T (t) - reactor temperature, Tf - inow (feed) temperature, Tj (t) - jacket temperature, U - overall heat transfer coefcient, H - heat of reaction, - density, S - area for heat exchange, cp - heat capacity. Only the substance A is present in the inow stream, and inow and outow mass rates are equal. It is assumed that the manipulated variable is Tj (t), whilst CA (t) is the output of interest. Therefore, the modelling task consists of identifying a model between Tj (t) and CA (t).

477

Input
300 300

Input

Tj (t)

Tj (t)
actual afne(3,3) 800 700 BS(2,2,2) HS(2,2,7) HBS(2,2,2,7)

290 280 600

290 280

900

1000

1100 9.5

20

40

60

80

100

Output CA (t)
9 8.5 8 20 40

Output

9.5

actual 120 140 afne(3,3) BS(2,2,2) HS(2,2,7) HBS(2,2,2,7)

CA (t)

9 8.5 8 600 700 800

time [s]

900

1000

1100

60

time [s]

80

100

120

140

Fig. 8. CSTR3 - Selected representative results of identication on validation data set 1 in the interval [600, 1100]s.

Fig. 9.

CSTR3 - Selected results of identication on validation data set 2.

V. C ONCLUSIONS V = 1, k0 = 9703 3600, H = 5960, cp = 500, U S = 150, E = 11843, R = 1.987 and the initial states of the process are CA0 = 8.5 and T0 = 305. 2) Identication results: The identication results obtained are collected in Table III, from where it is observed that the afne model achieved reasonable results close to 90% in terms of R2 in the case of all three data sets. Further increase of the T order of the afne structure does not provide any considerable improvements. First order HS structures yielded results that are better by approximately 8% on average in terms of the R2 criterion and approximately twice on average in terms T of the IAE criterion. This indicates a clear improvement and justies the need for a nonlinear model structure. The best tting among HS structures is obtained for a second order HS structure, i.e. HS(2,2,7) with a seventh order polynomial and 12 DoF in total. It is interesting to notice that these values are close to those produced by a relatively simple BS structure, i.e. BS(1,1,1), with only 4 DoF. A boundary of 99% in terms of the R2 criterion is exceed by a second order BS model with T 7 DoF. The tting is improved, if at all, only slightly by rst order HBS structures, and it is the second order HBS model, i.e. HBS(2,2,2,7), in the case of which the results improve more signicantly. However, this comes at the cost of 14 DoF, when compared to only 7 DoF in the case of BS(2,2,2). Representative graphical results generated from the estimated models are depicted in Figures 8 and 9, and demonstrate the performances of the selected models on arbitrarily chosen intervals of the validation data set 1 and 2, respectively. It is observed that the outputs of the estimated models BS(2,2,2) and HBS(2,2,2,7) are both virtually undistinguishable from the actual system output in both gures. Consequently, by taking into account the corresponding DoF and a pragmatic point of view, it is the second order BS that appears to be a preferable choice in this case. The paper has demonstrated feasibility of BS based modelling approach for approximating CSTRs. It has been shown that BS models are capable of capturing both, i.e. the dynamic and static behaviour of the exemplary CSTR systems considered. In the case of the CSTR process exhibiting the IM property, a BS structure with an additional nonlinear memoryless element transforming the input, i.e. a HBS structure, has shown to be an appropriate choice. R EFERENCES
[1] W. Bequette, Process Dynamics: Modeling, Analysis and Simulation, 1st ed., ser. Series in the Physical and Chemical Engineering Sciences. Prentice Hall PTR, 1998. [2] E. Meadows and J. Rawlings, Model predictive control. Prentice Hall, Englewood Cliffs, NJ, 1997, ch. 5, pp. 233310. [3] M. Pottmann and R. K. Pearson, Block-oriented NARMAX models with output multiplicities, American Institute of Chemical Engineers, vol. 44, no. 1, pp. 131140, 1998. [4] K. Murakami and D. E. Seborg, Constrained parameter estimation with applications to blending operations, J. of Process Control, vol. 10, no. 23, pp. 195202, 2000. [5] G. Stephanopoulos, Chemical Process Control: An Introduction to Theory and Practice. Prentice Hall, 1983. [6] I. Zajic, T. Larkowski, M. Sumislawska, D. Hill, and K. J. Burnham, Modelling of an air handling unit: a Hammerstein-bilinear model identication approach, in Proc. of 21st Int. Conf. on Systems Engineering, Las Vegas, USA, 2011, pp. 5961. [7] I. Zajic, T. Larkowski, D. Hill, and K. J. Burnham, Energy consumption analysis of HVAC system with respect to zone temperature and humidity set-point, in Proc. of 18th IFAC World Congress, Milan, Italy, 2011, pp. 45764581. [8] R. K. Pearson, Selecting nonlinear model structures for computer control, J. of Process Control, vol. 13, no. 1, pp. 126, 2003. [9] L. Ljung, System Identication - Theory for the User, 2nd ed. New Jersey, USA: Prentice Hall PTR, 1999. [10] T. Larkowski and K. J. Burnham, Bilinear approach to modelling of continuous stirred tank reactor process, in Proc. of 9th European Workshop on Advanced Control and Diagnosis, Budapest, Hungary, 2011, pp. 168173. [11] F. Le, I. Markovsky, C. T. Freeman, and E. Rogers, Identication of electrically stimulated muscle models of stroke patients, Control Engineering in Practice, vol. 18, no. 4, pp. 396407, 2010.

478